An agent can reach VictoriaLogs only through the gateway, and since the
machine query route landed the way to read it has been to hand-roll a
client_credentials token request and a curl, per query. This is the CLI
that closes that: `swarm-logs query '<LogsQL>'`, matched log lines on
stdout, so the answer pipes into grep like any other command's.
Built to the plan posted on the tracker thread: own crate, own
docs/tools reference generated off the clap tree, `query` as the one
verb, and the JSON error body surfaced on a non-200 rather than
swallowed. No `tail`: streaming is a different endpoint with a different
response shape, and folding it in here would be a fatter scope than the
ask.
Minting the token is NOT implemented here — swarm-queue-client already
owns the client_credentials request, its error type and its CA handling,
and a token-endpoint fix has to be findable in one place. What this crate
adds is the agent-shaped half: the client id arrives as a *file* beside
the secret, so nothing outside nix/agent-modules/queue.nix spells
`hive-<name>-agent` twice. That is the same problem hive-agent's
swarm_queue module solves, and swarm-logs/src/auth.rs is its `decide`
restated over this binary's inputs.
⚠️ The plan named one thing to verify empirically before calling the auth
settled: whether authelia's bearer policy for the logs vhost accepts the
agent client's audience. Measured from inside a container: it does not.
The client minted a token fine but with `aud: []` and `scp: []`, asking
for the logs URL as an audience answered `invalid_target`, and presenting
the audience-less token to the gateway answered a bare 401. So
swarm-authelia.nix's agentClients gains `authelia.bearer.authz` and the
query URL as a second audience — authelia authorises a bearer token by
the URL being requested, and that URL is now one binding read by three
places rather than three spellings of one address.
The URL reaches an agent the same way its queue coordinates do: computed
on the host (a container cannot derive a gateway address), forwarded by
hive_c0re::meta into the container's option set, and consumed by a new
agent module that installs the binary *wrapped* with its coordinates —
the shape swarm-controller.nix installs swarmctl in. Gated on the queue
credential as well as on the URL: a binary that can only answer 401 is
worse than no binary, because an agent reads a 401 as "no logs", which is
the exact confusion the store's machine route was added to end.
The rewrite in #4445 flattened this section's new-functionality-vs-
existing-functionality split into unconditional rules, because it had
been anchored to the current-state table that PR deleted. Restore the
split on its own terms, per #4445's own description of the rule: new
functionality has to match the target shape immediately; existing
functionality can move step by step, provided each step moves toward
the target shape.
Refs #4445
Two lines still framed the invariant as one file total: line 19 ('plus
exactly one file') and line 21 ('the only credential on disk') implied
a single mTLS cert page-wide, contradicting the earlier fix that made
line 26 say 'every host has at least one'. Agents and swarm-level
services each need their own identity certificate, so a host running
several holds several. Restate both as per-identity, keep the target
shape (store + certs on disk, nothing else) unchanged.
Also relax the touching-a-credential rule: moving a secret into bao
without resolving its renewal strategy in the same change is
acceptable while the migration to this shape is in progress, provided
a follow-up issue is opened to settle renewal.
Refs #4445
The citation for the nix option description was 529-531, but the quoted
phrase 'A path, never a value.' appears on line 533. The correct range
529-533 includes both the 'operator places by hand' claim and the
'path, never a value' assertion that the doc attributes to that file.
Refs #4445
Drop the current-state table and the prose describing how the swarm
currently breaks the target (unrotated secrets persisted to disk outside
the store, no agent store identity, the appservice token's second
minter). Scope the path-never-value rule to the mTLS client certificate
— the one credential a nix option can carry, now that everything else
goes through bao. Trim the migration checklist to unconditional rules
for any credential-touching change, since there's no current-state table
left for the old existing-functionality carve-outs to reference.
Refs #4445
The target section described the target as a shorter list of secret files —
the intro promised "what changes is how many paths there are", and the
per-secret contract pointed at the current-state table's file-shaped columns.
That is today's shape with fewer rows, not the target.
Rewrite it around the invariant: no secret the store holds is ever written to
disk. The mTLS client certificate is the sole credential on disk, and the
section now says why — it authenticates the puller to the store, so it is the
one credential that cannot be fetched from it. Per secret the target still
states minter, reader and renewal, but the reader is now a process pulling a
store path at runtime under its own certificate, not a path being read off
disk.
The current-state section and all of its citations are untouched: the current
state genuinely is files, and that contrast is what the page is for.
Two vale rewrites in c33cc441 changed meaning, not just wording:
- "on track to become" asserted a trajectory the page's own inventory
contradicts (no bao identity anywhere, all renewals NONE, all
secrets plaintext). Restored intent: "what it should become".
- "most often misread" claimed an observed reader frequency nobody
measured. Dropped the claim about readers entirely; states the fact
directly instead.
Both stay vale-clean; no suppressions.
Fixes the Microsoft.Contractions/write-good.ThereIs/alex.Condescending/
Microsoft.Auto errors vale flagged in CI (prose-lint-errors, required):
contractions throughout, no sentences opening with 'There is', dropped
the 'auto-mints' hyphenation and the 'easily' condescension flag, and
renamed the credentials table's 'what it is' header to 'kind' (also
flagged once styles were resynced). No technical claim or file:line
citation changed. Same fix applied to the one new line in
docs/README.md's pointer to this page.
Audited all 22 file:line citations against source on this head, per
argus's PR review. Fixed:
- matrix_account.rs writer cell pointed at path-computation line 163;
the actual store.write() call is at 189-197 (now anchored to the
put_matrix_account function name too).
- swarm-secret-publisher.nix appservice-token minter cited the
"put on every run" comment (280-286) instead of the mint itself
(271-278).
- glue-matrix-bao-token.nix persisted-token cell cited the
registration-restamp comment (186-188) instead of the printf+chmod
(180-181).
- glue-queue-agent-credential.nix persisted-files cell missed the
secret's chmod 0600 at 219, citing 224-231; corrected to 218-224.
- swarm-otel.nix persisted-file cell cited 763-767, one line short of
the chmod 0400 at 768; corrected to 764-768.
- hive-priv/src/main.rs:368,413 (forge/github token cell) landed in
unrelated DestroyContainer/ListContainers/WriteAgentGithubToken match
arms; replaced with the write_forge_token/write_github_token function
names as a greppable, edit-resistant anchor.
- swarm-bao.nix:527-528 (the "operator places by hand" quote) missed
the quoted text, which starts at 529; corrected to 529-531.
15 of 22 already resolved exactly; no claim lacked supporting code, so
nothing was removed. 22 of 22 citations checked.
The swarm's credential docs say where every file lives. They do not say
whether it should be a file at all, so a discussion about direction has
had nothing to point at and each one re-derived the same table.
This page carries that table with the three columns the target contract
is written in — minter, reader, renewal — plus the column the target is
really about: whether the value is persisted outside the store. Stating
it flatly is the point. All four stored families are plaintext files on
disk, the appservice token twice; every renewal cell reads NONE; no agent
container holds a store identity at all, so the hive reads on its behalf
and writes a file in; and the appservice token has a second, uncoordinated
local mint that can diverge from the published one.
The target section is marked as a target throughout, because its first
line is the one most easily misread as fact: every host needing a store
mTLS certificate is where this is going, while today only the store's own
host auto-mints and swarm-bao.nix calls it the credential an operator
places by hand everywhere else.
The progressive-enhancement rule is stated as a table of questions a
reviewer applies to a pull request rather than as prose, since a rule
nobody can check is a preference. New functionality matches the target
immediately; existing functionality moves stepwise, and the questions
distinguish a step from churn.
Indexed from the docs root and the swarm README. It supersedes
secrets.md when the migration completes — at which point that file is
deleted and this one moves into its place.
Add attribution-trailer lint row to the ci.md checks table (modeled
on the vale-errors row: not yet a required check) so the doc doesn't
assert stale completeness once attribution-trailers merges. Delete
the dead CI_COMMIT_REF_NAME elif in check-attribution-trailers.sh —
byte-identical to the else arm, never exercised, and its comment
overclaimed GitLab/Forgejo support nothing here can test.
Refs #4432
`glue-matrix-bao-token.nix` has read
`secret/swarm/hives/<hive>/matrix/appservice-token` since it landed, but
nothing ever wrote that path. The store was empty in every deployment, so
every read degraded to "keep what activation minted" and each hive stayed
the origin of a value the swarm has to agree on — two hives never
converged.
`swarm-secret-publish` is now the producer. It already holds a store
identity, already writes under the hive prefix, and already runs per
hive in the roster, so the mint is a third loop beside the two OIDC
copies rather than a second shape of this unit.
Idempotence comes from a record of its own, not from the store: this
principal is granted `create`/`update` with no `read`, so it cannot ask
whether a hive already has a token. It keeps what it minted under
`StateDirectory=` (0700 dir, 0600 file) and mints only when that file is
missing or empty; the `put` runs every time, because re-putting the same
bytes changes nothing for a reader while a mint whose publish failed must
not be left as a token this host holds and no hive can reach.
The token never becomes a nix literal and never reaches argv: the mint
redirects into a file, and the publish hands bao `value=@<path>` so bao
opens it itself — the same handling the OIDC loops use.
`hive-matrix.nix`'s activation mint stays as the genuine first-boot
fallback. It already fires only when the token file is absent, so it
cannot clobber a value the store delivered; `hs_token` has no swarm half
and is still minted there for real.
Refs #4402
Nothing reads it any more: hive-c0re creates accounts as the hive's
appservice, so the mint, the host file, the bind mount, the
`LoadCredential` entry and tuwunel's `registration_token_file` all go.
⚠️ `allow_registration` has to go to `false` in the same change, and not
as hardening. tuwunel refuses to START when registration is allowed with
no token configured — it demands
`yes_i_am_very_very_sure_…_open_registration_…` instead — so dropping the
token and leaving the flag true is not a lax homeserver, it is one that
does not boot. The flag is checked only for requests arriving without an
appservice token, so hive-c0re provisions exactly as before and everyone
else is refused outright.
The swarm secret store keeps its role, repointed at the credential that
replaced the token (`swarm/hives/<hive>/matrix/appservice-token`). Its
unit now also re-runs hive-matrix's own registration renderer after
writing the file: the token is half an agreement, and a registration
still naming the previous value authenticates nobody. The renderer is
shared through an internal option rather than copied, so the
registration's shape has one home.
Both spellings of `registrationTokenFile` become
`mkRemovedOptionModule` with a message naming what replaced them. A hive
that never set the option — the default — is unaffected; one that pinned
it fails to evaluate with instructions instead of a silent no-op.
An upgraded hive needs no intervention: the activation script has both
halves in place before the homeserver restarts, existing agents keep the
tokens their devices already hold, and the old token file is left on
disk read by nothing. docs/integrations/matrix.md spells the path out.
Refs #4402
A subagent inherited the parent's built-in tool list, which correctly has
no `Bash` -- the agent reaches a shell through the `bash` MCP server, not
the built-in. Subagents get no such server, so the intersection was empty
and they could not run a command at all: no commits, no pushes, no gates.
Add `subagent_builtin_tools_for`/`_arg`, which reuse the shared resolver
and append `Bash` only when `Execution` -- the group that gates the `bash`
MCP server -- is present. Only the subagent spawn path calls them, so the
harness's own `--tools`/`--allowedTools` are unchanged.
The capability transfers; the mechanism does not.
Refs #4422
The comments and docs around the empty-`--tools` assert stated a
mechanism: that an empty value parses as *unset* and therefore grants
MORE built-ins than omitting the flag. That claim came from a
measurement, and the installed `claude --help` (2.1.268) says the
opposite in as many words — `Use "" to disable all tools`. One of the two
is wrong and this repo is not the place that settles it.
The rule the code enforces is right under both readings, which is
precisely why it should not be argued from either: we never emit an empty
`--tools`, because what an empty one means is release-dependent and
nothing here wants a subagent with no built-in tools regardless. So the
prose now states the rule and the ambiguity, and asserts nothing about
`""` in either direction — replacing the claim with its opposite would be
the same mistake sourced differently.
No behaviour change: the assert, the resolution and every test are
untouched.
Two test comments also gained the limit they were quietly missing.
`no_spawned_tool_escapes_the_session` hardcodes ten names, so it is a
canary for tools someone thought to list, not a guarantee — the
guarantee is the subset property, which covers tools nobody has invented
yet. And `a_subagent_gets_no_builtin_its_parent_lacks` resolves both
sides in one process off one env var, so it catches a code divergence but
not the two real systemd units disagreeing about `HIVE_TOOL_GROUPS` —
which is what they did until the previous commit, and is not a thing a
unit test can reach.
Refs #4416
`build_config` spawned a subagent with `--dangerously-skip-permissions`
and no `--tools` at all, so it got claude's entire built-in set —
`SendMessage` and `ListAgents` (message peers, or the operator, as its
parent), `Task*` including `TaskStop`, which takes an *agent* id and so
reaches clean outside the run, `Cron*`, `RemoteTrigger` and
`EnterWorktree`/`ExitWorktree`. None of that is part of "do this bounded
task in this directory", and none of it is something the parent agent
itself can do: the harness has always passed `--tools`.
Pass the same one. The value comes from
`hive_sh4re::permissions::builtin_tools_arg()` — literally the function
the harness resolves its own session with — so the subagent's set is the
parent's set, `HIVE_TOOL_GROUPS` and all. That inheritance is the
requirement, not an implementation detail: a hardcoded subagent list
would hand `WebFetch`/`WebSearch` to the subagent of an agent without the
`web_tools` group, which is a privilege escalation, and would drift from
the parent's list the first time anyone added a tool to either.
`--tools` is the real gate: it holds under
`--dangerously-skip-permissions`, unlike `--allowedTools`, which only
auto-approves prompts. It does not filter MCP tools, so the
`goal_reached`/`need_help` signal surface is deliberately unnamed in it
and survives on `--strict-mcp-config` alone.
`build_config`'s doc comment claimed `strict_mcp_config` was *the* safety
property and that a subagent got "nothing implicit and nothing more".
That was false for built-ins, and is what hid this gap for as long as it
did; it now says which flag covers which half and that neither
substitutes for the other.
An empty `--tools` value parses as *unset* and grants more than omitting
the flag, so an empty resolution can only be a bug — `build_config`
asserts against it and a test pins the non-emptiness alongside the
subset-of-parent property.
Refs #4416
The `--tools` list a harness session gets is not a constant: the base set
plus whatever the agent's `HIVE_TOOL_GROUPS` add (today, `web_tools` →
`WebFetch`/`WebSearch`). That resolution lived in `hive-agent`'s
`mcp_config`, which is fine while the harness is the only thing that
spawns a `claude` — and it is not: `hive-subagent-mcp` spawns one too.
`hive-agent` is binary-only (no `src/lib.rs`, no lib target), so nothing
can depend on it to reach `builtin_tools_arg`. The alternative to a
shared home is a second list in the subagent daemon, which diverges on
the first tool anyone adds to either — and diverging upward is a
subagent holding a built-in its parent does not have.
So move the base list, the `HIVE_TOOL_GROUPS` parse and the resolution
into `hive_sh4re::permissions`, beside `ToolGroup` — whose
`builtin_tools()` was already half of the answer. `hive-agent`
re-exports them, so `mcp_config::builtin_tools_arg()` still reads the
same at the call site, and `allowed_tools_arg` now derives its built-in
half from the same function rather than repeating the merge loop.
Behaviour is unchanged. The parse is `strum::EnumString` rather than a
`serde_json::from_value` round-trip through a `Value::String`: same
`snake_case` names (a test pins the two derives against each other),
without `hive-sh4re` needing `serde_json` outside its dev-dependencies.
It is now a pure function of its input, so the fallbacks are testable
without mutating the environment — which under edition 2024 is `unsafe`
and racy across a test binary's threads.
Refs #4416
`goal_reached`/`need_help` took the session name as a tool argument, so
identity was an assertion by the caller and the only guard on it was
`occupancy()` — "does that name have a turn in flight", which two
concurrently running siblings both satisfy for each other. A subagent
could stop its sibling's run by naming it.
Identity moves into the URL. Each spawned run is minted an unguessable
token (`Uuid::new_v4`, the OS CSPRNG), the URL carrying it goes into that
one subagent's own `--mcp-config`, and the route resolves it back to a
session before dispatching to a handler bound to that session. Neither
tool takes a `name` any more: a subagent has no field in which to name a
sibling, and a sibling's name — which a brief may well mention — is not a
token.
One route with a path parameter, not a route per session: the `Router` is
built once at startup and subagents come and go for the daemon's whole
life. An unminted or revoked token gets a bare 404, the same answer either
way, so nothing enumerates. A run's token is revoked when the run ends
(`finish_turn`) or when a call never reached a spawn.
Two things fall out of that:
- the config file becomes one per session. A single shared path was
already a race between two `start`s; with a per-session URL in it, the
loser would read the winner's identity.
- `occupancy()` stops being the identity guard and is gone from the signal
path entirely rather than kept "just in case" — a revoked token can't
reach it, and it never answered the question it was standing in for.
It still backs `status`, which is what it was always actually for.
Refs #4403
Refs #4413
`start` takes an optional `goal`. With one set a session stops being a
single turn: when a turn ends and nothing has said to stop, the daemon
spawns another turn re-prompting the subagent toward that goal, up to
`max_turns` (default 5, per-session). Without a goal nothing changes —
one turn, one todo, same as before.
Four things end a run, each recorded distinctly and reported by `status`:
the turn ending with no goal, `goal_reached`, `need_help`, and the turn
cap. The last says so out loud rather than stopping quietly — the todo
states the harness limit was reached and the goal was never reported
reached. Every stop extends the done message rather than replacing it,
and lands in the session's report file when it has one. The path is
never inferred: it comes from `start`'s `report_file` or from the
subagent naming where it wrote.
`goal_reached` and `need_help` are the subagent's own, served on a second
route (`/signal/mcp`) that carries those two tools and nothing else, so
reporting on a run can't become starting one. `goal_reached` is built as
a label, never a gate: it is self-reported by a subagent that has just
been re-prompted with "you haven't reached the goal", which is exactly
the incentive to claim it — the same failure class as a build report
asserting the tests pass. Every surface that renders it says so.
`need_help` is the blocking signal, and shows in `status` as its own
state so a parent polling it sees the block without reading a file.
`status` also carries `turn N of M`: with 4330's last-event age, that
separates working from wedged from out of turns off one answer.
Two bugs the new tests caught: a `tokio::fs::File` was dropped without
flushing, so the report line was written to nothing, and the plain idle
answer dropped the turn counter.
Also documents `await_resume`'s third case — a closed channel with no
send, which fails open the same as `Underway` — per argus on #4411.
Refs #4403
`continue` returned "started" the instant `Claude::spawn` handed back a
pid, and a resume that matched nothing only surfaced later, as an
end-of-turn todo. By then the caller had moved on believing it had a
running subagent.
A pid is proof enough for `start`, which creates its session: the spawn
succeeding is the whole story. It is not proof for a resume — claude
exits non-zero a fraction of a second *after* the process exists. So
`continue` now waits for the first real answer and reports a miss as its
own `Err`, carrying claude's message and the directory searched.
The wait ends on whichever comes first, so a successful `continue` pays
no fixed delay: the turn's first non-terminal stream event settles it at
about the same moment a miss's exit would have. Measured on this box:
14 runs of the driver's own invocation against a missing session took
550-1087 ms spawn to exit, and a healthy turn's first event lands at
roughly 500 ms. The five-second cap is ~4.6x the slowest miss and is only
ever reached by a child that neither speaks nor exits.
The underway signal reads the event's kind, not its content: a missed
resume is not silent — it emits a terminal `result` event and stderr
before exiting — so "any sink callback" would have reported every miss as
a successful start. Liveness still counts all three callbacks.
The end-of-turn todo is unchanged for every failure later in the turn;
the only one it no longer repeats is the miss the caller was just handed.
Refs #4405
A session cannot be moved between directories, so a hint reading
"pass dir" could be misread as pointing an existing session at any
directory. Say instead that dir names the directory the session was
started in.
Refs #4405
`status` could only answer running / starting / idle / killed / none,
because every turn ran against `&NoopSink` and the whole stream-json
stream was discarded. "Running" describes a wedged subagent exactly as
well as a busy one, leaving a caller to tell them apart from `ps` output
and CPU-time deltas.
So the daemon now keeps a `name -> last_event_at` clock, bumped by
`LivenessSink` on every line of every stream — stream-json events, plain
stdout chatter and stderr alike — and `status` reports its age on a
running answer: a few seconds means working, an age climbing into the
minutes with no end-of-turn todo means wedged. Nothing is read out of the
content; classifying *what* a subagent is doing is a separate question
and waits on its own driver work. In memory with the rest of this
daemon's state, dropped when the turn ends, no persistence.
The clock is seeded at the spawn rather than at the first line, so a
subagent that wedged before emitting anything still reports a climbing
age rather than no age at all — the case an age is worth most in.
Separately, `continue`'s existence pre-check is gone. It could only
repeat the lookup `Claude::spawn` was about to do, and its message —
"no session named `x` exists" — was false in the common failure: the
session existed, just not under the claude home + cwd `build_store`
resolved from. claude's own `--resume` is the authority and exits
non-zero (`does not match any session title`) rather than quietly
starting a fresh session, so the turn fails on its own. `classify_end`
appends the one fact the CLI's message lacks — the directory searched:
claude error: no session matched the requested id or title (searched
<claude_home> for cwd <cwd>; if the session was started elsewhere,
pass `dir`)
The `dirs` map's durability is untouched; whether to persist it stays an
open operator decision.
Module doc, `docs/tools/subagent.md`, the `continue`/`status` tool
descriptions and the `base:claude-subagents` skill all updated — including
`continue`'s `dir` doc, which said "the daemon remembers it" without
saying that a restart is both when it forgets and when you most want it.
Refs #4330
Refs #4405
`swarmctl agent create <name> --hive <hive>` POSTs `/api/agents` to
swarm-controller over the daemon's unix socket and prints the queued
job's node id.
It deliberately does not wait. The endpoint queues a DAG whose last node
*publishes* a deploy message; the hive's `hive-c0re` then converges on
its own clock, out of the controller's sight. So even a fully settled
graph would not mean the agent is up, and there is nothing this CLI
could wait for that would let it claim otherwise. Printing the id is
exactly what the response says and all of what it says.
Transport is a bare hyper HTTP/1.1 client handshaked onto a tokio
`UnixStream` via `hyper_util::rt::TokioIo` — the same crate family
`hivectl/src/watch.rs` and `hive-agent/src/web_ui/proxy.rs` already use,
all of it already workspace-pinned. The request/response shapes are a
local mirror rather than a shared crate: the controller's own types are
private to its binary and this crate does not link it, the same
separation `hivectl` keeps from `hive-c0re`.
Errors are reduced to one actionable line — the controller answers
RFC 9457 problem+json, so an unknown `--hive` reaches the operator as
the roster of hives that would have worked rather than a body dump.
Response `warnings` are printed when non-empty.
The nix module wraps the binary with `SWARM_CONTROLLER_SOCKET`, read
from the same `socketPath` the daemon binds.
Refs #4399
swarm-controller's `InitAgentConfigRepo` node already covers config-repo
creation, so this deletes a duplicate rather than a capability; old
`init_config` rows are skipped by `collect_lenient` with no migration, by
operator decision.
Refs #4398
- swarm_agent_state.rs: HEADROOM in swarm_term.rs is private, so the
intra-doc link can never resolve; switch to plain backticks.
- docs/swarm/README.md: move closing period inside the quoted phrase
per Microsoft.Quotes.
Refs #3802
The swarm can already tell whether an agent is alive — the `agent-status`
KV bucket republishes once a minute — but not what it is doing right now.
A header bar wants the second thing, and a minute-old answer to "is this
agent thinking" is the wrong answer most of the time it is read.
`hive-agent` now publishes a turn-state header to
`$SWARM.agent-state.<hive>.<agent>`, a core subject beside the terminal
rows it already sends. It goes out **on transition, not on a timer**: the
publisher watches the event bus, rebuilds the header, and sends only when
the serialised result differs from the last one it sent — so a second
periodic writer, which is the problem this exists to fix, is not what
replaces the bucket.
The payload is the published contract a swarm-level renderer is written
against, so the test asserts on the serialised JSON keys rather than on
Rust field names. Two fields deliberately depart from the per-agent web
UI's `StateSnapshot`: `turn_state_since` is an ISO 8601 UTC string rather
than unix seconds, matching the sibling `$SWARM.term` subject's stamp, and
`agent_state` carries the swarm's own `AgentState` vocabulary rather than
a `paused` boolean, so a reader can compare actual against wanted without
translating. `turn_state` and `agent_state` stay two separate fields:
neither vocabulary contains the other's values.
Swarm-side, `GET /api/agents/{name}/state/stream` relays the subject as
SSE, resolving the agent's hive at request time exactly as the terminal
stream does and passing the bytes through without parsing them.
The broker grant is a second `--agent-publish-subject` rather than a
widening of the existing one, so the terminal family and the header family
stay independently revocable, and a `module-eval` arm pins the rendered
flag and its argument together — the doubled dollar included, since a
single one expands to nothing in `ExecStart` and yields a grant that
matches nothing.
Refs #3802
Refs #4374
- otelNoIdentity: name journaldUnits so the fixture trips the store-
identity path it's testing instead of swarm-otel's unrelated
journald-safety assertion (an empty list with log collection on is
refused as "collect everything", not "collect nothing").
- otelNoStores: give it a bao client identity. The secret gate moved
from deployCfg.authelia.enable to a real client cert/key pair, so a
fixture meaning "no telemetry stores" now needs its own secret
identity to keep exercising the exporter/authenticator wiring it
was written for.
- docs/swarm/secrets.md: two vale fixes — a contraction, and drop a
condescending "simply".
The swarm collector's OIDC client secret only existed where authelia
did: `swarm-otel-oidc-secret.service` copied the minted plaintext out
of authelia's container tree, reachable only because the two share a
host's network namespace. A swarm that placed authelia elsewhere
delivered nothing, and the option's own description said so —
"a deployment that places authelia elsewhere points this at a file it
delivers itself." Same gap as #3853 and #4234, and this is the
swarm-otel twin of #4234's fix for Grafana.
Mirrors PR #4361 (Grafana) almost exactly:
- `swarm-bao-otel-oidc.service` reads
`swarm/services/<client-id>/oidc/client` out of the store, in every
deployment, replacing the co-located copy unit outright — one
delivery route, not two, per the ruling that landed under #4234.
- Client registration moved out of `swarm-otel.nix`'s own `config`
block (gated on this host running the collector) into
`glue-swarm-otel-oidc-client.nix` (gated on this host running
authelia), the same split `glue-grafana-oidc-client.nix` made. It
was broken the same way: a split deployment registered the client
nowhere at all, so authelia never minted a secret for the publisher
to send on.
- The publisher's `services` prefix (write grant in `swarm-bao.nix`,
hive read grant in `policy::render`) already covers any service's
path — nothing to add there. `swarm-secret-publisher.nix` only grew
`serviceClientIds` by one entry.
One judgement call, stated rather than buried: the store-reading unit
renders only where this host holds a client identity
(`deploy.bao.clientCertFile`/`clientKeyFile`), rather than asserting
it the way `swarm-grafana.nix` does. Grafana's local login form is
disabled unconditionally, so a Grafana with no OIDC secret has no way
in at all — that earns a hard refusal. This collector without a
credential still receives every hive's telemetry; only its own pushes
to the stores go out unauthenticated and get refused there, an
already-supported degrade the module's own `haveCollectorSecret` flag
named before this change. So the reading unit follows the shape
`glue-matrix-bao-token.nix` and `glue-queue-agent-credential.nix` use
for their own optional readers: no unit when the identity is absent,
not a build refusal.
Fixtures mirror #4361's: `otelBaoWithAuthelia`/`otelBaoRemoteAuthelia`
are the positive pair (co-located and split, both reading through the
store), `otelNoIdentity` is the negative — no reading unit, no
assertion firing, `clientSecretFile` left null.
Refs #4258
An agent container cannot dial the store's loopback listener: the bridge
to-loopback DROP rule is there precisely to stop that, and the store
authenticates every reader by client certificate, so the usual answer —
a gateway vhost — is the one shape that cannot work. A terminating proxy
strips the certificate and bao sees nginx as the client for every hive.
nginx's stream module does not terminate. `ssl_preread` reads the SNI off
the ClientHello and splices the rest of the connection through byte for
byte, so bao completes the handshake itself and authenticates the client
it actually has. That is the no-vhost rule kept, not bent.
The listener binds the bridge IP rather than every address, because bao
already holds `127.0.0.1:<port>` in the same netns and a wildcard bind
there is EADDRINUSE — nginx would fail to start, taking the gateway with
it. Nothing moves as a result: the name already resolves two ways, so a
host-side reader still goes straight to loopback and an agent goes
through the passthrough, both on one `BAO_ADDR`.
Renders only inside the store's own `deploy.bao.enable` region; a host
that runs no store grows no listener and opens no port.
Per-agent certificates and per-agent policy are separate work.
Refs #4386
Moves the TermMsg rendering pipeline (Row.tsx, termMsg.ts, linkify.tsx,
markdown.ts) from @hive/agent into @hive/shared, so swarm-ui becomes a
second consumer of it instead of forking a copy -- CSS was already
shared (@hive/shared/terminal.css). marked+dompurify move from
@hive/agent's deps to @hive/shared's; swarm-ui picks them up
transitively, no new direct dep there.
New swarm-ui route /agents/:name/term (AgentTermPage), linked from
AgentsPage's detail panel via a "terminal" badge next to "link matrix
account". Consumes GET /api/agents/{name}/term/stream: unlike
@hive/agent's own useLiveStream (TermEnvelope-wrapped, history/backfill
dance), the swarm relay forwards one bare TermMsg per SSE event with no
envelope and no history endpoint -- useSwarmTermStream is a much
smaller hook for that shape (EventSource -> parse -> coalesce, nothing
to buffer/dedupe/backfill against).
Verified against a live SSE mock (screenshots in /agents/iris/state/screenshots/
3801-agents-detail-panel-terminal-badge.png and
3801-agent-term-page-live-rows.png -- real rows rendering through the
shared Row component, not just a build/typecheck pass).
The previous commit left two delivery paths and a three-way gate:
`swarm-grafana-oidc-secret.service` copied authelia's minted plaintext
out of its host tree wherever the two were co-located,
`swarm-bao-grafana-oidc.service` read the same value from the swarm
secret store wherever they were not, and `ssoConfigured && (ssoLocal ||
haveClientIdentity)` decided whether Grafana got an OIDC block at all.
Delete the co-located path. The store reader is now THE delivery unit,
in every deployment — the publisher on authelia's host writes
`swarm/services/<id>/oidc/client` whether the reader is a network away
or in the container next door. The ruling behind it: the store exists so
a host holds ONE out-of-band secret, its client certificate, and reads
everything else with it. Skipping the store when the producer happens to
be local saves a round trip and costs a second delivery unit, a second
way for the file to be wrong, and a gate to choose between them.
The gate goes too, and both of its questions become assertions, scoped
to hosts that run Grafana:
- `swarm.authelia.url` must be set. `auth.disable_login_form` is
unconditional — Grafana ships an admin/admin account on a public
vhost — so dropping the OIDC block when the swarm names no IdP
produced a container with no SSO and no password box, silently. An
eval-time refusal naming the option is the only report that reaches
anyone, the shape swarm-nats.nix already uses for the same option.
- `deploy.bao.clientCertFile` / `clientKeyFile` must be set. This
replaces a warning that nothing reads back, and its message names both
options and where the leaf comes from.
Fixtures follow. `grafanaWithAuthelia` gains the cert pair, because a
co-located host is a store reader like any other. The old
`grafanaRemoteAutheliaNoIdentity` is kept rather than deleted, renamed
`grafanaNoIdentity`: the shape is still reachable, only its deliverable
changed from silence to a refusal, and an arm now reads that refusal
back. Its mirror `grafanaNoSso` covers the other assertion, each fixture
wrong in exactly one way so an arm can name which refusal fired. Every
positive keeps an explicit negative — the one-delivery-unit arm asserts
the deleted unit is absent in both topologies rather than merely that
the store reader is present.
Refs #4234
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Grafana's OIDC client secret only existed where authelia did. One
`ssoLocal` gate — `grafana.enable && authelia.enable` — decided the
client registration, the minted secret's delivery and the whole
`auth.generic_oauth` block, so a swarm whose authelia runs on another
host got Grafana with no SSO wiring at all. The local login form is
disabled unconditionally, so that is no way in.
Split the one gate into the two questions it was conflating:
- `ssoConfigured` — does this SWARM have an identity provider
(`swarm.authelia.url`, which is swarm-wide and whose own description
makes null mean "no SSO configured"). With a delivery route present
this is what emits Grafana's OIDC block.
- `ssoLocal` — is authelia on THIS host, now spelled as the forge and
matrix modules spell it. It decides only which unit delivers the
secret.
Where authelia is elsewhere, `swarm-bao-grafana-oidc.service` reads the
secret from the swarm secret store, shaped after
glue-queue-agent-credential.nix: cert login fails loudly because a retry
fixes every state it fails on, the read degrades quietly because no
retry turns "no value there" into a value, and nothing writes a
stand-in. The producer is the publisher that already runs on authelia's
host, which gains the swarm's service clients beside the per-hive ones
at `swarm/services/<id>/oidc/client` — with the write grant in
swarm-bao.nix and the hive read grant in `policy::render` to match.
Registration moved to glue-grafana-oidc-client.nix. It has to be
declared where authelia's config is rendered, and swarm-grafana.nix's
config block hangs off this host running Grafana.
Two judgement calls stated rather than buried: a hive's read policy now
grants the whole `services` prefix, because a service's path names the
service and nothing swarm-wide records which hive runs it (cost recorded
in docs/trust-boundary/security.md); and the client is registered on any
authelia host, because no swarm-wide "this swarm has a Grafana" fact
exists to gate it on.
Refs #4234
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
argus's review on #4333 flagged one real vale error (Microsoft.Contractions):
"did not" in the new "A killed turn" section reads as "didn't" instead. The
rest of the diff's new prose (docs/tools/subagent.md and the tool
descriptions in hive-subagent-mcp/src/mcp.rs) has no other instance of the
same expansion, confirmed both by a local `vale --minAlertLevel=error` run
(clean) and by grepping the added lines. Also tidies session.rs's module doc:
the two `//!` runs split to dodge the 30-line comment-block lint had a bare
trailing `//!` right before the blank separator; dropped so the first run
ends on content.
Refs #4326
A subagent whose claude process died on a signal — the kernel's OOM
killer, a stopped unit, an `interrupt` — was indistinguishable from one
that finished its turn: its entry left the `running` map, `status` fell
through to "a session exists on disk" and answered `idle`, and the
end-of-turn todo said the subagent had "finished". The usual next move
on that reading is `continue`, which resumes work that was cut mid-turn
with nothing having recorded that it was cut.
The driver already preserves how the child ended — `RunningClaude::wait`
returns `Error::Exit` carrying the `ExitStatus`, whose `signal()` is the
whole answer — so this reads it rather than having to recover it:
`classify_end` turns the outcome into `Complete` / `Killed { signal }` /
`Failed`, and `State::finish_turn` remembers a kill against the name
(cleared by the next confirmed spawn under it).
What an agent sees as a result:
- `status` reports the session killed, naming the signal, instead of idle;
- the todo the daemon pushes without being asked says the subagent was
KILLED mid-turn rather than that it finished;
- `continue` still resumes such a session, but its reply says the
previous turn was killed, so no caller carries on from cut-off work
believing it was complete.
Refs #4326
A terminal row published on `$SWARM.term.<hive>.<agent>` goes out bare,
with no envelope around it and no server-side stamp, so a subscriber had
nothing to place the row in time with beyond its own receipt clock —
wrong by the queue's latency and meaningless for anything read later
than live.
`TermMsg` gains `ts`, ISO 8601 UTC. `classify` takes the event's own
unix-seconds stamp and applies it to every row that event expands into,
so a row replayed out of sqlite says when it happened rather than when
it was read, and a row that sat in a lagging subscriber's buffer does
not lie about its time. The oversize degrade keeps it; only the body is
ever spent.
`TermEnvelope` stops duplicating `ts` and keeps `seq`: the dedup counter
is a real transport concern, the event's time is not, now that it rides
on the row. Nothing in the frontend read `envelope.ts` — only the type
declared it.
Refs #4321
The queue's payload ceiling was justified by what the queue was about to
carry; it carries it now, so the comment says so.
The other two sites say "a hive with no queue configured". The swarm has
exactly one queue and a hive cannot lack it — only its coordinates, its
credential, or its ability to reach it. That wording is already used
everywhere else the absence is named; these two predate it.
The docs section on the agents' queue coordinates stopped at delivering
them and never said what the connection is for. It now names the subject
and the degrade rule, which is the part an operator reading an agent's
terminal at the swarm needs.
Refs #3805
The swarm always has exactly one queue; a hive can only lack its
address. Reworded every prose site this PR added that stated or
implied the opposite, to name what is actually absent (coordinates,
credential, or address) instead of the queue itself.
Refs #3805
The host end: `HIVE_C0RE_AGENT_QUEUE_CREDENTIAL_DIR` tells the daemon
where the reader unit put the files, and a new
`deploy.hive-controller.queue.agentNatsUrl` says where the queue is as an
agent *container* reaches it. That address defaults to the bridge one and
never to loopback — `statusPublish.natsUrl` beside it is loopback and
correct, because hive-c0re shares the host netns and an agent does not.
Paired with the swarm's token endpoint, gated together, and forwarded by
`hive_c0re::meta` as both an env var and an agent option: the harness
reads the variable at runtime, its unit is built from the option.
The agent end: `nix/agent-modules/queue.nix` declares that option pair
and, when set, has the harness unit inherit the two credentials by name.
Bare-id `LoadCredential=` is the terse form documented for inheriting
what the service manager received, and is non-fatal when the credential
is absent — which a hive whose publisher has not run yet needs.
No `HIVE_AGENT_OIDC_CA_FILE`: the meta flake already embeds the hive CA
and the swarm root into each container's trust store at build time, and
reqwest's rustls backend verifies against it.
Refs #3805
Ruled: swarm-bao-queue-agent.service must run before hive-c0re.service
and be wanted (not required) by it, so no agent container renders
ahead of the reader's attempt at its credential. An unreachable store
delays hive-c0re's start by the reader's own start-limit window rather
than failing it outright.
Refs #4314
The publisher on the authelia host has been writing
`secret/swarm/hives/<hive>/queue/agent` — the OIDC client secret agent
containers present to the swarm queue, plus the client id it belongs to —
and nothing read it. This is the reader: a oneshot `swarm-bao-queue-agent`
that logs in with the host's certificate and lands the two fields as two
files under `deploy.hive-controller.queue.agentCredentialDir`, the secret
`0600` and the client id `0644`.
Two files rather than one because that is the consumer's shape:
`swarm_queue_client::QueueConfig::from_env` takes the secret as a path and
the client id as a value, so the split here is what keeps the next slice
from parsing anything.
Same shape as the store's first reader, `glue-matrix-bao-token.nix` — a
cert login that fails loudly under `Restart=on-failure` because every state
it fails on is one a retry fixes, then reads that degrade quietly because no
retry turns "no value there" into a value. Unlike the matrix token there is
no local fallback and none is possible, so absent files mean this hive's
agents do not connect, which is the ordinary state of a swarm before the
publisher has run.
Nothing consumes the files yet and this unit is ordered `Before=` nothing.
The next slice bind-mounts them into agent containers through hive-c0re and
adds the ordering edge along with them.
Refs #3805
`policy::render()` became `render(hive)` when a hive gained read on its own
entry, so two places now describe a document that no longer exists: this
module's header said it "is the same for every hive and depends on nothing",
and the security doc said the grant reaches the agent-credential prefix and
nothing else.
The module header is the load-bearing one. It sits above `write_policy_for`
and says, to anyone about to touch that function, that the render is
hive-independent — which is an invitation to hoist it to a shared constant
and hand every hive the stanza naming one of them.
Agents have authelia *users*; they had no machine identity at all, so an
agent could not authenticate to the swarm queue as anything. This mints
one `agent-<hive>` OIDC client per hive beside the existing
`hive-<hive>` one, teaches the auth-callout responder an agent arm, and
opens the queue's client port on the bridge so a container can reach it.
One client per HIVE, not per agent: agents are created at runtime, and a
per-agent client would make creating one a config change plus an
authelia reload. The cost is that agents on a hive are indistinguishable
to the broker, which is deliberate and tracked separately.
The agent grant is deny-by-default twice over. An agent id matches no
hive rule, so it gets a hive's status-key grant from neither; and with
no agent subject configured the responder returns no grant at all rather
than an empty publish list, which would be a denial wearing a grant's
shape. What an agent may publish is a deployment's decision, taken
through `--agent-publish-subject` the same way `--hive-publish-subject`
already works.
`Policy::new` now refuses two prefixes where one contains the other. The
arms are tried in order, so that overlap does not error at match time -
it silently hands one principal the other's grant.
Not shipped here, and neither is reachable without it: no subject is
configured for agents anywhere in nix, and nothing yet delivers
`agent-<hive>.secret` into an agent container. Both belong to the stream
that will be the first consumer.
