Compare commits

...
Author SHA1 Message Date
müde
216eb8ae9b docs(turn-loop): split into index + claude-invocation/config/mcp sub-pages, condense the loop 2026-07-05 23:04:48 +02:00
müde
746f518250 fix(agent): show Compacting state during in-flight operator /compact 2026-07-05 22:54:20 +02:00
müde
59dd7955aa fix(stats): last_usage orders by id (insertion order); clarify Window default 2026-07-05 22:53:16 +02:00
müde
f2f104627d refactor(agent): drop vestigial install_configured param, non_empty_env helper, explicit infallible expect in client 2026-07-05 22:49:45 +02:00
müde
9cebc128e2 refactor(web_ui): Bus::effective_context_window, sigint_claude + per-connection hello, relocate fetch_reminder_stats, drop redundant guard 2026-07-05 22:47:14 +02:00
müde
77f31b44bb refactor(agent): fire_and_forget helper, drop dead default_model + compact_percent fallback, fix stale comment 2026-07-05 22:43:23 +02:00
müde
9721be7bc3 docs(turn_stats): mark unread columns as forward-capture, not consumed 2026-07-05 22:39:45 +02:00
müde
f490e07509 refactor(prompt): drop unreachable role-block defensive branches; keep marker grammar 2026-07-05 22:36:10 +02:00
müde
20ff891962 refactor(forge_notify): share render_body_excerpt across formatters; fix nits + docs 2026-07-05 22:32:00 +02:00
müde
dd7f8c5ebb docs(turn-loop): document PromptTooLong archive+requeue, TurnOutcome=Result, credential-file session detection 2026-07-05 22:27:09 +02:00
müde
90af3e8d0b refactor(mcp): split claude launch-config layer into mcp_config.rs 2026-07-05 22:24:46 +02:00
müde
e1e3bda94a style: apply nix fmt 2026-07-05 22:24:42 +02:00
müde
9a0d525fdf refactor(mcp): drop vestigial SocketReply, match hive_sh4re::Response directly; remove dead code 2026-07-05 22:12:45 +02:00
müde
543177ed74 refactor(mcp): collapse repeated failure-arm triplet into reply_err helper 2026-07-05 22:05:06 +02:00
müde
785a36b907 refactor(web_ui): share one broker_request helper; /send timeout now 409 not 500 2026-07-05 21:55:28 +02:00
müde
d83d03e4c0 fix(agent): key session detection off credential filenames; own clear_session in login.rs 2026-07-05 21:48:56 +02:00
müde
d004fdc0eb fix(stats): set busy_timeout on agent snapshot reader so a concurrent write can't blank the page 2026-07-05 21:43:25 +02:00
müde
2b5edef4d1 fix(agent): recover PromptTooLong via archive+requeue; model TurnOutcome as Result 2026-07-05 21:42:13 +02:00
31 changed files with 1720 additions and 1715 deletions

View file

@ -34,9 +34,10 @@ hand-maintained per-file tree drifts out of sync with the code.
operator dashboard (`dashboard.rs`). Largest crate.
- **`hive-ag3nt/`** — in-container harness; one `hive` binary for every
agent. Turn-loop *policy* layer (`turn.rs`) over the `hive-claude`
driver, embedded MCP server (`mcp.rs`), per-agent web UI (`web_ui/`
module dir), event + turn-stats sqlite sinks, login flow, system-prompt
renderer, forge-notify subscriber.
driver, embedded MCP server (`mcp.rs`) + its claude launch-config layer
(`mcp_config.rs`: tool-group/capability → `--allowedTools`, `--mcp-config`
render), per-agent web UI (`web_ui/` module dir), event + turn-stats
sqlite sinks, login flow, system-prompt renderer, forge-notify subscriber.
- **`hive-claude/`** — reusable, app-agnostic driver for headless
`claude --print`: spawns the CLI, streams + classifies stream-json,
parses per-turn `Telemetry`, and drives a durable self-compacting
@ -85,7 +86,10 @@ read them à la carte.
- **"How does the per-agent terminal classify + colour
events?"** → [`docs/terminal-rendering.md`](docs/terminal-rendering.md).
- **"How does claude get its prompt and what tools does it have?"** →
[`docs/turn-loop.md`](docs/turn-loop.md).
[`docs/turn-loop.md`](docs/turn-loop.md) (index: the loop, binary shape,
turn outcomes; sub-pages:
[`claude-invocation`](docs/turn-loop/claude-invocation.md),
[`config`](docs/turn-loop/config.md), [`mcp`](docs/turn-loop/mcp.md)).
- **"How do config changes flow from manager to operator to
container?"** → [`docs/approvals.md`](docs/approvals.md).
- **"What state survives destroy / purge / restart?"** →

View file

@ -360,7 +360,8 @@ the var is absent or empty.
**Updating the surface** — when a new `#[tool]` fn is added to `HiveServer`
in `hive-ag3nt/src/mcp.rs`, add its name to the matching `ToolGroup::tools()`
slice in `hive-sh4re/src/lib.rs`. That's the single source of truth;
`allowed_mcp_tools` reads it at session start.
`mcp_config::allowed_mcp_tools` (in `hive-ag3nt/src/mcp_config.rs`) reads it at
session start.
## Capabilities

View file

@ -213,7 +213,7 @@ Five shapes, distinguished by the notification's classification:
| Trigger | Wrapper |
| ----------------------------------- | --------------------------------------------------------------------------------- |
| Comment on issue / PR | `[comment on PR #N owner/repo] title\nurl: ...\n\nauthor: body\nassignee: ...` |
| Review submission | `[PR approved #N owner/repo] title\nurl: ...\n\nreviewer: body\nassignee: ...` |
| Review submission | `[PR approved #N owner/repo] title\nurl: ...\n\nauthor: body\nassignee: ...` |
| New issue / PR | `[new PR #N owner/repo] title\nurl: ...\n\n<body excerpt>\nassignee: ...` |
| Later activity (open, not creation) | `[activity on PR #N owner/repo] title\nurl: ...\n\n<body excerpt>\nassignee: ...` |
| State change | `[PR merged #N owner/repo] title\nurl: ...\nassignee: ...` |
@ -224,6 +224,10 @@ Review labels come from the Forgejo `state` field: `APPROVED` →
submitted yet — no peer-visible event). Unknown states fall back to
the generic comment wrapper.
A review submitted with **no body** renders `reviewed by: <author>` in
place of the `<author>: <body>` line — deliberately worded to not collide
with the meta-suffix `reviewer:` line (requested reviewers, below).
### "new" vs "activity on"
A review submitted with **no body** carries no `latest_comment_url`,

View file

@ -18,42 +18,37 @@ Each agent harness (`hive serve` — one binary for all agents) runs:
over stdin.
5. Stream stdout (JSON lines) into the bus as
`LiveEvent::Stream(value)`. Pump stderr as `Note`.
6. Wait for claude to exit. Compaction is two-pronged — *reactive*
on `Prompt is too long` and *proactive* on a context watermark
(see [Compaction](#compaction) below). **Rate-limit detection**:
on stderr the harness does a raw-line match for `429` /
`rate_limit` markers; on stdout it only fires on parsed
`{"type":"error"}` JSON events (avoiding false positives when
agents discuss `rate_limit_error` in conversation text). On
detection the harness sets the `rate_limited` sentinel
(`Bus::emit_status("rate_limited")`), sleeps
`HIVE_RATE_LIMIT_SLEEP_SECS` (default 300), then retries.
The dashboard and per-agent page show a `⊘ rate limited` badge
while the harness is parked. **Auth-failed detection**: both
stdout and stderr pumps also match
`AUTH_FAIL_MARKERS` (`"authentication_failed"`, `401`, etc.).
On the first 401, `drive_turn` retries the same prompt once
immediately (transient token-refresh races and brief API hiccups
can cause a 401 that clears on retry). Only if the retry also
returns `AuthFailed` does `drive_turn` bubble it up to the serve
loop, which then writes `{state_dir}/hyperhive-needs-login`,
emits `needs_login_idle` status, requeues the inflight message
(so it replays after re-auth), and parks in `wait_for_login`
the same path used at boot. The operator re-authenticates via
the per-agent web UI login flow; on success the sentinel is
cleared and the queued message drives the next turn normally.
**Mtime-snapshot resumption**: `wait_for_login`
snapshots the `~/.claude/` dir (newest file mtime + file count)
at entry and only resumes when that snapshot advances — not
just when credentials exist on disk. This prevents a silent
infinite-401 loop: stale credentials already on disk at the
time of the 401 no longer cause an immediate false-resume. The
`DirSnapshot` struct tracks both axes; either a mtime advance
OR a file-count change triggers resume (the count axis handles
filesystems where `modified()` errors on every file).
6. Wait for claude to exit and classify the turn's outcome from the
stream + exit — success, compaction, rate-limit, auth-failure, or
hard failure. The outcome drives the post-turn action (see
[Turn outcomes](#turn-outcomes)); compaction is handled inside the
session (see
[Compaction](turn-loop/claude-invocation.md#compaction)). Rate-limit
and auth-failure detection is described [below](#failure-detection-and-login).
7. Emit `LiveEvent::TurnEnd { ok, note }`. Sleep `poll_ms` to avoid
tight loops on transient failures.
### Failure detection and login
- **Rate limit** — a `429` / `rate_limit` marker on stderr, or a parsed
`{"type":"error"}` rate-limit event on stdout (conversation-text
mentions don't count), sets the `rate_limited` sentinel, parks for
`HIVE_RATE_LIMIT_SLEEP_SECS` (default 300), then retries. The UI shows
a `⊘ rate limited` badge while parked.
- **Auth failure (401)**`drive_turn` retries once (transient
token-refresh races clear on retry); a second `AuthFailed` writes
`{state_dir}/hyperhive-needs-login`, requeues the message, and parks in
`wait_for_login` — the same path as a cold boot with no session. The
operator re-auths via the per-agent web UI; the queued message then
drives the next turn.
- **Login detection** — both boot (`login::has_session`, Online vs
NeedsLogin) and `wait_for_login`'s resume check key off the credential
files in `login::CRED_FILE_NAMES` (the set `/logout` deletes).
`wait_for_login` resumes only when that set changes (a new file or a
newer mtime), so stale credentials on disk at the 401 don't trigger an
instant false-resume, and leftover session-history files don't read as a
live session after a logout + container recreate.
## Harness binary shape
One `hive` binary for all agents. The earlier split into
@ -95,7 +90,8 @@ loop (`serve_loop` / `handle_turn` / `wake`) has no per-role branches.
`HIVE_LABEL` (default `"hive"` for standalone runs; the meta
flake sets it unconditionally for any container-deployed agent;
see `docs/conventions.md::Hive identity` for the env stack),
opens turn-stats sqlite, prepares the on-boot files (see below),
opens turn-stats sqlite, prepares the on-boot files (see
[claude-invocation](turn-loop/claude-invocation.md#on-boot-files)),
installs claude plugins, spawns `forge_notify::run` + `web_ui::serve`,
and either drops into `serve_loop` directly (`Online`) or parks on
the login flow first (`NeedsLogin`).
@ -108,14 +104,16 @@ and the manager fall through to operator).
### Turn outcomes
`turn::TurnOutcome` drives the post-claude branch:
`turn::TurnOutcome` (`Result<bool, TurnError>``Ok(compacted)` on success,
else a `TurnError`) drives the post-claude branch:
| Outcome | Action |
| --- | --- |
| `Ok` / `Compacted` | `ack_turn` |
| `RateLimited` | sleep `HIVE_RATE_LIMIT_SLEEP_SECS` (default 300), requeue inflight, status back to `online` |
| `AuthFailed` | emit `needs_login_idle` sentinel, requeue inflight, park in `wait_for_login` |
| `Failed(err)` | route `[system] \`<qualified-label>\` claude turn failed:\n<err>` to `<parent>` via `send_to_parent` |
| `Ok(_)` (`false` normal / `true` compacted) | `ack_turn` |
| `Err(PromptTooLong)` | `drive_turn` archived the session (the lib already compacted + retried and it still overflowed); requeue inflight so the message redelivers into a fresh session that fits — no status park |
| `Err(RateLimited)` | sleep `HIVE_RATE_LIMIT_SLEEP_SECS` (default 300), requeue inflight, status back to `online` |
| `Err(AuthFailed)` | emit `needs_login_idle` sentinel, requeue inflight, park in `wait_for_login` |
| `Err(Failed(err))` | route `[system] \`<qualified-label>\` claude turn failed:\n<err>` to `<parent>` via `send_to_parent` |
After the outcome handler, the stats sink records a row and the
`hyperhive-continue` sentinel (dropped by the `request_next_turn`
@ -133,632 +131,21 @@ isn't needed (this is the `request_next_turn` contract — "no effect
if a new inbox message arrives before this turn ends"). The
`should_self_continue` predicate encodes exactly that decision.
## The claude invocation
```
claude --print --verbose --output-format stream-json --model <name> \
--effort <level> --resume <title> # or --name <title> on first use \
--system-prompt-file /run/hive/claude-system-prompt.md \
--mcp-config /run/hive/claude-mcp-config.json --strict-mcp-config \
--tools <builtins> --allowedTools <builtins+mcp>
# wake prompt piped over stdin
```
**Crate split.** The generic subprocess mechanics — spawning
`claude --print`, streaming + classifying stream-json, session
lookup/archive, and the durable-session compaction loop — live in the
reusable **`hive-claude`** crate (`hive_claude::{Claude, InfiniteSession,
Attach, CompactionPolicy, PercentPolicy, Telemetry, Sink, SessionStore}`;
see `hive-claude/README.md`). `hive_ag3nt::turn` is the hyperhive **policy
layer** on top: it builds the per-turn config from the bus, bridges the
output stream onto the event bus (`BusSink`), and owns the compaction /
auto-reset / retry decisions in `drive_turn`. The lib returns everything it
parsed from a turn (usage, cost, context window, resolved model) as
`Telemetry`, which the policy layer applies to the bus.
Hive-enforced settings ship at `/etc/claude-code/managed-settings.json`
(claude-code's canonical managed-settings path — precedence #1,
read-only, un-overridable), wired in `nix/templates/harness-base.nix`
from the `prompts/claude-settings.json` asset. `effortLevel` is
deliberately not in that file — effort is controlled live via the
`--effort` flag (`HIVE_DEFAULT_EFFORT` / the per-agent UI slider), which
managed scope would otherwise lock.
`<name>` is read from `Bus::model()` on each turn. The initial
default is set by `hyperhive.model` in the agent's `agent.nix`
(NixOS option; propagates via `HIVE_DEFAULT_MODEL` env var; falls
back to `"haiku"` if unset). The operator can flip it at runtime
with `/model <name>` in the web terminal — the next turn picks it
up. The choice is persisted to `/harness/hyperhive-model` so it
survives restart; override path: `HYPERHIVE_MODEL_FILE` env var
for tests.
Context-window size is looked up per-model via
`events::context_window_tokens(model)`. Resolution order (first
match wins):
1. `HIVE_CONTEXT_WINDOW_TOKENS_<KEY>` env var, where `KEY`
(lowercased) is a substring of the active model name. Injected
by the meta flake from `services.hyperhive.c0re.contextWindowTokens`
(host-level NixOS option, defaults: haiku=200k, sonnet=1M,
opus=1M). Override these for all agents at once without a
per-agent config change.
2. `HIVE_CONTEXT_WINDOW_TOKENS` — single global override for any
model (useful in dev / test).
3. Hard fallback: `200_000` (conservative; only reached outside
NixOS where the env vars aren't set).
The effective window drives watermarks and is exposed at runtime
via `/api/state.context_window_tokens` so the UI can show a
percentage-of-window ctx badge.
**Session identity — a constant title.** Every turn keys on one fixed,
harness-owned session title (`turn::session_title()`, default
`hive-session`, override `HIVE_SESSION_TITLE`). The durable
`hive_claude::InfiniteSession` (built once by the serve loop via
`turn::make_session`, then reused) `--resume <title>`s it; the *first* use
(bootstrap, post-archive, post-purge) misses and the session re-runs the
same prompt once with `--name <title>` to mint it. That single self-heal
rule is the whole identity system — there is **no** scraped session-id
file. Because the
title is constant, `/compact` and its post-compact retry provably target
the same session (killing the old "compact ran on a different/empty
session" bug), and a `choom` invocation in the same cwd can't hijack the
context (it won't carry our title). claude stores sessions in
`~/.claude/projects/<cwd-slug>/<uuid>.jsonl` (bind-mounted persistently);
`--name` writes the title into the file as a `custom-title` event, which
is what `--resume <title>` resolves against. We never pass bare
`--continue` (it resumes the *latest* session in the cwd — the hijack
vector). Auto-compact and auto-memory are disabled via the managed
settings at `/etc/claude-code/managed-settings.json` because hyperhive
owns compaction — see [Compaction](#compaction) below.
**Session reset** is available via `POST /api/new-session` (or
`/new-session` slash command). It does *not* touch the session inline —
that would race a mid-write claude process. Instead `Bus::request_session_reset()`
sets a one-shot flag consumed at the next turn boundary by `drive_turn`,
which **archives** the current session: the backing `<uuid>.jsonl` is
renamed to `<uuid>.jsonl.archived` (dropped out of claude's `*.jsonl`
resolution glob, history preserved on disk, only the file carrying *our*
title — any `choom` session sharing the cwd is left alone). The next
turn's `--resume <title>` then misses and self-heals into a fresh session.
### Compaction
claude's own in-session auto-compact is off (via the managed settings
at `/etc/claude-code/managed-settings.json`); hyperhive owns it. The
`hive_claude::InfiniteSession` keeps the session alive across the context
window with two triggers baked into its `run`:
- **Reactive** — claude-code prints `Prompt is too long`. The session is
*already* past the window, so no turn can run on it — the session
`/compact`s straight away and retries the same wake-up prompt once. No
notes-checkpoint turn is possible here: the detail is gone.
- **Proactive** — a turn finishes cleanly but the last inference's context
size crossed the policy watermark. While the session is still healthy it
runs one synthetic *notes-checkpoint* turn (`CHECKPOINT_PROMPT`
"context is filling up, flush durable state into `/state` now") and
*then* `/compact`s, so the agent can persist in-flight state before the
detail collapses into a summary.
The **when** is a `hive_claude::CompactionPolicy` injected by the harness:
`turn::make_session` builds a `PercentPolicy` that compacts once the
model-reported context fill reaches `HIVE_COMPACT_WATERMARK_PERCENT`
(default **75%**), falling back to `events::context_window_tokens(model)`
for the window on turns the model didn't report one. `0` disables proactive
compaction (the reactive path always applies). The proactive path is
best-effort — a failed checkpoint or `/compact` never fails the turn that
already succeeded.
The operator can force a compaction any time via `POST /api/compact`. It's
**deferred**: the handler sets `Bus::request_compact()` and returns
immediately; the harness runs the `/compact` at the next turn boundary
(end of the in-flight turn in `drive_turn`, or — when the agent is idle —
in `turn::run_pending_compact` on the serve loop's next empty poll). This
lets `/api/compact` work mid-turn instead of only when idle, without racing
a live claude process.
To disable proactive compaction for a specific agent, use the nix option:
```nix
hyperhive.autoCompact = false; # default true
```
Setting `autoCompact = false` sets `HIVE_COMPACT_WATERMARK_TOKENS=0`, which
the percent resolver still honours as a disable. Useful for large-context
models (sonnet/opus) where the 75% heuristic fires before the session is
actually full — the reactive path (compact-on-overflow at the hard limit)
still applies.
- **Auto session-reset** — a third path (`turn::maybe_auto_reset`,
pre-turn) that fires when both conditions hold: context is ≥ a watermark
(`HIVE_AUTO_RESET_WATERMARK_TOKENS`, default **50% of
`context_window_tokens(model)`**) AND the time since the last turn
exceeds the assumed prompt-cache TTL (`HIVE_CACHE_TTL_SECS`, default
`3600`). Claude's prompt cache goes cold after a while; once it's cold,
`--resume`-ing a large session pays the full re-upload cost with no
benefit over starting fresh. So `drive_turn` **archives** the current
session (same mechanism as the operator reset — rename `<uuid>.jsonl`
`.archived`) so the next turn's `--resume <title>` misses and starts
fresh. Unlike proactive compaction the session is dropped entirely, not
compacted — and *no* preceding checkpoint turn runs, because any turn
before the reset would just re-warm the cache and defeat the purpose.
Set `HIVE_AUTO_RESET_WATERMARK_TOKENS=0` to disable. Auto-reset and the
operator reset are mutually exclusive per turn (both archive → fresh
turn), so an explicit operator reset short-circuits the heuristic.
The child runs with `cwd = /state` (when the bind exists; falls
back to the parent's cwd in dev), so any relative path in a tool
call (`Read foo.md`, `Bash ls`, `Write notes.md`) lands in the
agent's durable bind-mounted dir. CLAUDE.md auto-load walks
upward from `/state` — drop a per-agent CLAUDE.md there if you
want long-term hints that survive destroy/recreate.
The wake prompt is intentionally minimal: the popped message's
`from`/`body`, prefixed with a `[msg #<id>]` broker-row-id marker
(so the agent knows what id to pass to `ack_until` for bulk triage;
transient pings show no marker because their sentinel id 0 has nothing
to ack), plus an inline `({unread} more pending — drain via …)` hint
when `unread > 0`. Claude drives any further `recv`/`send` itself via
the embedded MCP server.
Whenever hive-c0re starts / restarts / rebuilds a container, it
also drops a `system` message into the agent's inbox via
`Coordinator::kick_agent` — a one-line "you were just (re)started,
check /state/ for your notes, your session is intact". The
next turn picks it up like any other inbox message.
### On-boot files
`hive_ag3nt::turn::write_*` writes two files next to the per-agent
socket at `/run/hive/` once at startup:
- `claude-mcp-config.json` — by default re-invokes the running binary
as `mcp` stdio child (so the same binary serves as harness + MCP
server per turn). When `hyperhive.mcp.httpPort` is set in the
agent's NixOS config, the config instead points claude at the
persistent `hive-mcp-http` daemon (`http://127.0.0.1:{port}/mcp`)
— no stdio child per turn; trades the per-turn re-registration race
for a hard dependency on the daemon's uptime (`Restart=always`).
- `claude-system-prompt.md` — rendered from
`hive-ag3nt/prompts/system.md` by `hive_ag3nt::prompt::render`:
HTML-comment markers (`<!-- role:agent -->...<!-- /role:agent -->`,
same for `role:manager`) gate the role-specific blocks; everything
else is shared. Five placeholders are then
substituted: `{label}` (short agent name), `{qualified_label}`
(hive-qualified `name@domain` form), `{operator_pronouns}`,
`{hive_identity}` (e.g. `` on hive `pr1ma` ``; empty when
`hyperhive.hiveName` is unset), and `{swarm_identity}` (same
shape for the swarm). Pronouns come from `HIVE_OPERATOR_PRONOUNS`
env (set by the meta flake from
`services.hyperhive.c0re.operatorPronouns`, default `she/her`).
When `hyperhive.docs.enable` is set, `HIVE_DOCS_DIR` is present
in the environment and `render()` appends a one-sentence pointer
telling the agent the docs are mounted at that path (in lieu of
the old CLAUDE.md-in-docs-dir approach, which was dropped in
favour of this direct injection).
Passed via `--system-prompt-file`.
**Marker grammar.** `<!-- role:X -->` opens a block; matching
`<!-- /role:X -->` closes it. The renderer always uses role `agent`.
Blocks with other role tags are elided. Nesting is NOT supported —
a stray opener overrides until its closing tag (or end of file). A
mismatched closer is elided from the output but does NOT pop the
active role. Whitespace inside markers is tolerated
(`<!--role:foo-->` parses the same as `<!-- role:foo -->`).
Content outside any marker is always included.
**`hive_identity` / `swarm_identity` shape.** Each carries a
leading space + backticked name (` on hive \`pr1ma\``,
` in swarm \`constellat1on\``) when the corresponding env var
is set, otherwise empty string. The independence lets the
template drop one or both into the opener prose without
breaking single-hive deployments that never set the option;
the renderer also treats `Some("")` from a caller as `None` so
empty-string env vars and missing env vars round-trip the
same way.
The per-turn plumbing lives in `hive_ag3nt::turn`: `write_mcp_config` /
`write_system_prompt` (on-boot files), `make_session` (builds the durable
`InfiniteSession`, once), `drive_turn` (the policy state machine —
reset/auto-reset, the turn, 401-retry, deferred-compact-at-turn-end),
`run_pending_compact` (idle operator compact), `BusSink` (stream → bus +
`Telemetry` applied via `apply_telemetry`), `emit_turn_end`, `session_title`
/ `session_store` / `archive_session` (identity + turn-boundary reset). The
actual claude spawn, stream classification, and the reactive/proactive
compaction loop are in the `hive-claude` crate. Login-wait
(`wait_for_login`) lives in `hive_ag3nt::login`.
### Reference docs (`hyperhive.docs.enable`)
```nix
hyperhive.docs.enable = true; # default: false (true for the manager agent)
```
Makes the hyperhive `docs/` tree available inside the container at a nix
store path read from `$HIVE_DOCS_DIR`, and injects a single pointer
sentence into the agent's system prompt so it knows the docs exist and
where to find them. The tree is served by `claude --add-dir` so the full
markdown is readable during every turn.
Enabled by default only for the root/manager agent (`manager.nix`). Any
agent can opt in by adding the line above to its `agent.nix`.
The `docs/` source is a narrow flake input (`hyperhive-docs`) tracked
separately from the main `hyperhive` flake so editing docs re-locks only
that input — not every agent's container gets rebuilt on a doc-only change.
### Agent icon
```nix
hyperhive.icon = ./icon.svg; # default: null (falls back to shared hyperhive logo)
```
Path to an SVG file used as this agent's visual identity — shown in
the per-agent page header, as the page favicon, and uploaded to the
agent's Forgejo profile avatar (via the `forge-avatar-sync` boot
unit) and Matrix profile avatar (set by `hive-matrix-daemon` over its
live Client). Commit the SVG next to `agent.nix` in the config repo
and reference it as a relative path.
When `null` (the default), the agent falls back to the shared
hyperhive branding mark. The harness serves whichever icon is active
at `GET /icon` on the per-agent web port.
### `user.passwordlessSudo`
```nix
hyperhive.user.passwordlessSudo = true; # default
```
Grants the per-agent unix user passwordless `sudo` (`NOPASSWD: ALL`).
Enabled by default so claude's shell tools work for operations that
need root inside the container (`systemctl`, package managers in dev
shells, etc.) — the same privilege surface the previous root-user shape
had, now elevated explicitly rather than implicitly.
Set to `false` for agents that should be strictly unprivileged.
Any tool invocation that needs root then fails loudly with the standard
sudo rejection rather than silently succeeding — easier to audit.
`hyperhive.user.uid`, `hyperhive.user.gid`, and
`hyperhive.user.name` are the companion options; see
`docs/agent-hierarchy.md` — "Harness systemd unit shape" for the full
`user.*` surface.
### Dashboard links
```nix
hyperhive.dashboardLinks = [
{ label = "Stats"; icon = "📊"; url = "http://localhost:9001/stats"; }
{ label = "Scratchpad"; url = "http://localhost:8080"; }
];
```
Declares extra navigation links that appear on the agent's dashboard
card and in the per-agent page header alongside the built-in forge /
config / container links. Each entry has:
| Field | Required | Description |
|-------|----------|-------------|
| `label` | yes | Display text shown in the icon strip tooltip and meta-nav. |
| `url` | yes | Absolute URL — may include a different port (the dashboard renders it as a plain anchor). |
| `icon` | no | Emoji or short glyph prefix. Defaults to empty string. |
The list is written to `<state>/hyperhive-dashboard-links.json` by a
one-shot systemd unit at container boot. `hive-c0re` reads the file on
each container-view snapshot and attaches the links to the agent card
(`kind = External`) without any code change. Omitting the option
(default empty) produces no extra links.
### Custom static files
```nix
hyperhive.frontend.extraFiles = {
"games/bitburner" = {
source = ./bitburner-dist; # path relative to agent.nix
# target defaults to attribute name: "games/bitburner"
};
"my-page" = {
source = ./my-page.html;
target = "my-page.html"; # explicit override
};
};
```
Layers additional files over the default per-agent web UI dist. Each
attribute defines one overlay entry:
- **`source`** — a Nix path (file or directory) copied into the merged
static tree. Evaluated at nix build time; the resulting derivation is
pointed at by `HIVE_STATIC_DIR`.
- **`target`** — destination path within the merged tree, used as both
the served URL prefix (`/<target>/…`) and the on-disk layout.
Defaults to the attribute name. Forward slashes create nested layouts
(`"games/bitburner"` serves at `/games/bitburner/…`).
Constraints: `target` must start with an alphanumeric or `_` and
contain only alphanumerics, `_`, `.`, `/`, `-`. `..` segments are
rejected by a config assertion. The merge step refuses to overwrite
files already present in the default dist — pick a target name that
does not collide with existing paths (`static/`, `index.html`, etc.).
The default dist ships at `hyperhive.frontend.dist` (the
`hyperhive-frontend` package output, read-only). To replace the
entire UI rather than layer on top, override `frontend.dist` directly.
### Connectivity overrides
Two `hyperhive.forge.*` / `hyperhive.matrix.*` options override where
the per-agent daemons connect. Both rarely need changing on a standard
single-host deploy, but are useful for multi-hive or custom-network
setups.
```nix
hyperhive.forge.url = "http://localhost:3000"; # default
hyperhive.matrix.url = "http://localhost:8008"; # default
```
**`hyperhive.forge.url`** — base URL of the Forgejo instance. Used by
a one-shot boot unit (`tea-login`) that writes `~/.config/tea/config.yml`
directly from the agent's `forge-token`, so `tea` and `hive-forge`
work without an interactive auth step. The unit is a no-op when
`forge-token` is absent. Override when the agent should connect to a
Forgejo on a different host or port (e.g. a swarm peer's forge).
Validated: must be an `http://` or `https://` URL or the empty string.
**`hyperhive.matrix.url`** — homeserver URL used by
`hive-matrix-daemon` when connecting via the matrix-sdk. Default
(`localhost:8008`) is overridden by hive-c0re at deploy time to the
gateway-routed `matrix.<domain>` URL so isolated agents can reach the
homeserver. Override per-agent when an agent should talk to a
different homeserver — for example a remote hive's tuwunel reached
over a VPN, or an external Matrix server for a federation-only agent.
### Claude Code plugins
The harness installs Claude Code plugins before the serve loop opens.
Two per-agent `agent.nix` options control this:
```nix
hyperhive.claudeMarketplaces = [ "anthropics/claude-plugins-official" ]; # default
hyperhive.claudePlugins = [ "formatter@my-marketplace" ]; # default: []
hyperhive.claudePluginsAutoUpdate = false; # default
```
- **`claudeMarketplaces`** — list of marketplace sources passed to
`claude plugin marketplace add <source>`. The official Anthropic
marketplace is pre-configured by default; override or extend to add
custom marketplaces. Idempotent — re-adding an existing source is
a no-op.
- **`claudePlugins`** — list of plugin specs passed to
`claude plugin install <spec>`. Empty by default. Each spec is
installed on every boot (`install` is expected to be idempotent);
failures log a warning but do not abort boot.
- **`claudePluginsAutoUpdate`** — when `true`, runs
`claude plugin marketplace update` before installing plugins to pull
the latest index. Disabled by default to keep boot times short and
plugin versions pinned.
### `cargo.shortMessages`
```nix
hyperhive.cargo.shortMessages = true; # default
```
When enabled (the default), the harness injects a `cargo` shell
function into `/etc/hyperhive/bash-env.sh` that transparently appends
`--message-format short` to compile subcommands (`build`, `check`,
`clippy`, `test`, `run`, `doc`, `bench`, `install`, `rustc`, `fix`).
This suppresses the per-crate progress lines that flood the response
window, leaving only warnings and errors.
The function handles `+toolchain` selectors (`cargo +nightly build`)
and passes through cleanly when `--message-format` is already present.
Non-compile subcommands (`new`, `add`, third-party `cargo-*`) are
left untouched.
Set to `false` for agents that parse cargo's JSON output
programmatically and do not pass `--message-format json` themselves.
## MCP surface
The harness ships an embedded MCP server (rmcp 1.7). Claude launches
it as a stdio child via `--mcp-config`. The hyperhive socket name is
`hyperhive`, so the tools land in claude as `mcp__hyperhive__<tool>`.
Tool access is gated by tool groups (`HIVE_TOOL_GROUPS`). The default
preset (`AGENT_DEFAULT`) includes `messaging`, `meta`, `inbox`, and
`execution`. Privileged groups (`lifecycle`, `approvals`, `scheduling`,
`diagnostics`) are opt-in via the P3RM1SS10NS tab.
### Core tools (always available)
**Messaging** (`messaging` group): `send(to, body, in_reply_to?)`,
`recv(wait_seconds?, max?)`, `ask(question, options?, multi?,
ttl_seconds?, to?)`, `answer(id, answer)`, `ack_until(up_to)`.
- `send` — message a peer (logical name) or the operator
(`to: "operator"`). Use `to: "<parent>"` to address the topology
parent without hardcoding the label; the broker resolves the
sentinel at delivery time. Optional `in_reply_to: i64` links the
message to a prior id for thread rendering. Per-agent
`hyperhive.allowedRecipients` (default: empty = unrestricted) limits
which names `send` accepts — useful for sandboxing: set
`[ "operator" ]` to restrict a sub-agent to operator messages only
(the topology parent is always reachable regardless of this list —
that carve-out is structural, keyed on parent relationship, not name).
- `recv` — drain inbox. Without `wait_seconds` (or `0`) returns
immediately. Positive value parks the turn up to that many seconds
(cap 180) — incoming messages wake instantly. `max` (default 1, cap
5) drains up to N rows; `wait_seconds` applies to the first, then
drains up to `max` total. Each returned row is prefixed with
`[msg #<id>]` (broker row id; note the highest id seen, then pass
it to `ack_until` to bulk-triage the batch). **Graceful shutdown**: when the harness
receives a stop signal, the inbox becomes fenced and `recv` returns an
explicit `from: "graceful-stop"` message instead of an empty inbox.
This unmissably directs the agent to flush durable state (`/state`
files) and end the turn — the container exits when the turn completes.
The graceful-stop turn takes the same post-turn compaction path as any
other turn: if the context crossed the watermark the harness runs a
notes-checkpoint turn and then `/compact`. Compacting before shutdown
keeps a later cold start cheap instead of re-uploading a huge transcript.
- `ask` — surface a structured question to the operator (default) or
a peer agent (`to: "<agent>"`). Non-blocking — returns a question
id; the answer arrives as a `question_answered` system event in the
asker's inbox. `options` is advisory; `multi=true` renders as
checkboxes; `ttl_seconds` auto-cancels with answer `[expired]`.
- `answer` — respond to a `question_asked` event routed to this
agent. Strict authorisation: only the declared target can answer.
- `ack_until(up_to)` — bulk-mark inbox rows handled: every row with
broker id `<= up_to` is stamped as acked in a single UPDATE.
Recipient-scoped (agents can only ack their own rows). Use when a
restart redelivers a large backlog of already-handled messages: read
the highest `[msg #N]` from the set you've actually processed, then
`ack_until(N)` to prevent re-pop. Acked rows never redeliver.
Transient pings (sentinel id 0) have nothing to ack and show no marker.
**System messages** (from sender `system`): lifecycle and Q&A events
delivered as regular inbox messages (same `recv` path; body is a JSON
object with an `event` discriminant field). The **submitting agent**
(the root agent for top-level containers; an agent with the `approvals`
tool group for its own subtree) receives lifecycle events (`spawned`,
`rebuilt`, `killed`, `destroyed`, `container_crash`, `needs_login`,
`logged_in`, `config_ready`, `needs_update`, `approval_resolved`). Any
agent receives Q&A events when it is the declared target
(`question_asked`) or the asker (`question_answered`). Full payload
shapes and routing logic in
[`docs/approvals.md` § Helper events](../approvals.md#helper-events-to-the-submitting-agent).
**Inbox** (`inbox` group): `get_loose_ends(agent?)`,
`cancel_loose_end(kind, id)`, `remind(message, delay_seconds? |
at_unix_timestamp?)`, `request_next_turn()`.
- `get_loose_ends(agent?)` — list pending questions (asked/owed),
scheduled reminders, and active local tasks published by external MCP
daemons (e.g. running bash tasks from `hive-bash-mcp`). Each row
carries an id + kind for `cancel_loose_end`. Omit `agent` to list
your own threads. Pass `agent: "<name>"` to inspect a direct child
agent (always accessible per topology enforcement); non-children
require the `query_agent_state` capability. The `"*"` hive-wide
query is not available on the agent socket.
- `cancel_loose_end` — withdraw a `question` (posts `[cancelled by
<self>]`), hard-delete a `reminder`, or cancel a pending `approval`
row. Agents may only cancel rows they own; the `approval` kind is
further restricted to the root agent (`ruth`) server-side.
- `remind` — schedule a reminder in this agent's own inbox. Large
payloads spill to `/agents/<self>/state/reminders/`. Pending count
capped at 50 per agent (`HIVE_REMIND_MAX_PENDING_PER_AGENT`).
- `request_next_turn` — ask the harness to start another turn
immediately after this one ends, even if the inbox is empty.
Next turn fires with `from: "self"` and `body: "continue"`.
**Meta** (`meta` group): `set_status(text)`, `get_agent_meta(name?)`.
- `set_status` — set a free-text status string visible on the
dashboard. Single line, ≤ 200 chars. Persisted to
`{state_dir}/hyperhive-status`. Pass `""` to clear.
- `get_agent_meta` — fetch identity + status metadata for an agent:
`{ name, hyperhive_rev, running, status_text, status_set_at,
hive_name?, swarm_name?, matrix_accounts? }`. `matrix_accounts` is a
list of matrix identities the agent can act as (`name`, `user_id?`,
`homeserver`); omitted for agents with no matrix provisioning. Omit
`name` to query self.
### Privileged tools (by tool group)
- **Bash execution** (`execution`) — background shell tasks. See
[`docs/tools/bash.md`](tools/bash.md).
- **Lifecycle + config** (`lifecycle`, `approvals`) — manage child
agents, spawn new ones, apply config commits. See
[`docs/tools/lifecycle.md`](tools/lifecycle.md).
- **Scheduling + diagnostics** (`scheduling`, `diagnostics`) —
scheduled prompts, `get_logs`. See
[`docs/tools/scheduling.md`](tools/scheduling.md).
- **Forge repos** (`forge`) — `create_repo` — the only agent path to
create a repo under the `agents/` org (direct forge token creation is
disabled for agents). The repo is created in the c0re-owned `agents`
org; the calling agent gets write collaborator access; the default
branch is branch-protected (operator-team must approve merges, so the
agent cannot self-merge). Opt-in; not in any default preset.
See [`docs/tools/forge.md — Repo management`](tools/forge.md).
- **Web egress** (`web_tools`) — enables Claude's built-in `WebFetch`
and `WebSearch` tools (not MCP tools; added directly to the
`--allowedTools` list). Off by default; add the group in the
P3RM1SS10NS tab and rebuild to enable.
- **Capability-gated**`get_host_journal` (requires
`read_host_journal` capability set via the P3RM1SS10NS tab;
orthogonal to tool groups). Full list of capabilities and their
effects in [`docs/conventions.md#capabilities`](../docs/conventions.md).
Also documented in [`docs/tools/scheduling.md`](tools/scheduling.md).
- **Matrix MCP + extra servers**`mcp__matrix__*` tools and
per-agent extra MCP config. See
[`docs/tools/matrix.md`](tools/matrix.md).
### Waking the agent from inside the container
External MCP servers (and any other in-container process) can
inject a wake-up event into the agent's inbox via the per-agent
socket at `/run/hive/mcp.sock`. Two equivalent paths:
- **Shell out to `hive wake --from <label> --body <text>`**
(use `--body -` to read body from stdin). Already on the
container's `PATH` since the harness binary is in
`systemPackages`. Convenient for shell-script integrations and
co-process daemons (matrix bridge, webhook listeners, scrapers).
- **Speak the wire protocol directly** — JSON-line over the
unix socket: `{"cmd":"wake","from":"matrix","body":"new dm
from @alice"}\n`. Same shape as any other `AgentRequest`;
see `hive-sh4re::AgentRequest::Wake`.
The wake event lands in the broker as `{from:<label>,
to:<agent>, body}`, waking whatever `recv` call the harness
is currently blocked on. The next turn fires with the wake
prompt formed from that message.
Identity = socket: anything that can connect to
`/run/hive/mcp.sock` is implicitly trusted to inject these —
the bind-mount is the agent's own container only.
### Authoritative state
`hive_ag3nt::events::Bus` carries the current turn-loop state in
addition to the broadcast channel and the events history. Variants:
- `Idle` — sitting on `Recv` waiting for mail.
- `Thinking``claude --print` is running for a turn.
- `Compacting` — operator-triggered `/compact` is in flight.
The harness flips state at the relevant transitions
(`set_state(Thinking)` before `drive_turn`, `set_state(Idle)`
after; `set_state(Compacting)` around an idle operator compact in
`turn::run_pending_compact`). Exposed via `/api/state.turn_state` +
`turn_state_since` (unix seconds); the agent page renders this rather
than deriving from SSE events.
### Tool envelope
`mcp::run_tool_envelope`: every MCP tool handler logs the request,
runs the body, logs the result. Pre-/post-log only — the inbox
status hint moved to the wake prompt + UI header.
### Tool whitelist (`mcp::ALLOWED_BUILTIN_TOOLS`)
- Allowed built-ins: `Edit`, `Glob`, `Grep`, `Read`, `Write`.
- Tool-group-gated built-ins: `WebFetch`, `WebSearch` (added when the
`web_tools` tool group is enabled — see P3RM1SS10NS tab).
- Denied by omission or the managed-settings deny list
(`/etc/claude-code/managed-settings.json`): `Bash`, `Task`,
`NotebookEdit`, `TodoWrite`.
- Allowed MCP tools: as listed above (by tool group).
`Bash` is disallowed — shell execution goes through
`mcp__bash__run` (background tasks with structured output +
task-id tracking) instead of an interactive shell. The bash MCP server
(`run` / `status` / `kill`) uses `allowedTools = ["*"]` so all
`mcp__bash__*` tools are always available regardless of tool groups.
`WebFetch` / `WebSearch` are off by default; enable the `web_tools`
tool group in the P3RM1SS10NS tab and rebuild the agent to enable them.
## Sub-pages
The rest lives in three topic pages under [`turn-loop/`](turn-loop/):
- **[claude-invocation.md](turn-loop/claude-invocation.md)** — how the harness
spawns `claude --print` each turn, the two-pronged compaction (reactive +
proactive), and the on-boot files it materialises (`--mcp-config`,
`--system-prompt-file`).
- **[config.md](turn-loop/config.md)** — the optional per-agent knobs the meta
flake wires in (reference docs, icon, passwordless sudo, dashboard links,
custom static files, connectivity overrides, claude plugins, cargo message
filtering).
- **[mcp.md](turn-loop/mcp.md)** — the MCP tool surface claude sees: core tools,
privileged tool groups, self-wake, authoritative state, the tool envelope,
and the built-in tool whitelist.
Per-subsystem impl detail lives in each module's `//!` doc-comment; these pages
describe present-state behaviour + wiring, not line-level mechanics.

View file

@ -0,0 +1,241 @@
# The claude invocation
```
claude --print --verbose --output-format stream-json --model <name> \
--effort <level> --resume <title> # or --name <title> on first use \
--system-prompt-file /run/hive/claude-system-prompt.md \
--mcp-config /run/hive/claude-mcp-config.json --strict-mcp-config \
--tools <builtins> --allowedTools <builtins+mcp>
# wake prompt piped over stdin
```
**Crate split.** The generic subprocess mechanics — spawning
`claude --print`, streaming + classifying stream-json, session
lookup/archive, and the durable-session compaction loop — live in the
reusable **`hive-claude`** crate (`hive_claude::{Claude, InfiniteSession,
Attach, CompactionPolicy, PercentPolicy, Telemetry, Sink, SessionStore}`;
see `hive-claude/README.md`). `hive_ag3nt::turn` is the hyperhive **policy
layer** on top: it builds the per-turn config from the bus, bridges the
output stream onto the event bus (`BusSink`), and owns the compaction /
auto-reset / retry decisions in `drive_turn`. The lib returns everything it
parsed from a turn (usage, cost, context window, resolved model) as
`Telemetry`, which the policy layer applies to the bus.
Hive-enforced settings ship at `/etc/claude-code/managed-settings.json`
(claude-code's canonical managed-settings path — precedence #1,
read-only, un-overridable), wired in `nix/templates/harness-base.nix`
from the `prompts/claude-settings.json` asset. `effortLevel` is
deliberately not in that file — effort is controlled live via the
`--effort` flag (`HIVE_DEFAULT_EFFORT` / the per-agent UI slider), which
managed scope would otherwise lock.
`<name>` is read from `Bus::model()` on each turn. The initial
default is set by `hyperhive.model` in the agent's `agent.nix`
(NixOS option; propagates via `HIVE_DEFAULT_MODEL` env var; falls
back to `"haiku"` if unset). The operator can flip it at runtime
with `/model <name>` in the web terminal — the next turn picks it
up. The choice is persisted to `/harness/hyperhive-model` so it
survives restart; override path: `HYPERHIVE_MODEL_FILE` env var
for tests.
Context-window size is looked up per-model via
`events::context_window_tokens(model)`. Resolution order (first
match wins):
1. `HIVE_CONTEXT_WINDOW_TOKENS_<KEY>` env var, where `KEY`
(lowercased) is a substring of the active model name. Injected
by the meta flake from `services.hyperhive.c0re.contextWindowTokens`
(host-level NixOS option, defaults: haiku=200k, sonnet=1M,
opus=1M). Override these for all agents at once without a
per-agent config change.
2. `HIVE_CONTEXT_WINDOW_TOKENS` — single global override for any
model (useful in dev / test).
3. Hard fallback: `200_000` (conservative; only reached outside
NixOS where the env vars aren't set).
The effective window drives watermarks and is exposed at runtime
via `/api/state.context_window_tokens` so the UI can show a
percentage-of-window ctx badge.
**Session identity — a constant title.** Every turn keys on one fixed,
harness-owned session title (`turn::session_title()`, default
`hive-session`, override `HIVE_SESSION_TITLE`). The durable
`hive_claude::InfiniteSession` (built once by the serve loop via
`turn::make_session`, then reused) `--resume <title>`s it; the *first* use
(bootstrap, post-archive, post-purge) misses and the session re-runs the
same prompt once with `--name <title>` to mint it. That single self-heal
rule is the whole identity system — there is **no** scraped session-id
file. Because the
title is constant, `/compact` and its post-compact retry provably target
the same session (killing the old "compact ran on a different/empty
session" bug), and a `choom` invocation in the same cwd can't hijack the
context (it won't carry our title). claude stores sessions in
`~/.claude/projects/<cwd-slug>/<uuid>.jsonl` (bind-mounted persistently);
`--name` writes the title into the file as a `custom-title` event, which
is what `--resume <title>` resolves against. We never pass bare
`--continue` (it resumes the *latest* session in the cwd — the hijack
vector). Auto-compact and auto-memory are disabled via the managed
settings at `/etc/claude-code/managed-settings.json` because hyperhive
owns compaction — see [Compaction](#compaction) below.
**Session reset** is available via `POST /api/new-session` (or
`/new-session` slash command). It does *not* touch the session inline —
that would race a mid-write claude process. Instead `Bus::request_session_reset()`
sets a one-shot flag consumed at the next turn boundary by `drive_turn`,
which **archives** the current session: the backing `<uuid>.jsonl` is
renamed to `<uuid>.jsonl.archived` (dropped out of claude's `*.jsonl`
resolution glob, history preserved on disk, only the file carrying *our*
title — any `choom` session sharing the cwd is left alone). The next
turn's `--resume <title>` then misses and self-heals into a fresh session.
## Compaction
claude's own in-session auto-compact is off (via the managed settings
at `/etc/claude-code/managed-settings.json`); hyperhive owns it. The
`hive_claude::InfiniteSession` keeps the session alive across the context
window with two triggers baked into its `run`:
- **Reactive** — claude-code prints `Prompt is too long`. The session is
*already* past the window, so no turn can run on it — the session
`/compact`s straight away and retries the same wake-up prompt once. No
notes-checkpoint turn is possible here: the detail is gone. If the retry
*still* overflows, `run` surfaces `Error::PromptTooLong`; `drive_turn`
then archives the session (session lifecycle stays hive-side) and the
serve loop requeues the message so it redelivers into a fresh session
(see [Turn outcomes](../turn-loop.md#turn-outcomes) — the wake prompt itself is tiny, so
the overflow was the accumulated context the archive clears).
- **Proactive** — a turn finishes cleanly but the last inference's context
size crossed the policy watermark. While the session is still healthy it
runs one synthetic *notes-checkpoint* turn (`CHECKPOINT_PROMPT`
"context is filling up, flush durable state into `/state` now") and
*then* `/compact`s, so the agent can persist in-flight state before the
detail collapses into a summary.
The **when** is a `hive_claude::CompactionPolicy` injected by the harness:
`turn::make_session` builds a `PercentPolicy` that compacts once the
model-reported context fill reaches `HIVE_COMPACT_WATERMARK_PERCENT`
(default **75%**), falling back to `events::context_window_tokens(model)`
for the window on turns the model didn't report one. `0` disables proactive
compaction (the reactive path always applies). The proactive path is
best-effort — a failed checkpoint or `/compact` never fails the turn that
already succeeded.
The operator can force a compaction any time via `POST /api/compact`. It's
**deferred**: the handler sets `Bus::request_compact()` and returns
immediately; the harness runs the `/compact` at the next turn boundary
(end of the in-flight turn in `drive_turn`, or — when the agent is idle —
in `turn::run_pending_compact` on the serve loop's next empty poll). This
lets `/api/compact` work mid-turn instead of only when idle, without racing
a live claude process.
To disable proactive compaction for a specific agent, use the nix option:
```nix
hyperhive.autoCompact = false; # default true
```
Setting `autoCompact = false` sets `HIVE_COMPACT_WATERMARK_TOKENS=0`, which
the percent resolver still honours as a disable. Useful for large-context
models (sonnet/opus) where the 75% heuristic fires before the session is
actually full — the reactive path (compact-on-overflow at the hard limit)
still applies.
- **Auto session-reset** — a third path (`turn::maybe_auto_reset`,
pre-turn) that fires when both conditions hold: context is ≥ a watermark
(`HIVE_AUTO_RESET_WATERMARK_TOKENS`, default **50% of
`context_window_tokens(model)`**) AND the time since the last turn
exceeds the assumed prompt-cache TTL (`HIVE_CACHE_TTL_SECS`, default
`3600`). Claude's prompt cache goes cold after a while; once it's cold,
`--resume`-ing a large session pays the full re-upload cost with no
benefit over starting fresh. So `drive_turn` **archives** the current
session (same mechanism as the operator reset — rename `<uuid>.jsonl`
`.archived`) so the next turn's `--resume <title>` misses and starts
fresh. Unlike proactive compaction the session is dropped entirely, not
compacted — and *no* preceding checkpoint turn runs, because any turn
before the reset would just re-warm the cache and defeat the purpose.
Set `HIVE_AUTO_RESET_WATERMARK_TOKENS=0` to disable. Auto-reset and the
operator reset are mutually exclusive per turn (both archive → fresh
turn), so an explicit operator reset short-circuits the heuristic.
The child runs with `cwd = /state` (when the bind exists; falls
back to the parent's cwd in dev), so any relative path in a tool
call (`Read foo.md`, `Bash ls`, `Write notes.md`) lands in the
agent's durable bind-mounted dir. CLAUDE.md auto-load walks
upward from `/state` — drop a per-agent CLAUDE.md there if you
want long-term hints that survive destroy/recreate.
The wake prompt is intentionally minimal: the popped message's
`from`/`body`, prefixed with a `[msg #<id>]` broker-row-id marker
(so the agent knows what id to pass to `ack_until` for bulk triage;
transient pings show no marker because their sentinel id 0 has nothing
to ack), plus an inline `({unread} more pending — drain via …)` hint
when `unread > 0`. Claude drives any further `recv`/`send` itself via
the embedded MCP server.
Whenever hive-c0re starts / restarts / rebuilds a container, it
also drops a `system` message into the agent's inbox via
`Coordinator::kick_agent` — a one-line "you were just (re)started,
check /state/ for your notes, your session is intact". The
next turn picks it up like any other inbox message.
## On-boot files
`hive_ag3nt::turn::write_*` writes two files next to the per-agent
socket at `/run/hive/` once at startup:
- `claude-mcp-config.json` — by default re-invokes the running binary
as `mcp` stdio child (so the same binary serves as harness + MCP
server per turn). When `hyperhive.mcp.httpPort` is set in the
agent's NixOS config, the config instead points claude at the
persistent `hive-mcp-http` daemon (`http://127.0.0.1:{port}/mcp`)
— no stdio child per turn; trades the per-turn re-registration race
for a hard dependency on the daemon's uptime (`Restart=always`).
- `claude-system-prompt.md` — rendered from
`hive-ag3nt/prompts/system.md` by `hive_ag3nt::prompt::render`:
HTML-comment markers (`<!-- role:agent -->...<!-- /role:agent -->`,
same for `role:manager`) gate the role-specific blocks; everything
else is shared. Five placeholders are then
substituted: `{label}` (short agent name), `{qualified_label}`
(hive-qualified `name@domain` form), `{operator_pronouns}`,
`{hive_identity}` (e.g. `` on hive `pr1ma` ``; empty when
`hyperhive.hiveName` is unset), and `{swarm_identity}` (same
shape for the swarm). Pronouns come from `HIVE_OPERATOR_PRONOUNS`
env (set by the meta flake from
`services.hyperhive.c0re.operatorPronouns`, default `she/her`).
When `hyperhive.docs.enable` is set, `HIVE_DOCS_DIR` is present
in the environment and `render()` appends a one-sentence pointer
telling the agent the docs are mounted at that path (in lieu of
the old CLAUDE.md-in-docs-dir approach, which was dropped in
favour of this direct injection).
Passed via `--system-prompt-file`.
**Marker grammar.** `<!-- role:X -->` opens a block; any
`<!-- /role:X -->` closes the current block. The renderer always uses
role `agent`, so blocks with other role tags are elided. Nesting is NOT
supported — a stray opener with no closer runs until end of file.
Whitespace inside markers is tolerated (`<!--role:foo-->` parses the
same as `<!-- role:foo -->`). Content outside any marker is always
included. Today's `system.md` carries no markers (single agent role) —
the grammar stays wired for a future manager / multi-role prompt.
**`hive_identity` / `swarm_identity` shape.** Each carries a
leading space + backticked name (` on hive \`pr1ma\``,
` in swarm \`constellat1on\``) when the corresponding env var
is set, otherwise empty string. The independence lets the
template drop one or both into the opener prose without
breaking single-hive deployments that never set the option;
the renderer also treats `Some("")` from a caller as `None` so
empty-string env vars and missing env vars round-trip the
same way.
The per-turn plumbing lives in `hive_ag3nt::turn`: `write_mcp_config` /
`write_system_prompt` (on-boot files), `make_session` (builds the durable
`InfiniteSession`, once), `drive_turn` (the policy state machine —
reset/auto-reset, the turn, 401-retry, deferred-compact-at-turn-end),
`run_pending_compact` (idle operator compact), `BusSink` (stream → bus +
`Telemetry` applied via `apply_telemetry`), `emit_turn_end`, `session_title`
/ `session_store` / `archive_session` (identity + turn-boundary reset). The
actual claude spawn, stream classification, and the reactive/proactive
compaction loop are in the `hive-claude` crate. Login-wait
(`wait_for_login`) lives in `hive_ag3nt::login`.

199
docs/turn-loop/config.md Normal file
View file

@ -0,0 +1,199 @@
# Agent config knobs
Optional per-agent knobs the meta flake wires into the container from
`services.hyperhive.agents.<name>`, read at boot or per turn by the harness.
Absent means the default. (The claude spawn + compaction themselves live in
[claude-invocation](claude-invocation.md).)
## Reference docs (`hyperhive.docs.enable`)
```nix
hyperhive.docs.enable = true; # default: false (true for the manager agent)
```
Makes the hyperhive `docs/` tree available inside the container at a nix
store path read from `$HIVE_DOCS_DIR`, and injects a single pointer
sentence into the agent's system prompt so it knows the docs exist and
where to find them. The tree is served by `claude --add-dir` so the full
markdown is readable during every turn.
Enabled by default only for the root/manager agent (`manager.nix`). Any
agent can opt in by adding the line above to its `agent.nix`.
The `docs/` source is a narrow flake input (`hyperhive-docs`) tracked
separately from the main `hyperhive` flake so editing docs re-locks only
that input — not every agent's container gets rebuilt on a doc-only change.
## Agent icon
```nix
hyperhive.icon = ./icon.svg; # default: null (falls back to shared hyperhive logo)
```
Path to an SVG file used as this agent's visual identity — shown in
the per-agent page header, as the page favicon, and uploaded to the
agent's Forgejo profile avatar (via the `forge-avatar-sync` boot
unit) and Matrix profile avatar (set by `hive-matrix-daemon` over its
live Client). Commit the SVG next to `agent.nix` in the config repo
and reference it as a relative path.
When `null` (the default), the agent falls back to the shared
hyperhive branding mark. The harness serves whichever icon is active
at `GET /icon` on the per-agent web port.
## `user.passwordlessSudo`
```nix
hyperhive.user.passwordlessSudo = true; # default
```
Grants the per-agent unix user passwordless `sudo` (`NOPASSWD: ALL`).
Enabled by default so claude's shell tools work for operations that
need root inside the container (`systemctl`, package managers in dev
shells, etc.) — the same privilege surface the previous root-user shape
had, now elevated explicitly rather than implicitly.
Set to `false` for agents that should be strictly unprivileged.
Any tool invocation that needs root then fails loudly with the standard
sudo rejection rather than silently succeeding — easier to audit.
`hyperhive.user.uid`, `hyperhive.user.gid`, and
`hyperhive.user.name` are the companion options; see
`docs/agent-hierarchy.md` — "Harness systemd unit shape" for the full
`user.*` surface.
## Dashboard links
```nix
hyperhive.dashboardLinks = [
{ label = "Stats"; icon = "📊"; url = "http://localhost:9001/stats"; }
{ label = "Scratchpad"; url = "http://localhost:8080"; }
];
```
Declares extra navigation links that appear on the agent's dashboard
card and in the per-agent page header alongside the built-in forge /
config / container links. Each entry has:
| Field | Required | Description |
|-------|----------|-------------|
| `label` | yes | Display text shown in the icon strip tooltip and meta-nav. |
| `url` | yes | Absolute URL — may include a different port (the dashboard renders it as a plain anchor). |
| `icon` | no | Emoji or short glyph prefix. Defaults to empty string. |
The list is written to `<state>/hyperhive-dashboard-links.json` by a
one-shot systemd unit at container boot. `hive-c0re` reads the file on
each container-view snapshot and attaches the links to the agent card
(`kind = External`) without any code change. Omitting the option
(default empty) produces no extra links.
## Custom static files
```nix
hyperhive.frontend.extraFiles = {
"games/bitburner" = {
source = ./bitburner-dist; # path relative to agent.nix
# target defaults to attribute name: "games/bitburner"
};
"my-page" = {
source = ./my-page.html;
target = "my-page.html"; # explicit override
};
};
```
Layers additional files over the default per-agent web UI dist. Each
attribute defines one overlay entry:
- **`source`** — a Nix path (file or directory) copied into the merged
static tree. Evaluated at nix build time; the resulting derivation is
pointed at by `HIVE_STATIC_DIR`.
- **`target`** — destination path within the merged tree, used as both
the served URL prefix (`/<target>/…`) and the on-disk layout.
Defaults to the attribute name. Forward slashes create nested layouts
(`"games/bitburner"` serves at `/games/bitburner/…`).
Constraints: `target` must start with an alphanumeric or `_` and
contain only alphanumerics, `_`, `.`, `/`, `-`. `..` segments are
rejected by a config assertion. The merge step refuses to overwrite
files already present in the default dist — pick a target name that
does not collide with existing paths (`static/`, `index.html`, etc.).
The default dist ships at `hyperhive.frontend.dist` (the
`hyperhive-frontend` package output, read-only). To replace the
entire UI rather than layer on top, override `frontend.dist` directly.
## Connectivity overrides
Two `hyperhive.forge.*` / `hyperhive.matrix.*` options override where
the per-agent daemons connect. Both rarely need changing on a standard
single-host deploy, but are useful for multi-hive or custom-network
setups.
```nix
hyperhive.forge.url = "http://localhost:3000"; # default
hyperhive.matrix.url = "http://localhost:8008"; # default
```
**`hyperhive.forge.url`** — base URL of the Forgejo instance. Used by
a one-shot boot unit (`tea-login`) that writes `~/.config/tea/config.yml`
directly from the agent's `forge-token`, so `tea` and `hive-forge`
work without an interactive auth step. The unit is a no-op when
`forge-token` is absent. Override when the agent should connect to a
Forgejo on a different host or port (e.g. a swarm peer's forge).
Validated: must be an `http://` or `https://` URL or the empty string.
**`hyperhive.matrix.url`** — homeserver URL used by
`hive-matrix-daemon` when connecting via the matrix-sdk. Default
(`localhost:8008`) is overridden by hive-c0re at deploy time to the
gateway-routed `matrix.<domain>` URL so isolated agents can reach the
homeserver. Override per-agent when an agent should talk to a
different homeserver — for example a remote hive's tuwunel reached
over a VPN, or an external Matrix server for a federation-only agent.
## Claude Code plugins
The harness installs Claude Code plugins before the serve loop opens.
Two per-agent `agent.nix` options control this:
```nix
hyperhive.claudeMarketplaces = [ "anthropics/claude-plugins-official" ]; # default
hyperhive.claudePlugins = [ "formatter@my-marketplace" ]; # default: []
hyperhive.claudePluginsAutoUpdate = false; # default
```
- **`claudeMarketplaces`** — list of marketplace sources passed to
`claude plugin marketplace add <source>`. The official Anthropic
marketplace is pre-configured by default; override or extend to add
custom marketplaces. Idempotent — re-adding an existing source is
a no-op.
- **`claudePlugins`** — list of plugin specs passed to
`claude plugin install <spec>`. Empty by default. Each spec is
installed on every boot (`install` is expected to be idempotent);
failures log a warning but do not abort boot.
- **`claudePluginsAutoUpdate`** — when `true`, runs
`claude plugin marketplace update` before installing plugins to pull
the latest index. Disabled by default to keep boot times short and
plugin versions pinned.
## `cargo.shortMessages`
```nix
hyperhive.cargo.shortMessages = true; # default
```
When enabled (the default), the harness injects a `cargo` shell
function into `/etc/hyperhive/bash-env.sh` that transparently appends
`--message-format short` to compile subcommands (`build`, `check`,
`clippy`, `test`, `run`, `doc`, `bench`, `install`, `rustc`, `fix`).
This suppresses the per-crate progress lines that flood the response
window, leaving only warnings and errors.
The function handles `+toolchain` selectors (`cargo +nightly build`)
and passes through cleanly when `--message-format` is already present.
Non-compile subcommands (`new`, `add`, third-party `cargo-*`) are
left untouched.
Set to `false` for agents that parse cargo's JSON output
programmatically and do not pass `--message-format json` themselves.

201
docs/turn-loop/mcp.md Normal file
View file

@ -0,0 +1,201 @@
# MCP surface
The harness ships an embedded MCP server (rmcp 1.7). Claude launches
it as a stdio child via `--mcp-config`. The hyperhive socket name is
`hyperhive`, so the tools land in claude as `mcp__hyperhive__<tool>`.
Tool access is gated by tool groups (`HIVE_TOOL_GROUPS`). The default
preset (`AGENT_DEFAULT`) includes `messaging`, `meta`, `inbox`, and
`execution`. Privileged groups (`lifecycle`, `approvals`, `scheduling`,
`diagnostics`) are opt-in via the P3RM1SS10NS tab.
## Core tools (always available)
**Messaging** (`messaging` group): `send(to, body, in_reply_to?)`,
`recv(wait_seconds?, max?)`, `ask(question, options?, multi?,
ttl_seconds?, to?)`, `answer(id, answer)`, `ack_until(up_to)`.
- `send` — message a peer (logical name) or the operator
(`to: "operator"`). Use `to: "<parent>"` to address the topology
parent without hardcoding the label; the broker resolves the
sentinel at delivery time. Optional `in_reply_to: i64` links the
message to a prior id for thread rendering. Per-agent
`hyperhive.allowedRecipients` (default: empty = unrestricted) limits
which names `send` accepts — useful for sandboxing: set
`[ "operator" ]` to restrict a sub-agent to operator messages only
(the topology parent is always reachable regardless of this list —
that carve-out is structural, keyed on parent relationship, not name).
- `recv` — drain inbox. Without `wait_seconds` (or `0`) returns
immediately. Positive value parks the turn up to that many seconds
(cap 180) — incoming messages wake instantly. `max` (default 1, cap
5) drains up to N rows; `wait_seconds` applies to the first, then
drains up to `max` total. Each returned row is prefixed with
`[msg #<id>]` (broker row id; note the highest id seen, then pass
it to `ack_until` to bulk-triage the batch). **Graceful shutdown**: when the harness
receives a stop signal, the inbox becomes fenced and `recv` returns an
explicit `from: "graceful-stop"` message instead of an empty inbox.
This unmissably directs the agent to flush durable state (`/state`
files) and end the turn — the container exits when the turn completes.
The graceful-stop turn takes the same post-turn compaction path as any
other turn: if the context crossed the watermark the harness runs a
notes-checkpoint turn and then `/compact`. Compacting before shutdown
keeps a later cold start cheap instead of re-uploading a huge transcript.
- `ask` — surface a structured question to the operator (default) or
a peer agent (`to: "<agent>"`). Non-blocking — returns a question
id; the answer arrives as a `question_answered` system event in the
asker's inbox. `options` is advisory; `multi=true` renders as
checkboxes; `ttl_seconds` auto-cancels with answer `[expired]`.
- `answer` — respond to a `question_asked` event routed to this
agent. Strict authorisation: only the declared target can answer.
- `ack_until(up_to)` — bulk-mark inbox rows handled: every row with
broker id `<= up_to` is stamped as acked in a single UPDATE.
Recipient-scoped (agents can only ack their own rows). Use when a
restart redelivers a large backlog of already-handled messages: read
the highest `[msg #N]` from the set you've actually processed, then
`ack_until(N)` to prevent re-pop. Acked rows never redeliver.
Transient pings (sentinel id 0) have nothing to ack and show no marker.
**System messages** (from sender `system`): lifecycle and Q&A events
delivered as regular inbox messages (same `recv` path; body is a JSON
object with an `event` discriminant field). The **submitting agent**
(the root agent for top-level containers; an agent with the `approvals`
tool group for its own subtree) receives lifecycle events (`spawned`,
`rebuilt`, `killed`, `destroyed`, `container_crash`, `needs_login`,
`logged_in`, `config_ready`, `needs_update`, `approval_resolved`). Any
agent receives Q&A events when it is the declared target
(`question_asked`) or the asker (`question_answered`). Full payload
shapes and routing logic in
[`docs/approvals.md` § Helper events](../approvals.md#helper-events-to-the-submitting-agent).
**Inbox** (`inbox` group): `get_loose_ends(agent?)`,
`cancel_loose_end(kind, id)`, `remind(message, delay_seconds? |
at_unix_timestamp?)`, `request_next_turn()`.
- `get_loose_ends(agent?)` — list pending questions (asked/owed),
scheduled reminders, and active local tasks published by external MCP
daemons (e.g. running bash tasks from `hive-bash-mcp`). Each row
carries an id + kind for `cancel_loose_end`. Omit `agent` to list
your own threads. Pass `agent: "<name>"` to inspect a direct child
agent (always accessible per topology enforcement); non-children
require the `query_agent_state` capability. The `"*"` hive-wide
query is not available on the agent socket.
- `cancel_loose_end` — withdraw a `question` (posts `[cancelled by
<self>]`), hard-delete a `reminder`, or cancel a pending `approval`
row. Agents may only cancel rows they own; the `approval` kind is
further restricted to the root agent (`ruth`) server-side.
- `remind` — schedule a reminder in this agent's own inbox. Large
payloads spill to `/agents/<self>/state/reminders/`. Pending count
capped at 50 per agent (`HIVE_REMIND_MAX_PENDING_PER_AGENT`).
- `request_next_turn` — ask the harness to start another turn
immediately after this one ends, even if the inbox is empty.
Next turn fires with `from: "self"` and `body: "continue"`.
**Meta** (`meta` group): `set_status(text)`, `get_agent_meta(name?)`.
- `set_status` — set a free-text status string visible on the
dashboard. Single line, ≤ 200 chars. Persisted to
`{state_dir}/hyperhive-status`. Pass `""` to clear.
- `get_agent_meta` — fetch identity + status metadata for an agent:
`{ name, hyperhive_rev, running, status_text, status_set_at,
hive_name?, swarm_name?, matrix_accounts? }`. `matrix_accounts` is a
list of matrix identities the agent can act as (`name`, `user_id?`,
`homeserver`); omitted for agents with no matrix provisioning. Omit
`name` to query self.
## Privileged tools (by tool group)
- **Bash execution** (`execution`) — background shell tasks. See
[`docs/tools/bash.md`](../tools/bash.md).
- **Lifecycle + config** (`lifecycle`, `approvals`) — manage child
agents, spawn new ones, apply config commits. See
[`docs/tools/lifecycle.md`](../tools/lifecycle.md).
- **Scheduling + diagnostics** (`scheduling`, `diagnostics`) —
scheduled prompts, `get_logs`. See
[`docs/tools/scheduling.md`](../tools/scheduling.md).
- **Forge repos** (`forge`) — `create_repo` — the only agent path to
create a repo under the `agents/` org (direct forge token creation is
disabled for agents). The repo is created in the c0re-owned `agents`
org; the calling agent gets write collaborator access; the default
branch is branch-protected (operator-team must approve merges, so the
agent cannot self-merge). Opt-in; not in any default preset.
See [`docs/tools/forge.md — Repo management`](../tools/forge.md).
- **Web egress** (`web_tools`) — enables Claude's built-in `WebFetch`
and `WebSearch` tools (not MCP tools; added directly to the
`--allowedTools` list). Off by default; add the group in the
P3RM1SS10NS tab and rebuild to enable.
- **Capability-gated**`get_host_journal` (requires
`read_host_journal` capability set via the P3RM1SS10NS tab;
orthogonal to tool groups). Full list of capabilities and their
effects in [`docs/conventions.md#capabilities`](../conventions.md).
Also documented in [`docs/tools/scheduling.md`](../tools/scheduling.md).
- **Matrix MCP + extra servers**`mcp__matrix__*` tools and
per-agent extra MCP config. See
[`docs/tools/matrix.md`](../tools/matrix.md).
## Waking the agent from inside the container
External MCP servers (and any other in-container process) can
inject a wake-up event into the agent's inbox via the per-agent
socket at `/run/hive/mcp.sock`. Two equivalent paths:
- **Shell out to `hive wake --from <label> --body <text>`**
(use `--body -` to read body from stdin). Already on the
container's `PATH` since the harness binary is in
`systemPackages`. Convenient for shell-script integrations and
co-process daemons (matrix bridge, webhook listeners, scrapers).
- **Speak the wire protocol directly** — JSON-line over the
unix socket: `{"cmd":"wake","from":"matrix","body":"new dm
from @alice"}\n`. Same shape as any other `AgentRequest`;
see `hive-sh4re::AgentRequest::Wake`.
The wake event lands in the broker as `{from:<label>,
to:<agent>, body}`, waking whatever `recv` call the harness
is currently blocked on. The next turn fires with the wake
prompt formed from that message.
Identity = socket: anything that can connect to
`/run/hive/mcp.sock` is implicitly trusted to inject these —
the bind-mount is the agent's own container only.
## Authoritative state
`hive_ag3nt::events::Bus` carries the current turn-loop state in
addition to the broadcast channel and the events history. Variants:
- `Idle` — sitting on `Recv` waiting for mail.
- `Thinking``claude --print` is running for a turn.
- `Compacting` — operator-triggered `/compact` is in flight.
The harness flips state at the relevant transitions
(`set_state(Thinking)` before `drive_turn`, `set_state(Idle)`
after; `set_state(Compacting)` around an idle operator compact in
`turn::run_pending_compact`). Exposed via `/api/state.turn_state` +
`turn_state_since` (unix seconds); the agent page renders this rather
than deriving from SSE events.
## Tool envelope
`mcp::run_tool_envelope`: every MCP tool handler logs the request,
runs the body, logs the result. Pre-/post-log only — the inbox
status hint moved to the wake prompt + UI header.
## Tool whitelist (`mcp_config::ALLOWED_BUILTIN_TOOLS`)
- Allowed built-ins: `Edit`, `Glob`, `Grep`, `Read`, `Write`.
- Tool-group-gated built-ins: `WebFetch`, `WebSearch` (added when the
`web_tools` tool group is enabled — see P3RM1SS10NS tab).
- Denied by omission or the managed-settings deny list
(`/etc/claude-code/managed-settings.json`): `Bash`, `Task`,
`NotebookEdit`, `TodoWrite`.
- Allowed MCP tools: as listed above (by tool group).
`Bash` is disallowed — shell execution goes through
`mcp__bash__run` (background tasks with structured output +
task-id tracking) instead of an interactive shell. The bash MCP server
(`run` / `status` / `kill`) uses `allowedTools = ["*"]` so all
`mcp__bash__*` tools are always available regardless of tool groups.
`WebFetch` / `WebSearch` are off by default; enable the `web_tools`
tool group in the P3RM1SS10NS tab and rebuild the agent to enable them.

View file

@ -8,7 +8,6 @@ use std::path::{Path, PathBuf};
use std::sync::{Arc, Mutex};
use std::time::Duration;
use anyhow::Result;
use clap::{Parser, Subcommand};
use hive_ag3nt::events::{Bus, LiveEvent, TurnState};
@ -108,7 +107,7 @@ fn log_system_event(bus: &Bus, from: &str, body: &str) {
}
/// Body string for the turn-failure notification we route to
/// `<parent>` on `TurnOutcome::Failed`. Reads the hive-qualified
/// `<parent>` on `TurnError::Failed`. Reads the hive-qualified
/// identity so the receiver sees `agent@hive` rather than relying on
/// the caller threading a `label` through every turn-handling layer.
/// Falls back to `<unknown>` when `HIVE_LABEL` is missing so a
@ -277,39 +276,37 @@ trait Surface {
/// Talks `AgentRequest` / `AgentResponse`.
struct AgentSurface;
/// Issue an `Ok`-expecting fire-and-forget broker request, logging any
/// rejection / unexpected response / transport error under `label`. Shared by
/// the `Surface` methods that don't need the reply (`ack_turn`,
/// `requeue_inflight`, `graceful_stop_complete`).
async fn fire_and_forget(socket: &Path, req: AgentRequest, label: &str) {
match client::request::<_, AgentResponse>(socket, &req).await {
Ok(AgentResponse::Ok) => {}
Ok(AgentResponse::Err { message }) => {
tracing::warn!(%message, "{label} rejected by broker");
}
Ok(other) => tracing::warn!(?other, "{label} unexpected response"),
Err(e) => tracing::warn!(error = ?e, "{label} transport error"),
}
}
impl Surface for AgentSurface {
async fn ack_turn(socket: &Path) {
match client::request::<_, AgentResponse>(socket, &AgentRequest::AckTurn).await {
Ok(AgentResponse::Ok) => {}
Ok(AgentResponse::Err { message }) => {
tracing::warn!(%message, "ack_turn rejected by broker");
}
Ok(other) => tracing::warn!(?other, "ack_turn unexpected response"),
Err(e) => tracing::warn!(error = ?e, "ack_turn transport error"),
}
fire_and_forget(socket, AgentRequest::AckTurn, "ack_turn").await;
}
async fn requeue_inflight(socket: &Path) {
match client::request::<_, AgentResponse>(socket, &AgentRequest::RequeueInflight).await {
Ok(AgentResponse::Ok) => {}
Ok(AgentResponse::Err { message }) => {
tracing::warn!(%message, "requeue_inflight rejected by broker");
}
Ok(other) => tracing::warn!(?other, "requeue_inflight unexpected response"),
Err(e) => tracing::warn!(error = ?e, "requeue_inflight transport error"),
}
fire_and_forget(socket, AgentRequest::RequeueInflight, "requeue_inflight").await;
}
async fn graceful_stop_complete(socket: &Path) {
match client::request::<_, AgentResponse>(socket, &AgentRequest::GracefulStopComplete).await
{
Ok(AgentResponse::Ok) => {}
Ok(AgentResponse::Err { message }) => {
tracing::warn!(%message, "graceful_stop_complete rejected by broker");
}
Ok(other) => tracing::warn!(?other, "graceful_stop_complete unexpected response"),
Err(e) => tracing::warn!(error = ?e, "graceful_stop_complete transport error"),
}
fire_and_forget(
socket,
AgentRequest::GracefulStopComplete,
"graceful_stop_complete",
)
.await;
}
async fn inbox_unread(socket: &Path) -> u64 {
@ -432,7 +429,7 @@ async fn serve_main<S: Surface>(socket: &Path, poll_ms: u64) -> Result<()> {
// through `<parent>` via the `send_to_parent` failure-notify path.
// The broker resolves `<parent>` per `topology::parent_of`;
// root agents fall through to operator.
for failure in plugins::install_configured(socket).await {
for failure in plugins::install_configured().await {
S::send_to_parent(socket, failure).await;
}
tokio::spawn(hive_ag3nt::forge_notify::run(socket.to_path_buf()));
@ -549,8 +546,7 @@ async fn serve_loop<S: Surface>(
}
},
};
let ctrl =
handle_turn::<S>(socket, &bus, stats.as_ref(), files, &session, next).await;
let ctrl = handle_turn::<S>(socket, &bus, stats.as_ref(), files, &session, next).await;
if ctrl.auth_failed {
*login_state.lock().unwrap() = LoginState::NeedsLogin;
login::wait_for_login(
@ -601,13 +597,10 @@ async fn handle_turn<S: Surface>(
let outcome = turn::drive_turn(&prompt, files, bus, session).await;
turn::emit_turn_end(bus, &outcome);
bus.set_state(TurnState::Idle);
if matches!(
outcome,
turn::TurnOutcome::Ok | turn::TurnOutcome::Compacted
) {
if outcome.is_ok() {
S::ack_turn(socket).await;
}
if matches!(outcome, turn::TurnOutcome::RateLimited) {
if matches!(outcome, Err(turn::TurnError::RateLimited)) {
let secs = turn::rate_limit_sleep_secs();
bus.emit_status("rate_limited");
bus.emit(LiveEvent::Note {
@ -618,7 +611,7 @@ async fn handle_turn<S: Surface>(
S::requeue_inflight(socket).await;
bus.emit_status("online");
}
if matches!(outcome, turn::TurnOutcome::AuthFailed) {
if matches!(outcome, Err(turn::TurnError::AuthFailed)) {
bus.emit_status("needs_login_idle");
bus.emit(LiveEvent::Note {
text: "API 401 — waiting for re-login via web UI".into(),
@ -626,7 +619,15 @@ async fn handle_turn<S: Surface>(
tracing::warn!("auth-failed; parking until re-login");
S::requeue_inflight(socket).await;
}
if let turn::TurnOutcome::Failed(e) = &outcome {
if matches!(outcome, Err(turn::TurnError::PromptTooLong)) {
// `drive_turn` already archived the session; requeue the message so it
// redelivers into the fresh session (which fits — the wake prompt is
// tiny, the overflow was the now-cleared context). No status park: the
// agent is healthy, it just needs one more delivery.
tracing::warn!("prompt-too-long; session archived, requeueing message for a fresh turn");
S::requeue_inflight(socket).await;
}
if let Err(turn::TurnError::Failed(e)) = &outcome {
S::send_to_parent(socket, format_turn_failure(e)).await;
}
if let Some(stats) = stats {
@ -659,7 +660,7 @@ async fn handle_turn<S: Surface>(
tracing::info!(%pending, "pending messages after turn; fetching next");
}
TurnControl {
auth_failed: matches!(outcome, turn::TurnOutcome::AuthFailed),
auth_failed: matches!(outcome, Err(turn::TurnError::AuthFailed)),
continue_requested: consume_continue_sentinel(),
pending,
}

View file

@ -76,7 +76,9 @@ where
}
}
}
Err(last_err.unwrap_or_else(|| anyhow!("hive socket: retries exhausted")))
// Reaching here means the final attempt returned `Transient`, which always
// sets `last_err` — so this is infallible.
Err(last_err.expect("a transient failure on the final attempt set last_err"))
}
/// Transient = connect / IO error worth a retry (server restart, broken

View file

@ -433,7 +433,6 @@ impl EventStore {
}
}
/// Authoritative turn-loop state. The harness owns it; the web UI
/// reads via `/api/state` and renders. Lives alongside the bus
/// because everyone who has a `Bus` already has the right handle to
@ -467,12 +466,6 @@ pub fn configured_model() -> Option<&'static str> {
.map(|s| &*Box::leak(s.into_boxed_str()))
}
/// Return the model to use when no config and no persisted override exist.
#[must_use]
pub fn default_model() -> &'static str {
configured_model().unwrap_or(DEFAULT_MODEL)
}
/// Compiled-in fallback effort level — matches the `effortLevel` baked
/// into `prompts/claude-settings.json`.
pub const DEFAULT_EFFORT: &str = "medium";
@ -908,6 +901,16 @@ impl Bus {
*self.api_context_window.lock().unwrap()
}
/// The effective context window for `model`: the API-reported window if a
/// turn has completed ([`api_context_window`]), else the per-model default
/// ([`context_window_tokens`]). Single accessor so the state + dashboard
/// endpoints agree by construction.
#[must_use]
pub fn effective_context_window(&self, model: &str) -> u64 {
self.api_context_window()
.unwrap_or_else(|| context_window_tokens(model))
}
/// Walk a stream-json value for `tool_use` blocks and bump the
/// per-turn counter for each one we find. Called by the stdout
/// pump on every parsed line. Cheap when the line isn't an

View file

@ -6,8 +6,8 @@
//! off forge's own unread-state, and the agent reading the thread via the
//! CLI is what marks it read. A delivery-dedupe cursor (thread id →
//! last-delivered `updated_at`) stops the still-unread notification from
//! re-firing a wake every poll; self-echo and drop-listed notifications
//! are still marked read directly. The cursor is persisted as the
//! re-firing a wake every poll; self-echo notifications (the agent's own
//! writes) are still marked read directly. The cursor is persisted as the
//! `forge_cursor` field of the harness's consolidated `hyperhive-harness.json`
//! (via [`crate::events`]) and reloaded on boot so a container
//! rebuild/restart doesn't re-deliver the whole currently-unread backlog —
@ -247,7 +247,7 @@ fn truncate(s: &str, max: usize) -> String {
let end = s
.char_indices()
.map(|(i, _)| i)
.take_while(|&i| i <= max - 3)
.take_while(|&i| i <= max.saturating_sub(3))
.last()
.unwrap_or(0);
format!("{}", &s[..end])
@ -314,6 +314,20 @@ fn render_truncated_mentions(lines: &[&str]) -> String {
out
}
/// The escaped body excerpt + overflowed-mention suffix for one issue / PR /
/// comment body, computed in the one required order: truncate → diff the
/// overflowed `@mention` lines against the *unescaped* excerpt → heading-escape
/// the excerpt. Both notification formatters build their body block from this
/// pair so the ordering contract lives in exactly one place. Returns
/// `(escaped_excerpt, mentions_suffix)`; `mentions_suffix` is empty when the
/// body fit (nothing was truncated away).
fn render_body_excerpt(raw: &str) -> (String, String) {
let raw_excerpt = truncate(raw, BODY_TRUNCATE);
let mentions = render_truncated_mentions(&extract_truncated_mention_lines(raw, &raw_excerpt));
let excerpt = escape_md_headings(&raw_excerpt);
(excerpt, mentions)
}
/// Map a Forgejo review state to a readable action label.
/// Returns `None` for non-review states (regular comments have no `state` field;
/// `PENDING` means the review was saved but not submitted yet).
@ -378,7 +392,9 @@ async fn format_notification(
fetch_json(client, subject_api_url, token).await
};
let is_pr = matches!(notif_type, "Pull Request" | "Pull");
// Forgejo's notification `subject.type` is "Pull" / "Issue" (never
// "Pull Request") — see the comment in `build_meta_suffix` below.
let is_pr = notif_type == "Pull";
let meta_suffix = build_meta_suffix(subject.as_ref(), is_pr);
// Determine whether this notification was triggered by a comment/review or
@ -513,23 +529,17 @@ async fn format_comment_notification(
..
} = meta;
// Truncate → mention-overflow → escape, in that order. See
// `docs/forge.md::Body excerpt + truncation + heading escape` for
// why truncate comes before escape (mention diff compares against
// unescaped raw body).
let raw_excerpt = truncate(body_text, BODY_TRUNCATE);
let truncated_mentions = if body_text.len() > BODY_TRUNCATE {
render_truncated_mentions(&extract_truncated_mention_lines(body_text, &raw_excerpt))
} else {
String::new()
};
let body_for_embed = escape_md_headings(&raw_excerpt);
// Truncate → mention-overflow → escape (see `render_body_excerpt`).
let (body_for_embed, truncated_mentions) = render_body_excerpt(body_text);
if let Some(review_label) = review_state {
// Review submission on a PR.
let kind = format!("PR {review_label}{num}{repo}");
let mut out = format!("[{kind}] {title}\nurl: {url}");
if body_text.is_empty() {
write!(out, "\n\nreviewer: {author}").ok();
// Bodiless review: name who reviewed. `reviewed by:` (not
// `reviewer:`) to avoid colliding with the `reviewer:` line the
// meta suffix carries for a PR's *requested* reviewers.
write!(out, "\n\nreviewed by: {author}").ok();
} else {
write!(out, "\n\n{author}: {body_for_embed}{truncated_mentions}").ok();
}
@ -541,9 +551,6 @@ async fn format_comment_notification(
let mut out = format!(
"[{kind}] {title}\nurl: {url}\n\n{author}: {body_for_embed}{truncated_mentions}"
);
if out.ends_with('\n') {
out.pop();
}
out.push_str(meta_suffix);
Some(out)
}
@ -633,20 +640,15 @@ fn format_state_change_notification(
kind
};
// Include the start of the issue/PR description so the agent
// gets context without a follow-up fetch. Same truncate →
// mention-overflow → escape pipeline as comment bodies (see
// `docs/forge.md::Body excerpt + truncation + heading escape`).
// Include the start of the issue/PR description so the agent gets context
// without a follow-up fetch. Same body pipeline as comment bodies.
let body_block = subject
.as_ref()
.and_then(|s| s["body"].as_str())
.map(str::trim)
.filter(|s| !s.is_empty())
.map(|raw| {
let raw_excerpt = truncate(raw, BODY_TRUNCATE);
let truncated = extract_truncated_mention_lines(raw, &raw_excerpt);
let mentions = render_truncated_mentions(&truncated);
let excerpt = escape_md_headings(&raw_excerpt);
let (excerpt, mentions) = render_body_excerpt(raw);
format!("\n\n{excerpt}{mentions}")
})
.unwrap_or_default();
@ -712,8 +714,15 @@ fn parse_rfc3339_secs(s: &str) -> Option<i64> {
if !(rest.is_empty() || rest.starts_with('Z')) {
let sign = rest.as_bytes()[0];
let off = &rest[1..];
let oh: i64 = off.get(0..2)?.parse().ok()?;
let om: i64 = off.get(3..5).unwrap_or("00").parse().ok()?;
// Accept `HH:MM` (Forgejo's form) and bare `HHMM`; minutes optional.
// Both fields fail the same way — a present-but-unparseable component
// returns `None` rather than one silently defaulting.
let (hh, mm) = match off.split_once(':') {
Some((h, m)) => (h, m),
None => (off.get(0..2)?, off.get(2..4).unwrap_or("00")),
};
let oh: i64 = hh.parse().ok()?;
let om: i64 = if mm.is_empty() { 0 } else { mm.parse().ok()? };
let offset = oh * 3_600 + om * 60;
match sign {
b'+' => epoch -= offset,
@ -885,11 +894,11 @@ fn should_deliver(delivered: &HashMap<u64, String>, id: u64, updated_at: &str) -
}
/// Mark a notification thread as read. Best-effort — logs on failure but
/// does not abort the poll loop. Called only on the self-echo and
/// drop-listed paths (the agent's own writes / explicitly-suppressed
/// reasons) — delivered threads are deliberately left unread for the
/// read-before-comment guard, and a failed delivery is left unread + out
/// of the dedupe cursor so it resurfaces on the next poll tick.
/// does not abort the poll loop. Called only on the self-echo path (the
/// agent's own comment/review/creation writes) — delivered threads are
/// deliberately left unread for the read-before-comment guard, and a failed
/// delivery is left unread + out of the dedupe cursor so it resurfaces on the
/// next poll tick.
async fn mark_read(client: &reqwest::Client, forge_url: &str, token: &str, id: u64) {
let mark_url = format!("{forge_url}/api/v1/notifications/threads/{id}");
match client

View file

@ -14,14 +14,18 @@ pub fn label() -> String {
env::var("HIVE_LABEL").unwrap_or_default()
}
/// `env::var(key)` reduced to `Some(value)` only when the var is set and
/// non-empty — the shared shape of the hive/swarm display-name lookups below.
fn non_empty_env(key: &str) -> Option<String> {
env::var(key).ok().filter(|s| !s.is_empty())
}
/// The hive's canonical DNS domain when set, otherwise None. Single-hive
/// deployments where `HYPERHIVE_HIVE_DOMAIN` is unset return None — callers
/// then degrade gracefully to the short label.
#[must_use]
pub fn hive_domain() -> Option<String> {
env::var("HYPERHIVE_HIVE_DOMAIN")
.ok()
.filter(|s| !s.is_empty())
non_empty_env("HYPERHIVE_HIVE_DOMAIN")
}
/// Human display name of this hive (e.g. `pr1ma`). Distinct from
@ -32,9 +36,7 @@ pub fn hive_domain() -> Option<String> {
/// at their discretion.
#[must_use]
pub fn hive_name() -> Option<String> {
env::var("HYPERHIVE_HIVE_NAME")
.ok()
.filter(|s| !s.is_empty())
non_empty_env("HYPERHIVE_HIVE_NAME")
}
/// Human display name of the wider swarm this hive belongs to (e.g.
@ -43,9 +45,7 @@ pub fn hive_name() -> Option<String> {
/// `services.hyperhive.swarmName` option is unset.
#[must_use]
pub fn swarm_name() -> Option<String> {
env::var("HYPERHIVE_SWARM_NAME")
.ok()
.filter(|s| !s.is_empty())
non_empty_env("HYPERHIVE_SWARM_NAME")
}
/// One peer hive in the same swarm. Parsed from `HYPERHIVE_PEERS`.

View file

@ -9,6 +9,7 @@ pub mod identity;
pub mod login;
pub mod login_session;
pub mod mcp;
pub mod mcp_config;
pub mod mcp_loose_ends;
pub mod paths;
pub mod plugins;

View file

@ -2,11 +2,12 @@
//! provided by hive-c0re and persists across container destroy/recreate so
//! OAuth tokens survive.
//!
//! "Has session" today means "the dir contains at least one regular file."
//! That's a heuristic: a fresh bind-mount starts empty, and `claude auth login`
//! writes credentials into the dir. We may refine later (probe for the
//! specific credentials filename, or run a no-op `claude` call) once the
//! exact layout is locked in.
//! "Has session" means the dir contains at least one of the credential files
//! in [`CRED_FILE_NAMES`] — the same set `/logout` (`web_ui::auth`) deletes to
//! force re-login. Keying both off one constant keeps boot detection and
//! logout in agreement: logout deliberately preserves session-history files,
//! so a "contains any regular file" check would wrongly report `Online` after
//! a logout + container recreate and burn a turn 401-ing before it reroutes.
use std::path::{Path, PathBuf};
use std::sync::{Arc, Mutex};
@ -24,20 +25,61 @@ pub fn default_dir() -> PathBuf {
crate::paths::claude_dir()
}
/// Returns `true` if `dir` exists and contains any regular file. Used at
/// startup to decide whether to enter the turn loop (logged in) or stay in
/// the partial-run "needs login" state.
/// The credential files that constitute a logged-in claude session inside
/// [`default_dir`]. A session exists iff at least one is present; a login
/// "refresh" is a change to one of them. `/logout` (`web_ui::auth`) deletes
/// exactly these to force re-login while preserving session-history files —
/// so boot detection ([`has_session`]) and logout agree by construction.
/// Rationale + the previous wholesale-wipe shape we replaced live in
/// [`docs/web-ui/agent.md::Per-agent endpoints`](../../docs/web-ui/agent.md)
/// (the `/api/logout` bullet).
pub const CRED_FILE_NAMES: &[&str] = &[".credentials.json", "mcp-needs-auth-cache.json"];
/// Is `entry` a regular file whose name is one of [`CRED_FILE_NAMES`]?
fn is_cred_file(entry: &std::fs::DirEntry) -> bool {
entry.file_type().is_ok_and(|t| t.is_file())
&& entry
.file_name()
.to_str()
.is_some_and(|n| CRED_FILE_NAMES.contains(&n))
}
/// Returns `true` if `dir` exists and holds at least one credential file
/// (see [`CRED_FILE_NAMES`]). Used at startup to decide whether to enter the
/// turn loop (logged in) or stay in the partial-run "needs login" state.
#[must_use]
pub fn has_session(dir: &Path) -> bool {
let Ok(entries) = std::fs::read_dir(dir) else {
return false;
};
for entry in entries.flatten() {
if entry.file_type().is_ok_and(|t| t.is_file()) {
return true;
entries.flatten().any(|e| is_cred_file(&e))
}
/// Outcome of [`clear_session`]: which credential files were removed and any
/// non-fatal per-file errors (e.g. permission denied). A file that was already
/// absent is not reported — deletion is idempotent.
#[derive(Debug, Default)]
pub struct ClearedSession {
pub wiped: Vec<&'static str>,
pub warnings: Vec<String>,
}
/// Delete the credential files (see [`CRED_FILE_NAMES`]) from `dir`, forcing a
/// re-login on the next turn, while preserving the session-history files
/// alongside them so `claude --continue` keeps working after a fresh login.
/// Idempotent: an already-absent file is skipped, not reported. This is the
/// write-side counterpart to [`has_session`]; `/logout` (`web_ui::auth`) drives
/// it.
pub async fn clear_session(dir: &Path) -> ClearedSession {
let mut cleared = ClearedSession::default();
for name in CRED_FILE_NAMES {
match tokio::fs::remove_file(dir.join(name)).await {
Ok(()) => cleared.wiped.push(name),
Err(e) if e.kind() == std::io::ErrorKind::NotFound => {}
Err(e) => cleared.warnings.push(format!("{name}: {e}")),
}
}
false
cleared
}
/// Login state the harness reports to its web UI.
@ -105,16 +147,16 @@ pub async fn wait_for_login(
}
}
/// Snapshot of the credentials dir at a point in time: number of
/// regular files + newest `mtime` across them. The two axes are both
/// load-bearing for `wait_for_login`'s refresh check (`session_refreshed`):
/// mtime catches the common case (re-login overwrites an existing
/// credentials file in-place), `file_count` catches the pathological case
/// where `meta.modified()` errors on every file (exotic fs, NFS quirks)
/// so the mtime axis stays `None` forever but new files still trigger a
/// resume. Defaults to `{0, None}` on `read_dir` failure (missing or
/// unreadable dir) — `wait_for_login` then resumes when files first
/// appear.
/// Snapshot of the credential files (see [`CRED_FILE_NAMES`]) in the dir at a
/// point in time: how many are present + newest `mtime` across them. The two
/// axes are both load-bearing for `wait_for_login`'s refresh check
/// (`session_refreshed`): mtime catches the common case (re-login overwrites
/// an existing credentials file in-place), `file_count` catches the
/// pathological case where `meta.modified()` errors on every file (exotic fs,
/// NFS quirks) so the mtime axis stays `None` forever but a new credential
/// file still triggers a resume. Defaults to `{0, None}` on `read_dir` failure
/// (missing or unreadable dir) — `wait_for_login` then resumes when a
/// credential file first appears.
#[derive(Debug, Clone, Copy, Default, PartialEq, Eq)]
struct DirSnapshot {
file_count: usize,
@ -127,7 +169,7 @@ fn snapshot_dir(dir: &Path) -> DirSnapshot {
};
let mut snap = DirSnapshot::default();
for entry in entries.flatten() {
if !entry.file_type().is_ok_and(|t| t.is_file()) {
if !is_cred_file(&entry) {
continue;
}
snap.file_count += 1;
@ -165,7 +207,31 @@ mod tests {
use std::fs;
use std::time::{Duration, SystemTime};
use super::{DirSnapshot, session_refreshed, snapshot_dir};
use super::{DirSnapshot, has_session, session_refreshed, snapshot_dir};
#[test]
fn has_session_only_counts_credential_files() {
let dir = tempfile::tempdir().unwrap();
// Session-history files (what `/logout` preserves) must NOT read as a
// logged-in session — this is the logout+recreate 401 bug.
fs::write(dir.path().join("history.jsonl"), b"{}").unwrap();
fs::write(dir.path().join("some-project-uuid.json"), b"{}").unwrap();
assert!(!has_session(dir.path()));
// A real credential file does.
fs::write(dir.path().join(".credentials.json"), b"{}").unwrap();
assert!(has_session(dir.path()));
}
#[test]
fn snapshot_dir_ignores_non_credential_files() {
let dir = tempfile::tempdir().unwrap();
fs::write(dir.path().join("history.jsonl"), b"{}").unwrap();
let snap = snapshot_dir(dir.path());
assert_eq!(
snap.file_count, 0,
"history files must not count as session"
);
}
#[test]
fn snapshot_dir_empty_dir_is_default() {
@ -191,11 +257,11 @@ mod tests {
#[test]
fn snapshot_dir_picks_latest_mtime_and_counts_files() {
let dir = tempfile::tempdir().unwrap();
fs::write(dir.path().join("old.json"), b"{}").unwrap();
fs::write(dir.path().join(".credentials.json"), b"{}").unwrap();
// Sleep so the second file's mtime is strictly greater than
// the first on filesystems with low timestamp resolution.
std::thread::sleep(Duration::from_millis(20));
let newer_path = dir.path().join("newer.json");
let newer_path = dir.path().join("mcp-needs-auth-cache.json");
fs::write(&newer_path, b"{}").unwrap();
let snap = snapshot_dir(dir.path());
assert_eq!(snap.file_count, 2);
@ -204,13 +270,13 @@ mod tests {
}
#[test]
fn session_refreshed_first_login_flips_on_any_file() {
// Empty-dir snapshot → any file appearing means a fresh
fn session_refreshed_first_login_flips_on_cred_file() {
// Empty-dir snapshot → a credential file appearing means a fresh
// login landed. First-time login semantics.
let dir = tempfile::tempdir().unwrap();
let snapshot = snapshot_dir(dir.path());
assert!(!session_refreshed(snapshot, snapshot_dir(dir.path())));
fs::write(dir.path().join("credentials.json"), b"{}").unwrap();
fs::write(dir.path().join(".credentials.json"), b"{}").unwrap();
assert!(session_refreshed(snapshot, snapshot_dir(dir.path())));
}
@ -220,7 +286,7 @@ mod tests {
// must NOT immediately return — it would loop straight into
// another 401-failing turn.
let dir = tempfile::tempdir().unwrap();
fs::write(dir.path().join("credentials.json"), b"{}").unwrap();
fs::write(dir.path().join(".credentials.json"), b"{}").unwrap();
let snapshot = snapshot_dir(dir.path());
assert_eq!(snapshot.file_count, 1);
// No change to the file → loop must NOT exit.
@ -233,10 +299,10 @@ mod tests {
// flow lands a refreshed credentials file — its mtime bumps
// strictly past the snapshot and wait_for_login resumes.
let dir = tempfile::tempdir().unwrap();
fs::write(dir.path().join("credentials.json"), b"{}").unwrap();
fs::write(dir.path().join(".credentials.json"), b"{}").unwrap();
let snapshot = snapshot_dir(dir.path());
std::thread::sleep(Duration::from_millis(20));
fs::write(dir.path().join("credentials.json"), b"{\"v\":2}").unwrap();
fs::write(dir.path().join(".credentials.json"), b"{\"v\":2}").unwrap();
assert!(session_refreshed(snapshot, snapshot_dir(dir.path())));
}
@ -246,7 +312,7 @@ mod tests {
// skew between snapshot and probe) must keep waiting until a
// file's mtime actually exceeds it, not return on first poll.
let dir = tempfile::tempdir().unwrap();
fs::write(dir.path().join("credentials.json"), b"{}").unwrap();
fs::write(dir.path().join(".credentials.json"), b"{}").unwrap();
let snapshot = DirSnapshot {
file_count: 1,
newest_mtime: Some(SystemTime::now() + Duration::from_hours(1)),
@ -262,12 +328,12 @@ mod tests {
// here by forging a snapshot with file_count=1 + no mtime, then
// writing a second file.
let dir = tempfile::tempdir().unwrap();
fs::write(dir.path().join("a"), b"{}").unwrap();
fs::write(dir.path().join(".credentials.json"), b"{}").unwrap();
let forged = DirSnapshot {
file_count: 1,
newest_mtime: None,
};
fs::write(dir.path().join("b"), b"{}").unwrap();
fs::write(dir.path().join("mcp-needs-auth-cache.json"), b"{}").unwrap();
// Real snapshot has file_count=2, so refresh fires even
// though the mtime axis would be inconclusive.
assert!(session_refreshed(forged, snapshot_dir(dir.path())));

View file

@ -24,117 +24,6 @@ use rmcp::{
use crate::client;
/// Wire-protocol-agnostic view of a hyperhive socket response. Both flavors
/// of `AgentServer` convert into this so the tool formatters can be shared.
#[derive(Debug)]
pub enum SocketReply {
Ok,
Err(String),
/// Unified `recv` result: zero or more messages popped in one
/// round-trip. Empty vec = "(empty)" path; single-message = the
/// standard wake body; multi = batch render with per-message
/// separators. Per-row `id` is rendered as a `[msg #<id>]` marker
/// so claude can bulk-triage via `ack_until` (turn-level ack still
/// rides `AckTurn`); `redelivered` triggers the "may already be
/// handled" banner in `format_recv` for that specific row.
Messages(Vec<hive_sh4re::DeliveredMessage>),
Status(u64),
/// `ack_until` result: rows newly marked handled.
Acked(u64),
QuestionQueued(i64),
Recent(Vec<hive_sh4re::InboxRow>),
Logs(String),
HostJournal(String),
/// `list_schedules` result — returned by `list_schedules` (scheduling tool group).
Schedules(Vec<hive_sh4re::WireSchedule>),
/// `list_containers` result — descendant containers with running status.
Containers(Vec<hive_sh4re::ContainerInfo>),
LooseEnds(Vec<hive_sh4re::LooseEnd>),
PendingRemindersCount(u64),
ReminderRollup(hive_sh4re::ReminderStats),
AgentMeta {
name: String,
running: bool,
hyperhive_rev: Option<String>,
status_text: Option<String>,
status_set_at: Option<i64>,
hive_name: Option<String>,
swarm_name: Option<String>,
matrix_accounts: Vec<hive_sh4re::MatrixIdentity>,
},
/// `create_repo` result — the new repo's full name + clone URL.
RepoCreated {
full_name: String,
clone_url: String,
},
}
impl From<hive_sh4re::Response> for SocketReply {
fn from(r: hive_sh4re::Response) -> Self {
match r {
hive_sh4re::Response::Ok => Self::Ok,
hive_sh4re::Response::Err { message } => Self::Err(message),
hive_sh4re::Response::Messages { messages } => Self::Messages(messages),
hive_sh4re::Response::Status { unread } => Self::Status(unread),
hive_sh4re::Response::Acked { count } => Self::Acked(count),
hive_sh4re::Response::Recent { rows } => Self::Recent(rows),
hive_sh4re::Response::QuestionQueued { id } => Self::QuestionQueued(id),
hive_sh4re::Response::LooseEnds { loose_ends } => Self::LooseEnds(loose_ends),
hive_sh4re::Response::PendingRemindersCount { count } => {
Self::PendingRemindersCount(count)
}
hive_sh4re::Response::ReminderRollup(stats) => Self::ReminderRollup(stats),
hive_sh4re::Response::Logs { content } => Self::Logs(content),
hive_sh4re::Response::HostJournal { content } => Self::HostJournal(content),
hive_sh4re::Response::Schedules { schedules } => Self::Schedules(schedules),
hive_sh4re::Response::Containers { containers } => Self::Containers(containers),
// A graceful stop is pending — the inbox is fenced. Returning an
// *empty* inbox here is ambiguous: claude's "park on recv" habit
// makes it long-poll again instead of ending the turn, so the
// stop-checkpoint turn never finishes and the drain wait times out
// into a hard stop. Return a single explicit directive instead, so
// every recv during the stop unmissably tells claude to flush + end.
hive_sh4re::Response::GracefulStop => Self::Messages(vec![hive_sh4re::DeliveredMessage {
from: "graceful-stop".into(),
body: "⛔ GRACEFUL STOP IN PROGRESS — the container shuts down as soon as this turn \
ends. Flush anything worth keeping to your durable /state files, then END \
YOUR TURN now. Do NOT call recv again: the inbox is fenced and recv will \
only keep returning this same notice."
.into(),
id: 0,
redelivered: false,
in_reply_to: None,
}]),
hive_sh4re::Response::AgentMeta {
name,
running,
hyperhive_rev,
status_text,
status_set_at,
hive_name,
swarm_name,
matrix_accounts,
} => Self::AgentMeta {
name,
running,
hyperhive_rev,
status_text,
status_set_at,
hive_name,
swarm_name,
matrix_accounts,
},
hive_sh4re::Response::RepoCreated {
full_name,
clone_url,
} => Self::RepoCreated {
full_name,
clone_url,
},
}
}
}
/// Write (or remove) the status file in the agent's own `state/` directory.
/// Called by `AgentServer::set_status` for both agent and manager flavors
/// before dispatching the wire `SetStatus` request (which only triggers a
@ -180,16 +69,32 @@ fn write_status_file(text: &str) -> Result<(), String> {
result.map_err(|e| format!("set_status write failed: {e}"))
}
/// Render the three identical failure arms every data-returning tool handler
/// repeats: a broker `Err` → `"{tool} failed: {m}"`, an unexpected `Ok` variant
/// → `"{tool} unexpected response: …"`, and a transport error → `"{tool}
/// transport error: …"`. Handlers match their own happy-path variant and route
/// everything else here via a catch-all arm (`other => reply_err(other, tool)`),
/// so the triplet lives in exactly one place.
fn reply_err(resp: Result<hive_sh4re::Response, anyhow::Error>, tool: &str) -> String {
match resp {
Ok(hive_sh4re::Response::Err { message }) => format!("{tool} failed: {message}"),
Ok(other) => format!("{tool} unexpected response: {other:?}"),
Err(e) => format!("{tool} transport error: {e:#}"),
}
}
/// Format helper for "send-like" tools (anything that expects an `Ok`).
/// `tool` and `ok_msg` only appear in the result string; they don't change
/// behavior.
#[must_use]
pub fn format_ack(resp: Result<SocketReply, anyhow::Error>, tool: &str, ok_msg: String) -> String {
pub fn format_ack(
resp: Result<hive_sh4re::Response, anyhow::Error>,
tool: &str,
ok_msg: String,
) -> String {
match resp {
Ok(SocketReply::Ok) => ok_msg,
Ok(SocketReply::Err(m)) => format!("{tool} failed: {m}"),
Ok(other) => format!("{tool} unexpected response: {other:?}"),
Err(e) => format!("{tool} transport error: {e:#}"),
Ok(hive_sh4re::Response::Ok) => ok_msg,
other => reply_err(other, tool),
}
}
@ -204,14 +109,41 @@ pub fn format_ack(resp: Result<SocketReply, anyhow::Error>, tool: &str, ok_msg:
/// so the model can tell where one ends and the next begins;
/// per-message redelivery banners included.
#[must_use]
pub fn format_recv(resp: Result<SocketReply, anyhow::Error>, waited: bool) -> String {
pub fn format_recv(resp: Result<hive_sh4re::Response, anyhow::Error>, waited: bool) -> String {
match resp {
Ok(hive_sh4re::Response::Messages { messages }) => render_recv_messages(&messages, waited),
// A graceful stop is pending — the inbox is fenced. Render a single
// explicit directive (not an empty inbox, which claude's "park on recv"
// habit would long-poll again, stalling the stop-checkpoint turn until
// the drain wait times out into a hard stop) so every recv during the
// stop unmissably tells claude to flush + end.
Ok(hive_sh4re::Response::GracefulStop) => {
render_recv_messages(&[graceful_stop_message()], waited)
}
other => reply_err(other, "recv"),
}
}
/// The synthetic single-message directive rendered for a fenced (graceful-stop)
/// inbox — see the `GracefulStop` arm of [`format_recv`].
fn graceful_stop_message() -> hive_sh4re::DeliveredMessage {
hive_sh4re::DeliveredMessage {
from: "graceful-stop".into(),
body: "⛔ GRACEFUL STOP IN PROGRESS — the container shuts down as soon as this turn \
ends. Flush anything worth keeping to your durable /state files, then END \
YOUR TURN now. Do NOT call recv again: the inbox is fenced and recv will \
only keep returning this same notice."
.into(),
id: 0,
redelivered: false,
in_reply_to: None,
}
}
/// Render the popped-message payload of a successful `recv` (see `format_recv`
/// for the empty/single/batch shapes).
fn render_recv_messages(messages: &[hive_sh4re::DeliveredMessage], waited: bool) -> String {
use std::fmt::Write as _;
let messages = match resp {
Ok(SocketReply::Messages(m)) => m,
Ok(SocketReply::Err(m)) => return format!("recv failed: {m}"),
Ok(other) => return format!("recv unexpected response: {other:?}"),
Err(e) => return format!("recv transport error: {e:#}"),
};
if messages.is_empty() {
return if waited {
format!("(empty){IDLE_WAIT_HINT}")
@ -269,10 +201,9 @@ pub const IDLE_WAIT_HINT: &str = " — nothing arrived before the wait timed out
If you have other useful work (assigned issues, in-flight PRs, a docs sweep, \
notes to update), do that now rather than immediately parking on recv again.";
/// Inner renderer for a `Vec<LooseEnd>` already extracted from the
/// socket reply. Called by both `format_loose_ends` (which handles the
/// `Result<SocketReply>` wrapper) and the augmented `get_loose_ends`
/// handler (which injects the `UnreadMatrix` entry before formatting).
/// Inner renderer for a `Vec<LooseEnd>` already extracted from the socket
/// reply. Called by the `get_loose_ends` handler, which injects the
/// `UnreadMatrix` entry before formatting.
fn render_loose_ends(loose_ends: &[hive_sh4re::LooseEnd]) -> String {
use std::fmt::Write as _;
if loose_ends.is_empty() {
@ -351,21 +282,6 @@ fn render_loose_ends(loose_ends: &[hive_sh4re::LooseEnd]) -> String {
out
}
/// Format helper for `get_loose_ends`: renders a short bulleted list
/// of pending approvals + questions + reminders. Empty list collapses
/// to a clear marker so claude doesn't go hunting for a payload that
/// isn't there.
#[must_use]
pub fn format_loose_ends(resp: Result<SocketReply, anyhow::Error>) -> String {
let loose_ends = match resp {
Ok(SocketReply::LooseEnds(t)) => t,
Ok(SocketReply::Err(m)) => return format!("get_loose_ends failed: {m}"),
Ok(other) => return format!("get_loose_ends unexpected response: {other:?}"),
Err(e) => return format!("get_loose_ends transport error: {e:#}"),
};
render_loose_ends(&loose_ends)
}
/// Per-room unread entry returned by `matrix_unread_summary`. Mirrors
/// `hive-matrix-mcp`'s `RoomUnread` but defined locally to avoid a
/// cross-crate dep on the matrix-sdk crate tree.
@ -464,11 +380,11 @@ fn loose_end_kind_label(kind: hive_sh4re::CancelLooseEndKind) -> &'static str {
/// (it would be a stale snapshot from before the stop) so the status
/// line is implicitly `<none>` in that case — but the explicit
/// `running: no` line tells the caller WHY. See
/// `docs/turn-loop.md::Sub-agent tools` (`get_agent_meta`).
/// `docs/turn-loop/mcp.md::Core tools` (`get_agent_meta`).
#[must_use]
pub fn format_agent_meta(resp: Result<SocketReply, anyhow::Error>) -> String {
pub fn format_agent_meta(resp: Result<hive_sh4re::Response, anyhow::Error>) -> String {
match resp {
Ok(SocketReply::AgentMeta {
Ok(hive_sh4re::Response::AgentMeta {
name,
running,
hyperhive_rev,
@ -536,9 +452,7 @@ pub fn format_agent_meta(resp: Result<SocketReply, anyhow::Error>) -> String {
}
out
}
Ok(SocketReply::Err(m)) => format!("get_agent_meta failed: {m}"),
Ok(other) => format!("get_agent_meta unexpected response: {other:?}"),
Err(e) => format!("get_agent_meta transport error: {e:#}"),
other => reply_err(other, "get_agent_meta"),
}
}
@ -680,18 +594,18 @@ impl AgentServer {
Self { socket }
}
/// Issue any `Request` through the retry-aware client and pull
/// the reply through `SocketReply`. Returns the retry count so tool
/// handlers can annotate their result (see `annotate_retries`).
/// Issue any `Request` through the retry-aware client. Returns the raw
/// `Response` plus the retry count so tool handlers can annotate their
/// result (see `annotate_retries`).
///
/// `AgentRequest` / `ManagerRequest` / `Request` are all the same type
/// (hive-sh4re type aliases), so this single method covers both sockets.
async fn dispatch(
&self,
req: hive_sh4re::Request,
) -> (Result<SocketReply, anyhow::Error>, u32) {
) -> (Result<hive_sh4re::Response, anyhow::Error>, u32) {
match client::request_retried::<_, hive_sh4re::Response>(&self.socket, &req).await {
Ok((r, n)) => (Ok(SocketReply::from(r)), n),
Ok((r, n)) => (Ok(r), n),
Err(e) => (Err(e), 0),
}
}
@ -712,7 +626,7 @@ impl AgentServer {
let to = args.to.clone();
// Check per-agent allow-list (hyperhive.allowedRecipients). When no
// policy file is present (e.g. manager containers) the check is a no-op.
if let Err(refusal) = check_send_allowed(&to) {
if let Err(refusal) = crate::mcp_config::check_send_allowed(&to) {
return run_tool_envelope("send", log, async move { refusal }).await;
}
run_tool_envelope("send", log, async move {
@ -757,13 +671,11 @@ impl AgentServer {
})
.await;
let s = match resp {
Ok(SocketReply::QuestionQueued(id)) => format!(
Ok(hive_sh4re::Response::QuestionQueued { id }) => format!(
"question queued (id={id}); answer will arrive as a system \
`question_answered` event in your inbox"
),
Ok(SocketReply::Err(m)) => format!("ask failed: {m}"),
Ok(other) => format!("ask unexpected response: {other:?}"),
Err(e) => format!("ask transport error: {e:#}"),
other => reply_err(other, "ask"),
};
annotate_retries(s, retries)
})
@ -847,12 +759,10 @@ impl AgentServer {
.dispatch(hive_sh4re::Request::AckUntil { up_to: args.up_to })
.await;
let rendered = match resp {
Ok(SocketReply::Acked(count)) => {
Ok(hive_sh4re::Response::Acked { count }) => {
format!("acked {count} message(s) up to id {}", args.up_to)
}
Ok(SocketReply::Err(m)) => format!("ack_until failed: {m}"),
Ok(other) => format!("ack_until unexpected response: {other:?}"),
Err(e) => format!("ack_until transport error: {e:#}"),
other => reply_err(other, "ack_until"),
};
annotate_retries(rendered, retries)
})
@ -882,22 +792,8 @@ impl AgentServer {
.await;
// Extract the vec so we can augment before rendering.
let mut loose_ends = match resp {
Ok(SocketReply::LooseEnds(t)) => t,
Ok(SocketReply::Err(m)) => {
return annotate_retries(format!("get_loose_ends failed: {m}"), retries);
}
Ok(other) => {
return annotate_retries(
format!("get_loose_ends unexpected response: {other:?}"),
retries,
);
}
Err(e) => {
return annotate_retries(
format!("get_loose_ends transport error: {e:#}"),
retries,
);
}
Ok(hive_sh4re::Response::LooseEnds { loose_ends }) => loose_ends,
other => return annotate_retries(reply_err(other, "get_loose_ends"), retries),
};
// Prepend matrix unread entry for self-queries only (can't
// reach another agent's matrix daemon from here).
@ -1035,13 +931,11 @@ impl AgentServer {
.dispatch(hive_sh4re::Request::CreateRepo { repo: args.repo })
.await;
let s = match resp {
Ok(SocketReply::RepoCreated {
Ok(hive_sh4re::Response::RepoCreated {
full_name,
clone_url,
}) => format!("created repo {full_name} — clone: {clone_url}"),
Ok(SocketReply::Err(m)) => format!("create_repo failed: {m}"),
Ok(other) => format!("create_repo unexpected response: {other:?}"),
Err(e) => format!("create_repo transport error: {e:#}"),
other => reply_err(other, "create_repo"),
};
annotate_retries(s, retries)
})
@ -1198,7 +1092,7 @@ impl AgentServer {
run_tool_envelope("list_containers", String::new(), async move {
let (resp, retries) = self.dispatch(hive_sh4re::Request::ListDescendants).await;
let body = match resp {
Ok(SocketReply::Containers(containers)) => {
Ok(hive_sh4re::Response::Containers { containers }) => {
if containers.is_empty() {
"no descendant containers".to_owned()
} else {
@ -1212,9 +1106,7 @@ impl AgentServer {
.join("\n")
}
}
Ok(SocketReply::Err(m)) => format!("list_containers failed: {m}"),
Ok(other) => format!("list_containers unexpected response: {other:?}"),
Err(e) => format!("list_containers transport error: {e:#}"),
other => reply_err(other, "list_containers"),
};
annotate_retries(body, retries)
})
@ -1254,10 +1146,8 @@ impl AgentServer {
})
.await;
let result = match resp {
Ok(SocketReply::HostJournal(content)) => content,
Ok(SocketReply::Err(m)) => format!("get_host_journal failed: {m}"),
Ok(other) => format!("get_host_journal unexpected response: {other:?}"),
Err(e) => format!("get_host_journal transport error: {e:#}"),
Ok(hive_sh4re::Response::HostJournal { content }) => content,
other => reply_err(other, "get_host_journal"),
};
annotate_retries(result, retries)
})
@ -1381,16 +1271,14 @@ impl AgentServer {
})
.await;
let s = match resp {
Ok(SocketReply::Logs(content)) => {
Ok(hive_sh4re::Response::Logs { content }) => {
if content.is_empty() {
format!("(no journal output for {agent})")
} else {
content
}
}
Ok(SocketReply::Err(m)) => format!("get_logs failed: {m}"),
Ok(other) => format!("get_logs unexpected response: {other:?}"),
Err(e) => format!("get_logs transport error: {e:#}"),
other => reply_err(other, "get_logs"),
};
annotate_retries(s, retries)
})
@ -1576,11 +1464,11 @@ impl AgentServer {
run_tool_envelope("list_schedules", String::new(), async move {
let (resp, retries) = self.dispatch(hive_sh4re::Request::ListSchedules).await;
let body = match resp {
Ok(SocketReply::Schedules(schedules)) => serde_json::to_string(&schedules)
.unwrap_or_else(|e| format!("list_schedules: serialise: {e:#}")),
Ok(SocketReply::Err(m)) => format!("list_schedules: {m}"),
Ok(other) => format!("list_schedules unexpected response: {other:?}"),
Err(e) => format!("list_schedules transport error: {e:#}"),
Ok(hive_sh4re::Response::Schedules { schedules }) => {
serde_json::to_string(&schedules)
.unwrap_or_else(|e| format!("list_schedules: serialise: {e:#}"))
}
other => reply_err(other, "list_schedules"),
};
annotate_retries(body, retries)
})
@ -1595,27 +1483,23 @@ impl AgentServer {
)]
impl ServerHandler for AgentServer {}
/// Run an MCP server over stdio for the given flavor. Returns when the client disconnects.
/// Run the MCP server over stdio. Used by all roles. Returns when the client
/// disconnects.
///
/// # Errors
///
/// Returns an error if the MCP server fails to initialize or the transport
/// encounters a fatal error.
pub async fn serve_stdio(socket: PathBuf) -> Result<()> {
pub async fn serve_agent_stdio(socket: PathBuf) -> Result<()> {
let server = AgentServer::new(socket);
let service = server.serve(stdio()).await?;
service.waiting().await?;
Ok(())
}
/// Run the MCP server over stdio. Used by all roles.
pub async fn serve_agent_stdio(socket: PathBuf) -> Result<()> {
serve_stdio(socket).await
}
/// Run the MCP server over HTTP (rmcp streamable-http transport) on `addr`.
///
/// Unlike [`serve_stdio`] — a fresh stdio child claude respawns every turn —
/// Unlike [`serve_agent_stdio`] — a fresh stdio child claude respawns every turn —
/// this is meant to run as a long-lived in-container daemon. claude reconnects
/// to the stable URL each turn instead of respawning and re-registering a stdio
/// subprocess, which removes the per-turn MCP registration race that can strand
@ -1765,17 +1649,6 @@ pub struct CancelLooseEndArgs {
pub id: i64,
}
#[derive(Debug, serde::Deserialize, schemars::JsonSchema)]
pub struct GetLooseEndsArgs {
/// Whose loose ends to list. Omit (or `null`) for your own: approvals
/// you submitted + questions where you are asker/target + your own
/// pending reminders. Pass `"*"` for a hive-wide view of EVERY pending
/// approval, unanswered question, and reminder across the swarm. Pass a
/// specific agent name to inspect just that agent's threads.
#[serde(default)]
pub agent: Option<String>,
}
#[derive(Debug, serde::Deserialize, schemars::JsonSchema)]
pub struct AgentGetLooseEndsArgs {
/// Whose loose ends to list. Omit (or `null`) for your own. You may
@ -1937,448 +1810,26 @@ pub struct GetHostJournalArgs {
#[serde(default)]
pub until: Option<String>,
}
/// Name of the hyperhive MCP server inside claude's view. Claude prefixes
/// tools as `mcp__<this>__<tool>` (e.g. `mcp__hyperhive__send`).
pub const SERVER_NAME: &str = "hyperhive";
/// Built-in claude tools always present in every session. Anything not
/// in this list (or added by `extra_builtin_tools`) literally doesn't
/// exist in the session. Web egress (`WebFetch`/`WebSearch`) are
/// tool-group-gated (`web_tools`) — off by default. Nested agents
/// (`Task`) are intentionally omitted. `Bash` is disallowed — shell
/// execution goes through `mcp__bash__run` (background tasks
/// with structured output via `hive-bash-mcp`) instead of a raw interactive shell. `TodoWrite`
/// is omitted because the todo list lives in claude's in-process session
/// state and silently evaporates on /compact or session reset — agents
/// should plan in /state notes instead.
pub const ALLOWED_BUILTIN_TOOLS: &[&str] = &["Edit", "Glob", "Grep", "Read", "Write"];
/// Env var written by the meta renderer with a comma-separated list of
/// `hive_sh4re::ToolGroup` `snake_case` names (e.g. `"messaging,inbox,meta"`).
/// When present, the harness expands the groups into per-tool allow entries
/// instead of using the hardcoded flavor default. See `docs/conventions.md::Tool groups`.
const TOOL_GROUPS_ENV: &str = "HIVE_TOOL_GROUPS";
/// `HIVE_CAPABILITIES` env var injected by `meta::render_flake` when the
/// operator grants capabilities to this agent. Comma-separated
/// `hive_sh4re::Capability` `snake_case` names. Absent = no extra capabilities.
const CAPABILITIES_ENV: &str = "HIVE_CAPABILITIES";
/// Returns the MCP tool names (without `mcp__hyperhive__` prefix) that are
/// unlocked by the agent's current capability set. These are added to the
/// `--allowedTools` list so claude can call them without prompting, and
/// hive-c0re performs a second server-side capability check before executing.
fn allowed_capability_tools() -> Vec<String> {
let raw = match std::env::var(CAPABILITIES_ENV) {
Ok(v) if !v.trim().is_empty() => v,
_ => return vec![],
};
let mut tools = Vec::new();
for token in raw.split(',') {
let t = token.trim().to_ascii_lowercase();
match t.as_str() {
"read_host_journal" => tools.push("get_host_journal".to_owned()),
// infra_admin lets an agent restart hive infrastructure
// containers (hive-ci / hive-gateway / hive-forge) through the
// existing `restart` tool. Unlock it here so agents that hold
// the capability without the full `lifecycle` group can still
// call it; c0re re-checks the capability server-side and only
// honours infra-container names via this path.
"infra_admin" => tools.push("restart".to_owned()),
// manage_root_agent / query_agent_state don't expose new MCP
// tools: manage_root_agent gates existing lifecycle tools via
// topology enforcement; query_agent_state unlocks the `agent`
// field in get_loose_ends / count_pending_reminders /
// reminder_rollup (c0re enforces the cap server-side).
"manage_root_agent" | "query_agent_state" => {}
unknown => {
tracing::warn!(capability = %unknown, "unrecognised capability in HIVE_CAPABILITIES — skipped");
}
}
}
tools
}
/// Resolve the active tool groups for a harness session.
///
/// Reads `HIVE_TOOL_GROUPS` from the environment first. Each comma-separated
/// token is matched (case-insensitive) against the `ToolGroup` serde names
/// (`messaging`, `meta`, `inbox`, `lifecycle`, `approvals`, `scheduling`,
/// `diagnostics`, `execution`). Unrecognised tokens are logged and skipped.
/// Falls back to `AGENT_DEFAULT` when the env var is absent or empty.
fn effective_tool_groups() -> Vec<hive_sh4re::ToolGroup> {
let raw = match std::env::var(TOOL_GROUPS_ENV) {
Ok(v) if !v.trim().is_empty() => v,
_ => return hive_sh4re::ToolGroup::AGENT_DEFAULT.to_vec(),
};
let mut groups = Vec::new();
for token in raw.split(',') {
let t = token.trim().to_ascii_lowercase();
// Parse via serde_json (the canonical deserialization path).
if let Ok(g) =
serde_json::from_value::<hive_sh4re::ToolGroup>(serde_json::Value::String(t.clone()))
{
groups.push(g);
} else {
tracing::warn!(token = %t, "{TOOL_GROUPS_ENV}: unknown tool group, skipping");
}
}
if groups.is_empty() {
tracing::warn!(
"{TOOL_GROUPS_ENV} set but contained no recognised groups; \
falling back to AGENT_DEFAULT"
);
return hive_sh4re::ToolGroup::AGENT_DEFAULT.to_vec();
}
groups
}
/// Tool group an extra (out-of-process) MCP server is gated behind, if any.
///
/// Most `hyperhive.extraMcpServers` entries are ungated — available whenever
/// the operator declares them. The `bash` server is the exception: raw shell
/// execution is a privilege, so it is only exposed when the agent holds the
/// `Execution` tool group. Unlike the in-process hyperhive tools (gated at
/// dispatch) and the capability tools (re-checked server-side by hive-c0re),
/// an out-of-process server has **no** later enforcement point — once it is
/// in the claude MCP config the agent can call it. So this gate, applied at
/// config-render time, is the security boundary for those servers.
fn extra_server_required_group(server: &str) -> Option<hive_sh4re::ToolGroup> {
match server {
"bash" => Some(hive_sh4re::ToolGroup::Execution),
_ => None,
}
}
/// Whether an extra MCP server should be exposed to claude given the active
/// tool `groups`. A gated server (see [`extra_server_required_group`]) is
/// suppressed when the agent lacks its required group.
fn extra_server_enabled(server: &str, groups: &[hive_sh4re::ToolGroup]) -> bool {
extra_server_required_group(server).is_none_or(|required| groups.contains(&required))
}
#[cfg(test)]
mod extra_server_gate_tests {
use super::{extra_server_enabled, extra_server_required_group};
use hive_sh4re::ToolGroup;
#[test]
fn bash_is_gated_behind_execution() {
assert_eq!(
extra_server_required_group("bash"),
Some(ToolGroup::Execution)
);
// Suppressed without Execution, even if other groups are present.
assert!(!extra_server_enabled(
"bash",
&[ToolGroup::Messaging, ToolGroup::Inbox]
));
// Available once Execution is granted.
assert!(extra_server_enabled("bash", &[ToolGroup::Execution]));
}
#[test]
fn other_servers_are_ungated() {
assert_eq!(extra_server_required_group("matrix"), None);
assert_eq!(extra_server_required_group("scraper"), None);
// An ungated server is available regardless of (even empty) groups.
assert!(extra_server_enabled("matrix", &[]));
assert!(extra_server_enabled("scraper", &[ToolGroup::Messaging]));
}
}
/// MCP tools claude is allowed to call without prompting, derived from
/// the supplied tool groups. Adding a new `#[tool]` fn to a server impl
/// requires updating the matching `ToolGroup::tools()` slice in hive-sh4re
/// (single source of truth). See `docs/conventions.md::Tool groups`.
#[must_use]
pub fn allowed_mcp_tools(groups: &[hive_sh4re::ToolGroup]) -> Vec<String> {
// Collect all tool names, deduplicating while preserving order.
// Always-on tools (e.g. `set_status`) come first so they're present
// regardless of which groups the agent is granted — a misconfigured
// agent still has to be able to report its dashboard status.
let mut seen = std::collections::HashSet::new();
let mut out: Vec<String> = hive_sh4re::ToolGroup::ALWAYS_ON_TOOLS
.iter()
.copied()
.chain(groups.iter().flat_map(|g| g.tools().iter().copied()))
.filter(|t| seen.insert(*t))
.map(|t| format!("mcp__{SERVER_NAME}__{t}"))
.collect();
// Extra MCP servers declared via `hyperhive.extraMcpServers` in
// the agent's NixOS config. Each entry maps its `allowedTools`
// pattern list to `mcp__<server>__<pattern>` so claude can call
// them without per-tool operator approval. `["*"]` (the default)
// expands to `mcp__<server>__*` — every tool from that server.
for (server, spec) in load_extra_mcp() {
if server == SERVER_NAME || !extra_server_enabled(&server, groups) {
continue;
}
for pat in spec.allowed_tools {
out.push(format!("mcp__{server}__{pat}"));
}
}
out
}
/// Combined allow-list passed to `--allowedTools` (auto-approve) — covers
/// both the built-ins and the MCP surface.
#[must_use]
pub fn allowed_tools_arg() -> String {
let groups = effective_tool_groups();
// Base built-ins always present.
let mut all: Vec<String> = ALLOWED_BUILTIN_TOOLS
.iter()
.map(|s| (*s).to_owned())
.collect();
// Extra built-ins gated by tool groups (e.g. WebFetch/WebSearch via web_tools).
for group in &groups {
for tool in group.builtin_tools() {
if !all.iter().any(|t| t == *tool) {
all.push((*tool).to_owned());
}
}
}
all.extend(allowed_mcp_tools(&groups));
// Capability-gated MCP tools: added to --allowedTools when HIVE_CAPABILITIES
// includes the corresponding capability. hive-c0re performs a second
// server-side check, so this is a usability gate (no annoying prompts),
// not the security boundary.
for tool in allowed_capability_tools() {
all.push(format!("mcp__{SERVER_NAME}__{tool}"));
}
all.join(",")
}
/// Built-in tools list for `--tools` (which built-ins exist in this
/// session). Base set plus any group-gated built-ins (e.g.
/// `WebFetch`/`WebSearch` when the `web_tools` group is active).
#[must_use]
pub fn builtin_tools_arg() -> String {
let groups = effective_tool_groups();
let mut tools: Vec<&str> = ALLOWED_BUILTIN_TOOLS.to_vec();
for group in &groups {
for t in group.builtin_tools() {
if !tools.contains(t) {
tools.push(t);
}
}
}
tools.join(",")
}
/// Where the NixOS module writes the per-agent extra-MCP spec (see
/// `nix/templates/harness-base.nix`). Each entry becomes an additional
/// `mcpServers.<key>` block in the rendered claude config + a
/// `mcp__<key>__<tool>` pattern in `--allowedTools`.
const EXTRA_MCP_PATH: &str = "/etc/hyperhive/extra-mcp.json";
/// Where the NixOS module writes the per-agent send allow-list (see
/// `nix/templates/harness-base.nix`). Empty list = unrestricted (the
/// default). Non-empty list constrains `mcp__hyperhive__send`'s `to`
/// field; the manager is always implicitly permitted regardless of
/// the list contents.
const SEND_ALLOW_PATH: &str = "/etc/hyperhive/send-allow.json";
/// Enforce the per-agent send allow-list. Returns `Ok` when the
/// recipient is permitted (no list configured, `<parent>` sentinel
/// always allowed, or `to` is in the list); returns `Err(refusal)`
/// with a claude-readable string when blocked <20><><EFBFBD> the harness surfaces
/// the refusal as the tool result so claude knows the message didn't
/// land and can react (e.g. route via `<parent>` instead).
fn check_send_allowed(to: &str) -> Result<(), String> {
if to == hive_sh4re::PARENT_RECIPIENT {
// Always allow `<parent>` — the allow-list constrains peer
// chatter, not the structural reporting line; the operator
// can rewire who the parent IS via `set_parent` without
// having to remember to update the per-agent allow-list.
// The broker resolves the sentinel to the real parent label
// on the host side per topology.json (falls back to `operator`
// for root agents).
return Ok(());
}
let Ok(raw) = std::fs::read_to_string(SEND_ALLOW_PATH) else {
return Ok(()); // file missing → no policy configured → unrestricted
};
let allow: Vec<String> = match serde_json::from_str(&raw) {
Ok(v) => v,
Err(e) => {
tracing::warn!(
path = SEND_ALLOW_PATH,
error = ?e,
"send allow-list parse failed; falling back to unrestricted",
);
return Ok(());
}
};
if allow.is_empty() {
return Ok(()); // empty list = unrestricted (back-compat)
}
if allow.iter().any(|n| n == to) {
return Ok(());
}
Err(format!(
"send refused: recipient '{to}' not in hyperhive.allowedRecipients \
(configured in agent.nix). Allowed: {allow:?}. Your structural \
parent is always reachable route through `send(to: \"{}\", …)` \
if you need to reach someone outside the allow-list.",
hive_sh4re::PARENT_RECIPIENT
))
}
#[derive(Debug, serde::Deserialize)]
struct ExtraMcpServer {
command: String,
#[serde(default)]
args: Vec<String>,
#[serde(default)]
env: std::collections::BTreeMap<String, String>,
#[serde(default = "default_allowed_tools")]
#[serde(rename = "allowedTools")]
allowed_tools: Vec<String>,
}
fn default_allowed_tools() -> Vec<String> {
vec!["*".to_owned()]
}
/// Read + parse the extra-MCP spec. Returns an empty map when
/// the file is missing or unparsable (the agent has none configured,
/// or the file is malformed — both cases degrade to "no extra servers").
fn load_extra_mcp() -> std::collections::BTreeMap<String, ExtraMcpServer> {
let Ok(raw) = std::fs::read_to_string(EXTRA_MCP_PATH) else {
return std::collections::BTreeMap::new();
};
serde_json::from_str(&raw).unwrap_or_else(|e| {
tracing::warn!(
path = EXTRA_MCP_PATH,
error = ?e,
"extra-mcp spec parse failed; ignoring",
);
std::collections::BTreeMap::new()
})
}
/// Render the MCP config blob claude reads from `--mcp-config <path>`.
/// `agent_binary` is the path (or PATH-resolvable name) of the `hive-ag3nt`
/// executable; `socket` is the hyperhive per-agent socket bind-mounted into
/// the container (forwarded to the child as `--socket <path>`). Merges in
/// any extra MCP servers declared via `hyperhive.extraMcpServers` in the
/// agent's NixOS config.
#[must_use]
pub fn render_claude_config(agent_binary: &str, socket: &std::path::Path) -> String {
let mut servers = serde_json::Map::new();
// When the harness is configured to run the built-in server as a
// persistent streamable-http daemon (loopback port in
// `HYPERHIVE_MCP_HTTP_PORT`), point claude at the stable URL instead of
// respawning a fresh stdio child each turn. The URL survives the per-turn
// claude re-spawn, so there is no per-turn re-registration race for the
// hyperhive surface. Extra servers (matrix/bash) stay stdio bridges.
let hyperhive_entry = match std::env::var("HYPERHIVE_MCP_HTTP_PORT")
.ok()
.and_then(|p| p.trim().parse::<u16>().ok())
{
Some(port) => serde_json::json!({
"type": "http",
"url": format!("http://127.0.0.1:{port}/mcp"),
}),
None => serde_json::json!({
"command": agent_binary,
"args": ["--socket", socket.display().to_string(), "mcp"],
"env": {}
}),
};
servers.insert(SERVER_NAME.to_owned(), hyperhive_entry);
// Auto-inject HYPERHIVE_STATE_DIR so extra MCP servers can resolve the
// agent's durable state dir without the agent author hard-coding it.
// User-supplied env takes precedence — we only fill in the missing key.
let state_dir = crate::paths::state_dir();
// Gate tool-group-restricted extra servers (e.g. `bash` → `Execution`).
// This is the security boundary for them: an out-of-process server the
// agent isn't entitled to must not even appear in the MCP config, or the
// agent could call it directly (there is no later enforcement point).
let groups = effective_tool_groups();
for (name, mut spec) in load_extra_mcp() {
if name == SERVER_NAME {
tracing::warn!(
"extra MCP server name `{SERVER_NAME}` collides with the built-in surface; ignoring",
);
continue;
}
if !extra_server_enabled(&name, &groups) {
tracing::info!(
server = %name,
"extra MCP server suppressed: agent lacks the required tool group"
);
continue;
}
spec.env
.entry("HYPERHIVE_STATE_DIR".to_owned())
.or_insert_with(|| state_dir.display().to_string());
servers.insert(
name,
serde_json::json!({
"command": spec.command,
"args": spec.args,
"env": spec.env,
}),
);
}
let config = serde_json::json!({ "mcpServers": servers });
serde_json::to_string_pretty(&config).unwrap_or_else(|_| "{}".into())
}
#[cfg(test)]
mod tests {
use super::{IDLE_WAIT_HINT, SocketReply, format_recv};
use super::{SERVER_NAME, allowed_mcp_tools};
use hive_sh4re::ToolGroup;
use super::{IDLE_WAIT_HINT, format_recv};
#[test]
fn empty_recv_after_wait_appends_idle_hint() {
let out = format_recv(Ok(SocketReply::Messages(vec![])), true);
let out = format_recv(
Ok(hive_sh4re::Response::Messages { messages: vec![] }),
true,
);
assert!(out.starts_with("(empty)"));
assert!(out.contains(IDLE_WAIT_HINT));
}
#[test]
fn empty_recv_without_wait_has_no_hint() {
let out = format_recv(Ok(SocketReply::Messages(vec![])), false);
let out = format_recv(
Ok(hive_sh4re::Response::Messages { messages: vec![] }),
false,
);
assert_eq!(out, "(empty)");
}
fn qualified(tool: &str) -> String {
format!("mcp__{SERVER_NAME}__{tool}")
}
#[test]
fn set_status_present_with_no_groups() {
// An agent with zero tool groups (or any group set that omits
// `meta`) must still be able to report its dashboard status.
let tools = allowed_mcp_tools(&[]);
assert!(
tools.contains(&qualified("set_status")),
"set_status missing from empty-group allow-list: {tools:?}"
);
}
#[test]
fn set_status_present_without_meta_group() {
let tools = allowed_mcp_tools(&[ToolGroup::Messaging, ToolGroup::Inbox]);
assert!(tools.contains(&qualified("set_status")));
// get_agent_meta stays gated behind `meta` — only set_status is always-on.
assert!(!tools.contains(&qualified("get_agent_meta")));
}
#[test]
fn no_duplicate_set_status_when_meta_granted() {
let tools = allowed_mcp_tools(&[ToolGroup::Meta]);
let count = tools
.iter()
.filter(|t| **t == qualified("set_status"))
.count();
assert_eq!(count, 1, "set_status duplicated: {tools:?}");
assert!(tools.contains(&qualified("get_agent_meta")));
}
}

View file

@ -0,0 +1,439 @@
//! Claude launch-config layer: resolves the agent's tool-group / capability
//! set into the `--allowedTools` / `--tools` argument strings and renders the
//! `--mcp-config` blob claude reads at spawn (built-in hyperhive server +
//! any `hyperhive.extraMcpServers`). Pure config-string generation consumed by
//! [`crate::turn`] when it builds the claude command — it never touches the
//! running MCP server ([`crate::mcp`]). The `send` allow-list check
//! ([`check_send_allowed`]) lives here too since it's driven by the same
//! `/etc/hyperhive/*.json` operator config.
/// Name of the hyperhive MCP server inside claude's view. Claude prefixes
/// tools as `mcp__<this>__<tool>` (e.g. `mcp__hyperhive__send`).
pub const SERVER_NAME: &str = "hyperhive";
/// Built-in claude tools always present in every session. Anything not
/// in this list (or added by `extra_builtin_tools`) literally doesn't
/// exist in the session. Web egress (`WebFetch`/`WebSearch`) are
/// tool-group-gated (`web_tools`) — off by default. Nested agents
/// (`Task`) are intentionally omitted. `Bash` is disallowed — shell
/// execution goes through `mcp__bash__run` (background tasks
/// with structured output via `hive-bash-mcp`) instead of a raw interactive shell. `TodoWrite`
/// is omitted because the todo list lives in claude's in-process session
/// state and silently evaporates on /compact or session reset — agents
/// should plan in /state notes instead.
pub const ALLOWED_BUILTIN_TOOLS: &[&str] = &["Edit", "Glob", "Grep", "Read", "Write"];
/// Env var written by the meta renderer with a comma-separated list of
/// `hive_sh4re::ToolGroup` `snake_case` names (e.g. `"messaging,inbox,meta"`).
/// When present, the harness expands the groups into per-tool allow entries
/// instead of using the hardcoded flavor default. See `docs/conventions.md::Tool groups`.
const TOOL_GROUPS_ENV: &str = "HIVE_TOOL_GROUPS";
/// `HIVE_CAPABILITIES` env var injected by `meta::render_flake` when the
/// operator grants capabilities to this agent. Comma-separated
/// `hive_sh4re::Capability` `snake_case` names. Absent = no extra capabilities.
const CAPABILITIES_ENV: &str = "HIVE_CAPABILITIES";
/// Returns the MCP tool names (without `mcp__hyperhive__` prefix) that are
/// unlocked by the agent's current capability set. These are added to the
/// `--allowedTools` list so claude can call them without prompting, and
/// hive-c0re performs a second server-side capability check before executing.
fn allowed_capability_tools() -> Vec<String> {
let raw = match std::env::var(CAPABILITIES_ENV) {
Ok(v) if !v.trim().is_empty() => v,
_ => return vec![],
};
let mut tools = Vec::new();
for token in raw.split(',') {
let t = token.trim().to_ascii_lowercase();
match t.as_str() {
"read_host_journal" => tools.push("get_host_journal".to_owned()),
// infra_admin lets an agent restart hive infrastructure
// containers (hive-ci / hive-gateway / hive-forge) through the
// existing `restart` tool. Unlock it here so agents that hold
// the capability without the full `lifecycle` group can still
// call it; c0re re-checks the capability server-side and only
// honours infra-container names via this path.
"infra_admin" => tools.push("restart".to_owned()),
// manage_root_agent / query_agent_state don't expose new MCP
// tools: manage_root_agent gates existing lifecycle tools via
// topology enforcement; query_agent_state unlocks the `agent`
// field in get_loose_ends / count_pending_reminders /
// reminder_rollup (c0re enforces the cap server-side).
"manage_root_agent" | "query_agent_state" => {}
unknown => {
tracing::warn!(capability = %unknown, "unrecognised capability in HIVE_CAPABILITIES — skipped");
}
}
}
tools
}
/// Resolve the active tool groups for a harness session.
///
/// Reads `HIVE_TOOL_GROUPS` from the environment first. Each comma-separated
/// token is matched (case-insensitive) against the `ToolGroup` serde names
/// (`messaging`, `meta`, `inbox`, `lifecycle`, `approvals`, `scheduling`,
/// `diagnostics`, `execution`). Unrecognised tokens are logged and skipped.
/// Falls back to `AGENT_DEFAULT` when the env var is absent or empty.
fn effective_tool_groups() -> Vec<hive_sh4re::ToolGroup> {
let raw = match std::env::var(TOOL_GROUPS_ENV) {
Ok(v) if !v.trim().is_empty() => v,
_ => return hive_sh4re::ToolGroup::AGENT_DEFAULT.to_vec(),
};
let mut groups = Vec::new();
for token in raw.split(',') {
let t = token.trim().to_ascii_lowercase();
// Parse via serde_json (the canonical deserialization path).
if let Ok(g) =
serde_json::from_value::<hive_sh4re::ToolGroup>(serde_json::Value::String(t.clone()))
{
groups.push(g);
} else {
tracing::warn!(token = %t, "{TOOL_GROUPS_ENV}: unknown tool group, skipping");
}
}
if groups.is_empty() {
tracing::warn!(
"{TOOL_GROUPS_ENV} set but contained no recognised groups; \
falling back to AGENT_DEFAULT"
);
return hive_sh4re::ToolGroup::AGENT_DEFAULT.to_vec();
}
groups
}
/// Tool group an extra (out-of-process) MCP server is gated behind, if any.
///
/// Most `hyperhive.extraMcpServers` entries are ungated — available whenever
/// the operator declares them. The `bash` server is the exception: raw shell
/// execution is a privilege, so it is only exposed when the agent holds the
/// `Execution` tool group. Unlike the in-process hyperhive tools (gated at
/// dispatch) and the capability tools (re-checked server-side by hive-c0re),
/// an out-of-process server has **no** later enforcement point — once it is
/// in the claude MCP config the agent can call it. So this gate, applied at
/// config-render time, is the security boundary for those servers.
fn extra_server_required_group(server: &str) -> Option<hive_sh4re::ToolGroup> {
match server {
"bash" => Some(hive_sh4re::ToolGroup::Execution),
_ => None,
}
}
/// Whether an extra MCP server should be exposed to claude given the active
/// tool `groups`. A gated server (see [`extra_server_required_group`]) is
/// suppressed when the agent lacks its required group.
fn extra_server_enabled(server: &str, groups: &[hive_sh4re::ToolGroup]) -> bool {
extra_server_required_group(server).is_none_or(|required| groups.contains(&required))
}
#[cfg(test)]
mod extra_server_gate_tests {
use super::{extra_server_enabled, extra_server_required_group};
use hive_sh4re::ToolGroup;
#[test]
fn bash_is_gated_behind_execution() {
assert_eq!(
extra_server_required_group("bash"),
Some(ToolGroup::Execution)
);
// Suppressed without Execution, even if other groups are present.
assert!(!extra_server_enabled(
"bash",
&[ToolGroup::Messaging, ToolGroup::Inbox]
));
// Available once Execution is granted.
assert!(extra_server_enabled("bash", &[ToolGroup::Execution]));
}
#[test]
fn other_servers_are_ungated() {
assert_eq!(extra_server_required_group("matrix"), None);
assert_eq!(extra_server_required_group("scraper"), None);
// An ungated server is available regardless of (even empty) groups.
assert!(extra_server_enabled("matrix", &[]));
assert!(extra_server_enabled("scraper", &[ToolGroup::Messaging]));
}
}
/// MCP tools claude is allowed to call without prompting, derived from
/// the supplied tool groups. Adding a new `#[tool]` fn to a server impl
/// requires updating the matching `ToolGroup::tools()` slice in hive-sh4re
/// (single source of truth). See `docs/conventions.md::Tool groups`.
#[must_use]
pub fn allowed_mcp_tools(groups: &[hive_sh4re::ToolGroup]) -> Vec<String> {
// Collect all tool names, deduplicating while preserving order.
// Always-on tools (e.g. `set_status`) come first so they're present
// regardless of which groups the agent is granted — a misconfigured
// agent still has to be able to report its dashboard status.
let mut seen = std::collections::HashSet::new();
let mut out: Vec<String> = hive_sh4re::ToolGroup::ALWAYS_ON_TOOLS
.iter()
.copied()
.chain(groups.iter().flat_map(|g| g.tools().iter().copied()))
.filter(|t| seen.insert(*t))
.map(|t| format!("mcp__{SERVER_NAME}__{t}"))
.collect();
// Extra MCP servers declared via `hyperhive.extraMcpServers` in
// the agent's NixOS config. Each entry maps its `allowedTools`
// pattern list to `mcp__<server>__<pattern>` so claude can call
// them without per-tool operator approval. `["*"]` (the default)
// expands to `mcp__<server>__*` — every tool from that server.
for (server, spec) in load_extra_mcp() {
if server == SERVER_NAME || !extra_server_enabled(&server, groups) {
continue;
}
for pat in spec.allowed_tools {
out.push(format!("mcp__{server}__{pat}"));
}
}
out
}
/// Combined allow-list passed to `--allowedTools` (auto-approve) — covers
/// both the built-ins and the MCP surface.
#[must_use]
pub fn allowed_tools_arg() -> String {
let groups = effective_tool_groups();
// Base built-ins always present.
let mut all: Vec<String> = ALLOWED_BUILTIN_TOOLS
.iter()
.map(|s| (*s).to_owned())
.collect();
// Extra built-ins gated by tool groups (e.g. WebFetch/WebSearch via web_tools).
for group in &groups {
for tool in group.builtin_tools() {
if !all.iter().any(|t| t == *tool) {
all.push((*tool).to_owned());
}
}
}
all.extend(allowed_mcp_tools(&groups));
// Capability-gated MCP tools: added to --allowedTools when HIVE_CAPABILITIES
// includes the corresponding capability. hive-c0re performs a second
// server-side check, so this is a usability gate (no annoying prompts),
// not the security boundary.
for tool in allowed_capability_tools() {
all.push(format!("mcp__{SERVER_NAME}__{tool}"));
}
all.join(",")
}
/// Built-in tools list for `--tools` (which built-ins exist in this
/// session). Base set plus any group-gated built-ins (e.g.
/// `WebFetch`/`WebSearch` when the `web_tools` group is active).
#[must_use]
pub fn builtin_tools_arg() -> String {
let groups = effective_tool_groups();
let mut tools: Vec<&str> = ALLOWED_BUILTIN_TOOLS.to_vec();
for group in &groups {
for t in group.builtin_tools() {
if !tools.contains(t) {
tools.push(t);
}
}
}
tools.join(",")
}
/// Where the NixOS module writes the per-agent extra-MCP spec (see
/// `nix/templates/harness-base.nix`). Each entry becomes an additional
/// `mcpServers.<key>` block in the rendered claude config + a
/// `mcp__<key>__<tool>` pattern in `--allowedTools`.
const EXTRA_MCP_PATH: &str = "/etc/hyperhive/extra-mcp.json";
/// Where the NixOS module writes the per-agent send allow-list (see
/// `nix/templates/harness-base.nix`). Empty list = unrestricted (the
/// default). Non-empty list constrains `mcp__hyperhive__send`'s `to`
/// field; the manager is always implicitly permitted regardless of
/// the list contents.
const SEND_ALLOW_PATH: &str = "/etc/hyperhive/send-allow.json";
/// Enforce the per-agent send allow-list. Returns `Ok` when the
/// recipient is permitted (no list configured, `<parent>` sentinel
/// always allowed, or `to` is in the list); returns `Err(refusal)`
/// with a claude-readable string when blocked <20><><EFBFBD> the harness surfaces
/// the refusal as the tool result so claude knows the message didn't
/// land and can react (e.g. route via `<parent>` instead).
pub fn check_send_allowed(to: &str) -> Result<(), String> {
if to == hive_sh4re::PARENT_RECIPIENT {
// Always allow `<parent>` — the allow-list constrains peer
// chatter, not the structural reporting line; the operator
// can rewire who the parent IS via `set_parent` without
// having to remember to update the per-agent allow-list.
// The broker resolves the sentinel to the real parent label
// on the host side per topology.json (falls back to `operator`
// for root agents).
return Ok(());
}
let Ok(raw) = std::fs::read_to_string(SEND_ALLOW_PATH) else {
return Ok(()); // file missing → no policy configured → unrestricted
};
let allow: Vec<String> = match serde_json::from_str(&raw) {
Ok(v) => v,
Err(e) => {
tracing::warn!(
path = SEND_ALLOW_PATH,
error = ?e,
"send allow-list parse failed; falling back to unrestricted",
);
return Ok(());
}
};
if allow.is_empty() {
return Ok(()); // empty list = unrestricted (back-compat)
}
if allow.iter().any(|n| n == to) {
return Ok(());
}
Err(format!(
"send refused: recipient '{to}' not in hyperhive.allowedRecipients \
(configured in agent.nix). Allowed: {allow:?}. Your structural \
parent is always reachable route through `send(to: \"{}\", …)` \
if you need to reach someone outside the allow-list.",
hive_sh4re::PARENT_RECIPIENT
))
}
#[derive(Debug, serde::Deserialize)]
struct ExtraMcpServer {
command: String,
#[serde(default)]
args: Vec<String>,
#[serde(default)]
env: std::collections::BTreeMap<String, String>,
#[serde(default = "default_allowed_tools")]
#[serde(rename = "allowedTools")]
allowed_tools: Vec<String>,
}
fn default_allowed_tools() -> Vec<String> {
vec!["*".to_owned()]
}
/// Read + parse the extra-MCP spec. Returns an empty map when
/// the file is missing or unparsable (the agent has none configured,
/// or the file is malformed — both cases degrade to "no extra servers").
fn load_extra_mcp() -> std::collections::BTreeMap<String, ExtraMcpServer> {
let Ok(raw) = std::fs::read_to_string(EXTRA_MCP_PATH) else {
return std::collections::BTreeMap::new();
};
serde_json::from_str(&raw).unwrap_or_else(|e| {
tracing::warn!(
path = EXTRA_MCP_PATH,
error = ?e,
"extra-mcp spec parse failed; ignoring",
);
std::collections::BTreeMap::new()
})
}
/// Render the MCP config blob claude reads from `--mcp-config <path>`.
/// `agent_binary` is the path (or PATH-resolvable name) of the `hive-ag3nt`
/// executable; `socket` is the hyperhive per-agent socket bind-mounted into
/// the container (forwarded to the child as `--socket <path>`). Merges in
/// any extra MCP servers declared via `hyperhive.extraMcpServers` in the
/// agent's NixOS config.
#[must_use]
pub fn render_claude_config(agent_binary: &str, socket: &std::path::Path) -> String {
let mut servers = serde_json::Map::new();
// When the harness is configured to run the built-in server as a
// persistent streamable-http daemon (loopback port in
// `HYPERHIVE_MCP_HTTP_PORT`), point claude at the stable URL instead of
// respawning a fresh stdio child each turn. The URL survives the per-turn
// claude re-spawn, so there is no per-turn re-registration race for the
// hyperhive surface. Extra servers (matrix/bash) stay stdio bridges.
let hyperhive_entry = match std::env::var("HYPERHIVE_MCP_HTTP_PORT")
.ok()
.and_then(|p| p.trim().parse::<u16>().ok())
{
Some(port) => serde_json::json!({
"type": "http",
"url": format!("http://127.0.0.1:{port}/mcp"),
}),
None => serde_json::json!({
"command": agent_binary,
"args": ["--socket", socket.display().to_string(), "mcp"],
"env": {}
}),
};
servers.insert(SERVER_NAME.to_owned(), hyperhive_entry);
// Auto-inject HYPERHIVE_STATE_DIR so extra MCP servers can resolve the
// agent's durable state dir without the agent author hard-coding it.
// User-supplied env takes precedence — we only fill in the missing key.
let state_dir = crate::paths::state_dir();
// Gate tool-group-restricted extra servers (e.g. `bash` → `Execution`).
// This is the security boundary for them: an out-of-process server the
// agent isn't entitled to must not even appear in the MCP config, or the
// agent could call it directly (there is no later enforcement point).
let groups = effective_tool_groups();
for (name, mut spec) in load_extra_mcp() {
if name == SERVER_NAME {
tracing::warn!(
"extra MCP server name `{SERVER_NAME}` collides with the built-in surface; ignoring",
);
continue;
}
if !extra_server_enabled(&name, &groups) {
tracing::info!(
server = %name,
"extra MCP server suppressed: agent lacks the required tool group"
);
continue;
}
spec.env
.entry("HYPERHIVE_STATE_DIR".to_owned())
.or_insert_with(|| state_dir.display().to_string());
servers.insert(
name,
serde_json::json!({
"command": spec.command,
"args": spec.args,
"env": spec.env,
}),
);
}
let config = serde_json::json!({ "mcpServers": servers });
serde_json::to_string_pretty(&config).unwrap_or_else(|_| "{}".into())
}
#[cfg(test)]
mod tests {
use super::{SERVER_NAME, allowed_mcp_tools};
use hive_sh4re::ToolGroup;
fn qualified(tool: &str) -> String {
format!("mcp__{SERVER_NAME}__{tool}")
}
#[test]
fn set_status_present_with_no_groups() {
// An agent with zero tool groups (or any group set that omits
// `meta`) must still be able to report its dashboard status.
let tools = allowed_mcp_tools(&[]);
assert!(
tools.contains(&qualified("set_status")),
"set_status missing from empty-group allow-list: {tools:?}"
);
}
#[test]
fn set_status_present_without_meta_group() {
let tools = allowed_mcp_tools(&[ToolGroup::Messaging, ToolGroup::Inbox]);
assert!(tools.contains(&qualified("set_status")));
// get_agent_meta stays gated behind `meta` — only set_status is always-on.
assert!(!tools.contains(&qualified("get_agent_meta")));
}
#[test]
fn no_duplicate_set_status_when_meta_granted() {
let tools = allowed_mcp_tools(&[ToolGroup::Meta]);
let count = tools
.iter()
.filter(|t| **t == qualified("set_status"))
.count();
assert_eq!(count, 1, "set_status duplicated: {tools:?}");
assert!(tools.contains(&qualified("get_agent_meta")));
}
}

View file

@ -11,8 +11,6 @@
//! plugin specs resolve against current index data. Marketplace update
//! failures are non-fatal — stale index is better than no install attempt.
use std::path::Path;
use tokio::process::Command;
const PLUGINS_PATH: &str = "/etc/hyperhive/claude-plugins.json";
@ -103,8 +101,7 @@ async fn update_marketplaces() {
/// notification, see `Surface::send_to_parent`). Wire-agnostic: the
/// caller picks the recipient via the same `<parent>` sentinel that
/// failure-notify uses everywhere else.
pub async fn install_configured(socket: &Path) -> Vec<String> {
let _ = socket; // Reserved for future telemetry; currently unused.
pub async fn install_configured() -> Vec<String> {
let Ok(raw) = tokio::fs::read_to_string(PLUGINS_PATH).await else {
return Vec::new();
};

View file

@ -2,7 +2,7 @@
//! HTML-comment markers gating role-specific blocks; this module
//! assembles the final prompt (always "agent" role — there is only one
//! role). Marker grammar + placeholder substitution rules in
//! `docs/turn-loop.md::On-boot files` (`claude-system-prompt.md`).
//! `docs/turn-loop/claude-invocation.md::On-boot files` (`claude-system-prompt.md`).
use std::path::{Path, PathBuf};
@ -16,7 +16,7 @@ use anyhow::{Context, Result};
/// [`hive_sh4re::assets::prompt_template`]
/// (`$HIVE_ASSETS_DIR/prompts/system.md`). Substitution placeholders +
/// marker grammar documented in
/// `docs/turn-loop.md::On-boot files` (`claude-system-prompt.md`).
/// `docs/turn-loop/claude-invocation.md::On-boot files` (`claude-system-prompt.md`).
#[must_use]
pub fn render(
template: &str,
@ -55,14 +55,12 @@ pub fn render(
}
}
/// Walk `template` line-by-line. Inside a `<!-- role:X -->` block,
/// suppress all lines unless `X == target`. Marker lines themselves are
/// always elided from the output. Unbalanced openers (no matching
/// closer) hold the suppression state until end-of-file. A mismatched
/// closer (`<!-- /role:manager -->` inside a `role:agent` block) is
/// elided from the output but does NOT reset the active role — keeps
/// the suppression conservative so a typo can't dump wrong-flavor
/// content.
/// Walk `template` line-by-line, stripping `<!-- role:X -->` /
/// `<!-- /role:X -->` marker lines and suppressing the lines inside a block
/// unless `X == target`. A close marker ends the current block. Production
/// `system.md` carries no markers today (single agent role — see the module
/// doc), so this is effectively a passthrough; the marker grammar stays wired
/// for a future manager / multi-role prompt.
fn filter_role_blocks(template: &str, target: &str) -> String {
let mut out = String::with_capacity(template.len());
// None = outside any block; Some(role) = inside role-tagged block.
@ -73,19 +71,11 @@ fn filter_role_blocks(template: &str, target: &str) -> String {
active_role = Some(role);
continue;
}
if let Some(close_role) = parse_close_marker(trimmed) {
if active_role == Some(close_role) {
active_role = None;
}
// Mismatched close: elide the marker line but keep the
// active role intact so wrong-flavor content stays gated.
if parse_close_marker(trimmed).is_some() {
active_role = None;
continue;
}
let include = match active_role {
None => true,
Some(role) => role == target,
};
if include {
if active_role.is_none_or(|role| role == target) {
out.push_str(line);
out.push('\n');
}
@ -245,46 +235,6 @@ shared closer
assert_eq!(parse_close_marker("just text"), None);
}
#[test]
fn mismatched_close_keeps_active_role() {
// `<!-- role:agent -->` block with a stray `<!-- /role:manager -->`
// closer inside: the manager-tagged close must NOT pop the agent
// gate, else manager-target output would leak the agent block's
// text (or vice-versa). Stray marker line itself is still elided.
let template = "shared\n\
<!-- role:agent -->\n\
agent line 1\n\
<!-- /role:manager -->\n\
agent line 2\n\
<!-- /role:agent -->\n\
shared end\n";
let manager = filter_role_blocks(template, "manager");
// Both agent lines stay gated out for the manager target; the
// stray close didn't accidentally pop the role. Stray marker
// itself elided from the output.
assert!(!manager.contains("agent line 1"));
assert!(!manager.contains("agent line 2"));
assert!(!manager.contains("<!--"));
assert!(manager.contains("shared"));
assert!(manager.contains("shared end"));
// Agent target still sees both lines (matched closer pops at the end).
let agent = filter_role_blocks(template, "agent");
assert!(agent.contains("agent line 1"));
assert!(agent.contains("agent line 2"));
assert!(agent.contains("shared end"));
}
#[test]
fn unbalanced_opener_suppresses_until_eof() {
// Stray opener with no closer — content stays suppressed for
// the wrong-role target right through to end-of-file. Real-
// file safety net: a typo in a closer doesn't accidentally
// dump wrong-flavor content into the active prompt.
let template = "shared\n<!-- role:manager -->\nm-only\nstill m-only\n";
let agent = filter_role_blocks(template, "agent");
assert_eq!(agent, "shared\n");
}
#[test]
fn render_substitutes_label_and_pronouns() {
// Real template's first agent line — keeps the renderer

View file

@ -5,7 +5,7 @@
use crate::events::Bus;
use crate::mcp::REDELIVERY_HINT;
use crate::turn::TurnOutcome;
use crate::turn::{TurnError, TurnOutcome};
use crate::turn_stats::TurnStatRow;
/// Assemble the per-turn wake prompt string. The role/tools/etc. live in the
@ -100,12 +100,12 @@ pub fn build_row(args: TurnRowArgs<'_>) -> TurnStatRow {
serde_json::to_string(&tool_calls).ok()
};
let (result_kind, note) = match outcome {
TurnOutcome::Ok => ("ok", None),
TurnOutcome::Compacted => ("compacted", None),
TurnOutcome::PromptTooLong => ("prompt_too_long", None),
TurnOutcome::RateLimited => ("rate_limited", None),
TurnOutcome::AuthFailed => ("auth_failed", None),
TurnOutcome::Failed(e) => ("failed", Some(format!("{e:#}"))),
Ok(false) => ("ok", None),
Ok(true) => ("compacted", None),
Err(TurnError::PromptTooLong) => ("prompt_too_long", None),
Err(TurnError::RateLimited) => ("rate_limited", None),
Err(TurnError::AuthFailed) => ("auth_failed", None),
Err(TurnError::Failed(e)) => ("failed", Some(format!("{e:#}"))),
};
let wake_from = if wake_from.starts_with("bash-task-") {
"bash-task".to_owned()

View file

@ -42,6 +42,7 @@ impl Window {
"7d" => Self::Week,
"30d" => Self::Month,
"all" => Self::All,
// Default (incl. `label()`'s own canonical `"24h"`/`"1d"`).
_ => Self::Day,
}
}
@ -229,11 +230,15 @@ fn empty_snapshot(window: Window) -> Snapshot {
}
fn snapshot(path: &Path, window: Window) -> Result<Snapshot> {
// Read-only open so an in-flight writer (the harness's own
// turn_stats sink) never blocks us and we can't corrupt the db
// via a query bug.
// Read-only open so we can't corrupt the db via a query bug.
let conn = Connection::open_with_flags(path, OpenFlags::SQLITE_OPEN_READ_ONLY)
.with_context(|| format!("open {} read-only", path.display()))?;
// turn_stats is rollback-journal (not WAL): a read landing while the
// harness's own sink is mid-INSERT gets SQLITE_BUSY, which propagates up
// and blanks the whole stats page. Wait out the brief write instead —
// matches hive-c0re's host-side reader (`hive_stats::read_agent`).
conn.busy_timeout(std::time::Duration::from_millis(500))
.with_context(|| format!("set busy_timeout on {}", path.display()))?;
let now = now_secs();
// Fixed windows look back a constant span; `all` starts at the earliest
// recorded turn (`MIN(started_at)`, falling back to `now` on an empty

View file

@ -12,7 +12,7 @@ use hive_claude::{Config, InfiniteSession, PercentPolicy, Sink};
use serde_json::Value;
use crate::events::{Bus, LiveEvent};
use crate::mcp;
use crate::mcp_config;
// Hive-enforced claude settings ship at `/etc/claude-code/managed-settings.json`
// (wired in `nix/templates/harness-base.nix` from the `prompts/claude-settings.json`
@ -92,8 +92,8 @@ pub struct TurnFiles {
}
impl TurnFiles {
/// Write all three files into the per-agent runtime dir alongside
/// `socket`. Idempotent — overwrites whatever was there.
/// Write the two per-turn files (MCP config + system prompt) into the
/// agent's config dir. Idempotent — overwrites whatever was there.
///
/// # Errors
///
@ -121,7 +121,7 @@ pub async fn write_mcp_config(socket: &Path) -> Result<PathBuf> {
let exe = std::env::current_exe()
.ok()
.map_or_else(|| "hive".into(), |p| p.display().to_string());
let body = mcp::render_claude_config(&exe, socket);
let body = mcp_config::render_claude_config(&exe, socket);
tokio::fs::write(&path, body).await?;
tracing::info!(path = %path.display(), "wrote claude MCP config");
Ok(path)
@ -139,29 +139,39 @@ pub async fn write_system_prompt(socket: &Path, label: &str) -> Result<PathBuf>
crate::prompt::write_system_prompt(socket, label).await
}
/// One claude turn's outcome. The harness uses this to decide whether to
/// transparently kick off a compaction and retry.
/// One claude turn's outcome: `Ok(compacted)` on success, or a [`TurnError`]
/// the serve loop must act on. The `compacted` bool is `true` when a
/// compaction ran this turn (reactively on overflow, or proactively per the
/// policy — or an operator `/compact` at turn end); it's recorded as
/// `result_kind = "compacted"` in turn stats so the stats page can distinguish
/// those turns. Both `Ok(true)` and `Ok(false)` are ack'd; the error cases
/// each map to a distinct serve-loop action (see [`emit_turn_end`] and the
/// `hive` serve loop).
pub type TurnOutcome = std::result::Result<bool, TurnError>;
/// The ways a turn can end without a usable result. Each is deliberately *not*
/// a generic failure — the serve loop reacts to each differently (requeue,
/// park, escalate).
#[derive(Debug)]
pub enum TurnOutcome {
Ok,
/// Turn completed and proactive context-size compaction fired afterwards.
/// Treated like `Ok` for ack and failure-notification purposes; recorded
/// as `result_kind = "compacted"` in turn stats so the stats page can
/// distinguish normal turns from turns that triggered a compaction.
Compacted,
pub enum TurnError {
/// claude saw "Prompt is too long" and even a reactive compact + retry
/// (inside [`InfiniteSession::run`]) couldn't bring it back under the
/// window. Rare; the serve loop treats it like `Ok` (acks the turn).
/// window. Rare. [`drive_turn`] archives the session (so the next turn
/// starts fresh) and the serve loop requeues the in-flight message, which
/// redelivers into that fresh session — the wake prompt itself is tiny, so
/// the overflow was the accumulated context, which the archive clears.
PromptTooLong,
/// The Anthropic API refused the request due to a rate limit, per-account
/// usage cap, or exhausted credit balance. The serve loop should park for
/// `rate_limit_sleep_secs()` and retry — NOT bubble up as a crash.
/// `rate_limit_sleep_secs()` and requeue — NOT bubble up as a crash.
RateLimited,
/// The Anthropic API rejected the request with 401 (OAuth session
/// expired or revoked). The serve loop should flip the container
/// into `needs_login_idle` and stop driving turns until the
/// operator re-auths via the per-agent web UI.
AuthFailed,
/// A hard failure with no recovery — the serve loop escalates it to the
/// parent (`send_to_parent`).
Failed(anyhow::Error),
}
@ -226,8 +236,10 @@ fn compact_percent() -> u8 {
if env_u64("HIVE_COMPACT_WATERMARK_TOKENS") == Some(0) {
return 0;
}
let pct = env_u64("HIVE_COMPACT_WATERMARK_PERCENT").unwrap_or(u64::from(DEFAULT_COMPACT_PERCENT));
u8::try_from(pct.min(100)).unwrap_or(DEFAULT_COMPACT_PERCENT)
let pct =
env_u64("HIVE_COMPACT_WATERMARK_PERCENT").unwrap_or(u64::from(DEFAULT_COMPACT_PERCENT));
// `min(100)` is ≤ 100, so this `try_from` is infallible.
u8::try_from(pct.min(100)).expect("value clamped to <= 100 fits in u8")
}
/// The agent's durable session type: the constant-title [`InfiniteSession`]
@ -309,24 +321,36 @@ pub async fn drive_turn(
text: format!("created fresh session titled \"{}\"", session_title()),
});
}
if progress.compacted {
TurnOutcome::Compacted
} else {
TurnOutcome::Ok
}
Ok(progress.compacted)
}
Err(e) => error_to_turn(e),
};
if matches!(outcome, Err(TurnError::PromptTooLong)) {
// The lib already compacted + retried and the session is still over the
// window. Archive it here (session lifecycle stays hive-side) so the
// requeued message — handled by the serve loop — redelivers into a
// fresh session that fits.
bus.emit(LiveEvent::Note {
text: "context still over the window after compaction — archiving session so the \
retried message starts fresh"
.into(),
});
archive_session(bus);
return Err(TurnError::PromptTooLong);
}
// Operator `/compact` (`POST /api/compact`) deferred to the turn boundary:
// run it now that the turn is done, so it works mid-turn rather than only
// when the agent is idle. Only on a healthy turn — no point spawning a
// compaction after a rate-limited / auth-failed / crashed one.
if bus.take_compact() && matches!(outcome, TurnOutcome::Ok | TurnOutcome::Compacted) {
if bus.take_compact() && outcome.is_ok() {
bus.emit(LiveEvent::Note {
text: "operator: /compact — running at turn end".into(),
});
// Reflect `Compacting` in the UI like the idle path (`run_pending_compact`)
// does; the serve loop resets to `Idle` once this turn returns.
bus.set_state(crate::events::TurnState::Compacting);
let _ = session.compact(&config, &sink).await;
return TurnOutcome::Compacted;
return Ok(true);
}
outcome
}
@ -374,28 +398,35 @@ fn maybe_auto_reset(bus: &Bus) {
/// semantics stay consistent across every agent role.
pub fn emit_turn_end(bus: &Bus, outcome: &TurnOutcome) {
match outcome {
TurnOutcome::Ok | TurnOutcome::Compacted | TurnOutcome::PromptTooLong => {
Ok(_) => {
bus.emit(LiveEvent::TurnEnd {
ok: true,
note: None,
});
tracing::info!("turn finished");
}
TurnOutcome::RateLimited => {
Err(TurnError::PromptTooLong) => {
bus.emit(LiveEvent::TurnEnd {
ok: false,
note: Some("context too long after compaction — session archived, retrying".into()),
});
tracing::warn!("turn prompt-too-long; archived session and requeueing");
}
Err(TurnError::RateLimited) => {
bus.emit(LiveEvent::TurnEnd {
ok: false,
note: Some("rate limited — parking until quota resets".into()),
});
tracing::warn!("turn rate-limited");
}
TurnOutcome::AuthFailed => {
Err(TurnError::AuthFailed) => {
bus.emit(LiveEvent::TurnEnd {
ok: false,
note: Some("authentication failed (401) — waiting for re-login".into()),
});
tracing::warn!("turn auth-failed (401)");
}
TurnOutcome::Failed(e) => {
Err(TurnError::Failed(e)) => {
let note = format!("{e:#}");
bus.emit(LiveEvent::TurnEnd {
ok: false,
@ -489,8 +520,8 @@ fn claude_config(bus: &Bus, files: &TurnFiles) -> Config {
system_prompt_file: Some(files.system_prompt.clone()),
mcp_config: Some(files.mcp_config.clone()),
strict_mcp_config: true,
tools: Some(mcp::builtin_tools_arg()),
allowed_tools: Some(mcp::allowed_tools_arg()),
tools: Some(mcp_config::builtin_tools_arg()),
allowed_tools: Some(mcp_config::allowed_tools_arg()),
add_dirs,
..Config::default()
}
@ -503,11 +534,13 @@ fn claude_config(bus: &Bus, files: &TurnFiles) -> Config {
fn error_to_turn(err: hive_claude::Error) -> TurnOutcome {
use hive_claude::Error;
match err {
Error::PromptTooLong => TurnOutcome::PromptTooLong,
Error::RateLimited => TurnOutcome::RateLimited,
Error::AuthFailed => TurnOutcome::AuthFailed,
Error::SessionNotFound => TurnOutcome::Ok,
other => TurnOutcome::Failed(other.into()),
Error::PromptTooLong => Err(TurnError::PromptTooLong),
Error::RateLimited => Err(TurnError::RateLimited),
Error::AuthFailed => Err(TurnError::AuthFailed),
// A resume-miss the lib couldn't self-heal is benign — treat it as a
// clean (non-compacted) turn; the next turn creates the session fresh.
Error::SessionNotFound => Ok(false),
other => Err(TurnError::Failed(other.into())),
}
}
@ -603,4 +636,3 @@ fn archive_session(bus: &Bus) {
}
}
}

View file

@ -1,8 +1,14 @@
//! Per-turn analytics sink. One sqlite row per claude turn captures:
//! identity (`model`, `wake_from`, `result_kind`), timing (`started_at`,
//! `ended_at`, `duration_ms`), cost (token counts), behaviour (tool-call
//! count + per-tool breakdown), and post-turn snapshot metrics
//! (`open_threads_count`, `open_reminders_count`).
//! `ended_at`, `duration_ms`), cost (token counts), and behaviour (tool-call
//! count + per-tool breakdown).
//!
//! **Captured but not yet read** (written every turn, no reader today —
//! kept for a future chart / backfill, not consumed by `stats::snapshot`
//! or the host rollup): `tool_call_count` (the snapshot recomputes tool
//! totals from `tool_call_breakdown_json` instead), `open_threads_count` +
//! `open_reminders_count` (planned: a loose-ends-over-time trend), and
//! `note` (failure detail for `result_kind = "failed"`).
//!
//! Lives next to `hyperhive-events.sqlite` in the agent's state dir
//! so the host-side state vacuum sweep can reach both. Schema is
@ -103,14 +109,19 @@ pub struct TurnStatRow {
pub last_output_tokens: u64,
pub last_cache_read_input_tokens: u64,
pub last_cache_creation_input_tokens: u64,
/// Captured, not yet read — the snapshot recomputes tool totals from
/// `tool_call_breakdown_json` (see the module doc).
pub tool_call_count: u64,
/// Per-tool breakdown as JSON: `{"Read":12,"Bash":3,...}`. None
/// when no tools were called (saves a sqlite write of `"{}"`).
pub tool_call_breakdown_json: Option<String>,
/// Post-turn loose-ends snapshot. Captured, not yet read — planned to
/// feed a loose-ends-over-time trend on the stats page.
pub open_threads_count: Option<u64>,
pub open_reminders_count: Option<u64>,
/// `"ok" | "failed" | "prompt_too_long"`.
pub result_kind: &'static str,
/// Failure detail for `result_kind = "failed"`. Captured, not yet read.
pub note: Option<String>,
/// FK to `sessions.id` for the fresh claude session this turn belongs
/// to. `None` on pre-capture rows (and when the stats db couldn't mint
@ -278,7 +289,10 @@ impl TurnStats {
last_input_tokens, last_output_tokens,
last_cache_read_input_tokens, last_cache_creation_input_tokens
FROM turn_stats
ORDER BY started_at DESC
-- `id` (AUTOINCREMENT) is monotonic with insertion, so this is
-- the most-recently-inserted row even among same-second turns
-- (which `started_at DESC` would order arbitrarily).
ORDER BY id DESC
LIMIT 1",
[],
|row| {

View file

@ -8,9 +8,7 @@ use axum::{
};
use serde::Deserialize;
use crate::client;
use super::{AppState, SOCKET_FETCH_TIMEOUT, error_response};
use super::{AppState, error_response};
#[derive(Deserialize)]
pub(super) struct SendForm {
@ -25,41 +23,28 @@ pub(super) async fn post_send(
if body.is_empty() {
return error_response(StatusCode::BAD_REQUEST, "send: `body` required");
}
let result = match tokio::time::timeout(
SOCKET_FETCH_TIMEOUT,
client::request::<_, hive_sh4re::Response>(
&state.socket,
&hive_sh4re::Request::OperatorMsg { body },
),
)
.await
{
Ok(Ok(hive_sh4re::Response::Ok)) => Ok(()),
Ok(Ok(hive_sh4re::Response::Err { message })) => Err(message),
Ok(Ok(other)) => Err(format!("unexpected response: {other:?}")),
Ok(Err(e)) => Err(format!("transport: {e:#}")),
Err(_) => Err("timed out — hive-c0re busy, retry".to_owned()),
};
match result {
match super::broker_request(&state.socket, &hive_sh4re::Request::OperatorMsg { body }).await {
// 200 instead of 303 → the client doesn't refetch /api/state.
// The operator message becomes a broker `Sent` (already shown
// server-side in the dashboard); on the agent side, the
// resulting `TurnStart` SSE event drives the terminal + the
// inbox row gets consumed by the time `TurnEnd` fires the
// existing turn-end refresh.
Ok(()) => (axum::http::StatusCode::OK, "ok").into_response(),
Err(e) => error_response(
Ok(hive_sh4re::Response::Ok) => (axum::http::StatusCode::OK, "ok").into_response(),
Ok(hive_sh4re::Response::Err { message }) => error_response(
StatusCode::INTERNAL_SERVER_ERROR,
&format!("send failed: {e}"),
&format!("send failed: {message}"),
),
Ok(other) => error_response(
StatusCode::INTERNAL_SERVER_ERROR,
&format!("send failed: unexpected response: {other:?}"),
),
Err(e) => super::broker_error_response(&e, "send"),
}
}
pub(super) async fn post_cancel_turn(State(state): State<AppState>) -> Response {
let out = tokio::process::Command::new("pkill")
.args(["-INT", "claude"])
.output()
.await;
let out = super::sigint_claude().await;
let note = match out {
Ok(o) if o.status.success() => "operator: /cancel — sent SIGINT to claude".to_owned(),
Ok(o) if o.status.code() == Some(1) => {

View file

@ -74,17 +74,9 @@ pub(super) async fn post_login_cancel(State(state): State<AppState>) -> Response
(axum::http::StatusCode::OK, "ok").into_response()
}
/// OAuth credential filenames inside `paths::claude_dir()`. Wiping
/// only these (and not the rest of `~/.claude/`) preserves session
/// history so `claude --continue` keeps working after a fresh login.
/// Rationale + the previous wholesale-wipe shape we replaced live in
/// [`docs/web-ui/agent.md::Per-agent endpoints`](../../../docs/web-ui/agent.md)
/// (the `/api/logout` bullet).
const CRED_FILE_NAMES: &[&str] = &[".credentials.json", "mcp-needs-auth-cache.json"];
/// Operator-driven `/logout`: SIGINT claude, delete the credential
/// files in `CRED_FILE_NAMES`, flip `LoginState::NeedsLogin`. The
/// turn loop's next iteration parks into `wait_for_login` which
/// files (via [`crate::login::clear_session`]), flip `LoginState::NeedsLogin`.
/// The turn loop's next iteration parks into `wait_for_login` which
/// resumes when a fresh credentials file appears via `/login/code`.
/// Always returns 200 with a body describing what happened. See
/// [`docs/web-ui/agent.md::Per-agent endpoints`](../../../docs/web-ui/agent.md)
@ -92,36 +84,20 @@ const CRED_FILE_NAMES: &[&str] = &[".credentials.json", "mcp-needs-auth-cache.js
/// preservation invariants.
pub(super) async fn post_logout(State(state): State<AppState>) -> Response {
// Step 1: SIGINT claude (best-effort, matches `post_cancel_turn`).
let _ = tokio::process::Command::new("pkill")
.args(["-INT", "claude"])
.output()
.await;
// Step 2: delete OAuth credential files only — preserve session
// history files alongside them.
let _ = super::sigint_claude().await;
// Step 2: delete OAuth credential files only — login::clear_session owns
// the file set and preserves session-history files alongside them.
let dir = crate::paths::claude_dir();
let mut warnings: Vec<String> = Vec::new();
let mut wiped: Vec<&str> = Vec::new();
for name in CRED_FILE_NAMES {
let path = dir.join(name);
match tokio::fs::remove_file(&path).await {
Ok(()) => wiped.push(name),
Err(e) if e.kind() == std::io::ErrorKind::NotFound => {
// Already gone — operator clicked /logout while
// already logged out, or the file simply didn't exist
// for this agent. Idempotent.
}
Err(e) => warnings.push(format!("{name}: {e}")),
}
}
let wipe_summary = if wiped.is_empty() {
let cleared = crate::login::clear_session(&dir).await;
let wipe_summary = if cleared.wiped.is_empty() {
"no credential files present (already logged out)".to_owned()
} else {
format!("wiped {}", wiped.join(", "))
format!("wiped {}", cleared.wiped.join(", "))
};
let warn_suffix = if warnings.is_empty() {
let warn_suffix = if cleared.warnings.is_empty() {
String::new()
} else {
format!(" (warnings: {})", warnings.join("; "))
format!(" (warnings: {})", cleared.warnings.join("; "))
};
// Step 3: flip LoginState + emit Note. Turn loop sees the flip on
// its next iteration and parks into wait_for_login.

View file

@ -259,6 +259,17 @@ fn read_gui_vnc_port() -> Option<u16> {
std::env::var("HIVE_GUI_VNC_PORT").ok()?.parse().ok()
}
/// SIGINT any running `claude` process in this container (best-effort). Shared
/// by `/api/cancel` and `/api/logout`. Returns the `pkill` `Output` so callers
/// can inspect the exit status (0 = signalled, 1 = no process matched) or
/// ignore it.
async fn sigint_claude() -> std::io::Result<std::process::Output> {
tokio::process::Command::new("pkill")
.args(["-INT", "claude"])
.output()
.await
}
fn error_response(status: StatusCode, message: &str) -> Response {
// Plain text — JS app surfaces in `alert()`, HTML wrapping would just
// be noise. Status is per-caller: 400 for bad input, 409 for a
@ -267,3 +278,52 @@ fn error_response(status: StatusCode, message: &str) -> Response {
// in its alert, so a benign "busy, retry" must not read as a 500.
(status, message.to_owned()).into_response()
}
/// Why a deadline-bounded broker request via the per-agent socket didn't
/// yield a response. Kept distinct so action handlers pick the right status
/// code (see [`broker_error_response`]) while decorative fetches `.ok()` both.
enum BrokerError {
/// Outran [`SOCKET_FETCH_TIMEOUT`] — hive-c0re is busy or stalled. A
/// retryable state conflict (→ 409), not a server fault.
Timeout,
/// The socket transport itself failed (connect / encode / decode).
Transport(anyhow::Error),
}
/// Issue a broker request over the per-agent socket, bounded by
/// [`SOCKET_FETCH_TIMEOUT`] so a busy or stalled hive-c0re degrades the
/// response instead of hanging it. Callers match the returned [`Response`]
/// variant themselves; the error side distinguishes a retryable timeout from
/// a transport failure. This is the one shared broker-call scaffold — every
/// web-UI handler that talks to the broker goes through it.
async fn broker_request(
socket: &Path,
req: &hive_sh4re::Request,
) -> std::result::Result<hive_sh4re::Response, BrokerError> {
match tokio::time::timeout(
SOCKET_FETCH_TIMEOUT,
crate::client::request::<_, hive_sh4re::Response>(socket, req),
)
.await
{
Ok(Ok(resp)) => Ok(resp),
Ok(Err(e)) => Err(BrokerError::Transport(e)),
Err(_) => Err(BrokerError::Timeout),
}
}
/// Map a [`BrokerError`] to an operator-facing error response: a timeout is a
/// retryable "busy" conflict (409), a transport failure is a 500. `action`
/// prefixes the message (e.g. `"send"`, `"get_loose_ends"`).
fn broker_error_response(err: &BrokerError, action: &str) -> Response {
match err {
BrokerError::Timeout => error_response(
StatusCode::CONFLICT,
&format!("{action}: timed out — hive-c0re busy, retry"),
),
BrokerError::Transport(e) => error_response(
StatusCode::INTERNAL_SERVER_ERROR,
&format!("{action}: transport: {e:#}"),
),
}
}

View file

@ -3,11 +3,10 @@
use axum::extract::State;
use serde::Serialize;
use crate::client;
use crate::login::LoginState;
use crate::login_session::drop_if_finished;
use super::{AppState, SOCKET_FETCH_TIMEOUT};
use super::AppState;
pub(super) async fn api_state(State(state): State<AppState>) -> axum::Json<StateSnapshot> {
// Capture seq *before* any reads so the dedupe contract is
@ -37,10 +36,7 @@ pub(super) async fn api_state(State(state): State<AppState>) -> axum::Json<State
let inbox = recent_inbox(&state.socket).await;
let (turn_state, turn_state_since) = state.bus.state_snapshot();
let model = state.bus.model();
let context_window_tokens = state
.bus
.api_context_window()
.unwrap_or_else(|| crate::events::context_window_tokens(&model));
let context_window_tokens = state.bus.effective_context_window(&model);
let ctx_usage = state.bus.last_ctx_usage();
let cost_usage = state.bus.last_cost_usage();
let effort = state.bus.effort();
@ -88,10 +84,7 @@ pub(super) async fn api_dashboard_state(
let (status_text, status_set_at) = read_own_status();
let rate_limited = state.bus.is_rate_limited();
let model = state.bus.model();
let context_window_tokens = state
.bus
.api_context_window()
.unwrap_or_else(|| crate::events::context_window_tokens(&model));
let context_window_tokens = state.bus.effective_context_window(&model);
// Full context-window size = input + cache-read + cache-creation. Using
// raw `input_tokens` here reported only the *uncached* sliver, which is
// ~0 once prompt caching kicks in — so every card showed `ctx·0k`. Match
@ -371,57 +364,24 @@ struct ExtraLink {
/// failure — the inbox section is decorative, not authoritative.
async fn recent_inbox(socket: &std::path::Path) -> Vec<hive_sh4re::InboxRow> {
const LIMIT: u64 = 30;
// Deadline-bounded: `/api/state` must render even when hive-c0re is
// busy — an empty inbox section beats a hung snapshot.
match tokio::time::timeout(
SOCKET_FETCH_TIMEOUT,
client::request::<_, hive_sh4re::Response>(
socket,
&hive_sh4re::Request::Recent { limit: LIMIT },
),
)
.await
{
Ok(Ok(hive_sh4re::Response::Recent { rows })) => rows,
// Deadline-bounded (via `broker_request`): `/api/state` must render even
// when hive-c0re is busy — an empty inbox section beats a hung snapshot.
match super::broker_request(socket, &hive_sh4re::Request::Recent { limit: LIMIT }).await {
Ok(hive_sh4re::Response::Recent { rows }) => rows,
_ => Vec::new(),
}
}
/// Fetch reminder activity stats from the broker via the per-agent /
/// manager socket. Returns None on any transport / decode failure — the
/// stats are decorative, not authoritative.
pub(super) async fn fetch_reminder_stats(
socket: &std::path::Path,
window_secs: u64,
) -> Option<hive_sh4re::ReminderStats> {
match tokio::time::timeout(
SOCKET_FETCH_TIMEOUT,
client::request::<_, hive_sh4re::Response>(
socket,
&hive_sh4re::Request::ReminderRollup {
since_secs: window_secs,
agent: None,
},
),
)
.await
{
Ok(Ok(hive_sh4re::Response::ReminderRollup(stats))) => Some(stats),
_ => None,
}
}
/// Read `HIVE_AVAILABLE_MODELS` (comma-separated short names injected by
/// `services.hyperhive.availableModels`) and return the parsed list.
/// Falls back to `["haiku", "sonnet", "opus"]` when the env var is absent
/// or resolves to an empty list after trimming.
fn available_models() -> Vec<String> {
const DEFAULT: &[&str] = &["haiku", "sonnet", "opus"];
let raw = match std::env::var("HIVE_AVAILABLE_MODELS") {
Ok(v) if !v.trim().is_empty() => v,
_ => return DEFAULT.iter().map(ToString::to_string).collect(),
};
let models: Vec<String> = raw
// Absent / empty / all-whitespace env all funnel to the single
// emptiness check below — no separate up-front guard needed.
let models: Vec<String> = std::env::var("HIVE_AVAILABLE_MODELS")
.unwrap_or_default()
.split(',')
.map(|s| s.trim().to_string())
.filter(|s| !s.is_empty())

View file

@ -5,10 +5,7 @@ use axum::http::StatusCode;
use axum::response::{IntoResponse, Response};
use serde::Deserialize;
use crate::client;
use super::state::fetch_reminder_stats;
use super::{AppState, SOCKET_FETCH_TIMEOUT, error_response};
use super::{AppState, error_response};
#[derive(Deserialize)]
pub(super) struct StatsQuery {
@ -29,6 +26,27 @@ pub(super) async fn api_stats(
axum::Json(snapshot)
}
/// Fetch reminder activity stats from the broker via the per-agent / manager
/// socket. Returns None on any transport / decode failure — the stats are
/// decorative, not authoritative.
async fn fetch_reminder_stats(
socket: &std::path::Path,
window_secs: u64,
) -> Option<hive_sh4re::ReminderStats> {
match super::broker_request(
socket,
&hive_sh4re::Request::ReminderRollup {
since_secs: window_secs,
agent: None,
},
)
.await
{
Ok(hive_sh4re::Response::ReminderRollup(stats)) => Some(stats),
_ => None,
}
}
/// Proxy this agent's loose-ends list via the per-agent socket. The
/// web UI surfaces the result as a collapsible section in the page
/// so the operator can see at a glance what's pending against the
@ -37,42 +55,25 @@ pub(super) async fn api_stats(
/// the `mcp__hyperhive__get_loose_ends` tool sees from inside the
/// container.
pub(super) async fn api_loose_ends(State(state): State<AppState>) -> Response {
let loose_ends: Vec<hive_sh4re::LooseEnd> = match tokio::time::timeout(
SOCKET_FETCH_TIMEOUT,
client::request::<_, hive_sh4re::Response>(
&state.socket,
&hive_sh4re::Request::GetLooseEnds { agent: None },
),
match super::broker_request(
&state.socket,
&hive_sh4re::Request::GetLooseEnds { agent: None },
)
.await
{
Ok(Ok(hive_sh4re::Response::LooseEnds { loose_ends })) => loose_ends,
Ok(Ok(hive_sh4re::Response::Err { message })) => {
return error_response(
StatusCode::INTERNAL_SERVER_ERROR,
&format!("get_loose_ends: {message}"),
);
Ok(hive_sh4re::Response::LooseEnds { loose_ends }) => {
axum::Json(serde_json::json!({ "loose_ends": loose_ends })).into_response()
}
Ok(Ok(other)) => {
return error_response(
StatusCode::INTERNAL_SERVER_ERROR,
&format!("unexpected response: {other:?}"),
);
}
Ok(Err(e)) => {
return error_response(
StatusCode::INTERNAL_SERVER_ERROR,
&format!("transport: {e:#}"),
);
}
Err(_) => {
return error_response(
StatusCode::CONFLICT,
"get_loose_ends: timed out — hive-c0re busy, retry",
);
}
};
axum::Json(serde_json::json!({ "loose_ends": loose_ends })).into_response()
Ok(hive_sh4re::Response::Err { message }) => error_response(
StatusCode::INTERNAL_SERVER_ERROR,
&format!("get_loose_ends: {message}"),
),
Ok(other) => error_response(
StatusCode::INTERNAL_SERVER_ERROR,
&format!("get_loose_ends: unexpected response: {other:?}"),
),
Err(e) => super::broker_error_response(&e, "get_loose_ends"),
}
}
/// `GET /api/bash-tasks` — snapshot of this agent's in-flight bash tasks.

View file

@ -56,15 +56,22 @@ pub(super) async fn events_stream(
) -> Sse<impl Stream<Item = Result<Event, Infallible>>> {
tracing::info!("sse: client subscribed");
let rx = state.bus.subscribe();
// Drop a "hello" note into the bus so every new subscriber sees at
// least one event immediately and can clear the connecting placeholder.
state.bus.emit(crate::events::LiveEvent::Note {
text: "live stream attached".into(),
});
let stream = BroadcastStream::new(rx).filter_map(|res| {
// Prime THIS connection with a one-off "hello" so it can clear the
// connecting placeholder immediately. Injected into this subscriber's own
// stream rather than emitted to the bus — a bus emit would spam every
// already-connected client with a spurious note each time anyone opens
// the stream.
let hello = Event::default().data(
serde_json::to_string(&crate::events::LiveEvent::Note {
text: "live stream attached".into(),
})
.unwrap_or_default(),
);
let live = BroadcastStream::new(rx).filter_map(|res| {
let ev = res.ok()?;
let json = serde_json::to_string(&ev).ok()?;
Some(Ok(Event::default().data(json)))
});
let stream = tokio_stream::once(Ok(hello)).chain(live);
Sse::new(stream).keep_alive(KeepAlive::default())
}

View file

@ -58,7 +58,10 @@ impl Claude {
})?;
if let Some(mut stdin) = child.stdin.take() {
stdin.write_all(prompt.as_bytes()).await.map_err(Error::Stdin)?;
stdin
.write_all(prompt.as_bytes())
.await
.map_err(Error::Stdin)?;
// Best-effort flush/close; claude sees EOF and starts the turn.
stdin.shutdown().await.ok();
}

View file

@ -41,7 +41,11 @@ impl CompactionPolicy for PercentPolicy {
if self.percent == 0 {
return false;
}
let Some(window) = usage.context_window.or(self.default_window).filter(|&w| w > 0) else {
let Some(window) = usage
.context_window
.or(self.default_window)
.filter(|&w| w > 0)
else {
return false;
};
usage.context_tokens.saturating_mul(100) >= u64::from(self.percent) * window

View file

@ -4,7 +4,9 @@ use std::sync::Mutex;
use serde_json::Value;
use crate::{Attach, Claude, CompactionPolicy, Config, Error, Result, SessionStore, Sink, Telemetry};
use crate::{
Attach, Claude, CompactionPolicy, Config, Error, Result, SessionStore, Sink, Telemetry,
};
/// A named claude session that outlives the model's context window by
/// compacting itself. Bundles the three things a durable session needs: