Under network isolation, agents lost DNS resolution of the hive's own
zones (forge.<domain>, matrix.<domain>): the hyperhive-isolated-dns
oneshot points resolv.conf at the bridge resolver, but resolvconf
regenerated it from host-tracking right after (dhcpcd re-triggering that
when the veth comes up), reverting to the non-authoritative host
resolver. Agents then couldn't reach the in-hive forge or matrix through
the gateway, which broke an isolate-by-default rollout.
Take resolvconf and dhcpcd out of the resolv.conf loop: disable
resolvconf and tell dhcpcd not to touch resolv.conf (without disabling
dhcpcd itself, so the veth still gets its address). Then the last writer
owns it -- the nixos-container host-copy in shared netns, or the oneshot
in isolated mode. Mirrors the approach the matrix container already
takes. Eval-checked; flake check passes.
Forgejo 15 has no REST endpoint to download an Actions artifact — the
only path is the web UI download route, which is keyed by the run's
global id rather than the per-repo run number shown in run-page URLs.
The REST artifacts list route keys off the run number instead, so the
two can't be chained directly.
artifact-get takes the run number (what pr-status surfaces as a CI
context target_url), translates it to the global run id via the REST
runs list by matching each run's html_url tail, then GETs the web
download route with the agent's forge token. Saves the artifact zip to
a path (default /tmp/forge-artifact-<name>.zip) or streams to stdout
with -o -.
The artifact name is percent-encoded into the path. The encoder that
list already used for query-string filters is promoted to a shared
verbs::pct_encode helper so both call sites stay in sync.
Lets an agent pull a CI-built artifact (e.g. a paper PDF) into /shared
without host access.
Forgejo's Actions artifact API hands the runner an upload URL built from
forgejo's ROOT_URL — the public forge domain (forge.<domain>), not the
runner's 127.0.0.1:<httpPort> registration URL. The hive-ci container shares
host netns but has no resolver entry for the hive's own domains (the
gateway's localHostsEntry only touches the host), so actions/upload-artifact
fails with 'getaddrinfo ENOTFOUND forge.<domain>' while normal CI (checkout
via the localhost registration URL) is unaffected. Add a networking.extraHosts
entry mapping the forge domain to loopback so the upload reaches the local
forge via the gateway.
pr-status and pr-merge both computed 'latest non-comment review verdict per
reviewer' independently (identical oldest-first, COMMENT/PENDING-skipping,
supersede-by-later loop). Extract it to a shared verbs::latest_reviews helper
so the verdict semantics live in one place and can't drift between the
health view and the pre-merge changes-requested gate. Pure dedup, no
behaviour change.
Forgejo's local Actions-artifact storage defaults to
/var/lib/forgejo/data/actions_artifacts but Forgejo does not pre-create it.
The artifact endpoint ingests the chunked upload, then the merge-chunks step
fails with 'lstat .../actions_artifacts/tmpNNN: no such file or directory',
so every upload-artifact step dies after a successful build. Pre-create the
dir (forgejo-owned) via tmpfiles. actions.ENABLED already registers the
endpoints; this gives them somewhere to write.
Adds `hive-forge pr-merge <n> [--method merge|rebase] [--keep-branch]
[--force]` wrapping POST /repos/{owner}/{repo}/pulls/{n}/merge, so agents on
the peer-review-and-merge workflow have a CLI path instead of the raw API.
- Methods: merge (default) | rebase. Squash is intentionally not offered.
- Deletes the head branch after merge unless --keep-branch.
- Safe by default: refuses unless the PR is mergeable, CI is not red, and no
review's current verdict requests changes (latest-per-reviewer wins, so a
later approval clears an earlier request-changes). --force overrides and
also sets Forgejo's force_merge.
- New client helper post_no_content for the 200-empty-body merge response.
The C0R3 page (/core.html) pulled the four SYST3M panel styles (rebuild
queue, meta inputs, container load) by @importing the whole ~39kb
dashboard.css, dragging the entire dashboard stylesheet into the C0R3
CSS bundle.
Extract those section-internal rules into a new system-sections.css and
@import it from both core.css and dashboard.css. core.html now loads only
the small shared file (its CSS bundle drops from ~40kb to 5.3kb) while
the dashboard keeps the rules — a couple are still referenced from
dashboard JS (the schedules view uses .rqe-source*). The container-load
table's .hive-stats-table already lives in common.css, linked directly
by core.html, so it is unaffected.
Behaviour-preserving: identical rules, same custom properties; the moved
selectors are uniquely named to these panels and not redefined
elsewhere, so the cascade is unchanged. Verified the built bundles:
core.css carries the section rules and no dashboard chrome; dashboard.css
still carries them via the import.
Completes the audit log end-to-end. The backend now emits an
audit_entry_added event on /dashboard/stream per privileged action (the
new row flattened at top level). logs.js already subscribes to that stream
for the BUILD tab; add a branch that prepends the new row to the AUDIT
table's cached list (de-duped by id against the cold fetch), bumps the
'latest N of M' total, and re-renders while the AUDIT tab is in view — so
a restart shows up without a manual refresh. Docs updated to match.
argus: record() is pub and returns Option<AuditEntry>; the hive rule
requires #[must_use] on public fns returning a non-trivial value. Call
sites already consume the return (handle_restart_infra via if-let, tests
via let _).
Follow-up to the audit-log backend + surface. Emits a DashboardEvent on
each agent-initiated privileged action so the audit view live-appends off
/dashboard/stream instead of polling.
- new DashboardEvent::AuditEntryAdded { seq, <flattened AuditEntry> } —
serde tag `audit_entry_added`; the AuditEntry fields flatten to the top
level so the wire shape matches one /api/audit-log `entries` row exactly.
- Coordinator::emit_audit_entry helper (stamps seq like the others).
- audit_log::record now returns the canonical inserted AuditEntry (id + ts
assigned) so the streamed event is the same row that was stored — no
drift. Best-effort unchanged (None on a sqlite blip).
- handle_restart_infra records + emits for every attempt (ok/err/denied),
threading the coordinator through.
Tests: kind_tag round-trip now covers the new variant; added a flatten
test pinning the top-level wire shape (kind/seq/id/…/detail, no nesting).
Pairs with iris's audit view (the /dashboard/stream listener half).
The LOGS page gained a fourth sub-tab (AUDIT) surfacing the
agent-initiated privileged-action audit trail. Document the sub-tab
(lazy fetch, filterable table, outcome/denied badge colouring, the
latest-N-of-total clamp, 30s relative-time ticker) alongside BUILD /
AGENT / SYSTEM, add #audit to the hash-routing list, and add the
GET /api/audit-log endpoint to the API reference.
Adds an AUDIT sub-tab to /logs.html (alongside BUILD / AGENT / SYSTEM),
consuming GET /api/audit-log ({ entries, total }). A read-only filterable
table: when / agent / action / target / outcome / detail, newest-first,
with a 'latest 500 of N' header from total and a client-side substring
filter. Outcome badges colour ok green / err red, with an err whose detail
starts 'denied:' rendered amber + labelled 'denied' (capability refusal
reads distinct from an execution failure). Lazy-fetched on tab show (like
SYSTEM); a 30s ticker keeps the relative timestamps honest.
The audit_log store + endpoint landed in the prior audit-log backend work;
this is the operator-visible surface for it. Resolves#1647.
argus review: the four public Result-returning methods (open, list_recent,
count_total, vacuum) need # Errors sections per the hive Rust doc rules;
open had no doc comment. Added all four.
Per iris's dashboard-side ask: return { entries, total } instead of a
bare array so the UI can show 'latest 500 of N' rather than silently
capping at the clamp. Adds AuditLog::count_total().
Adds a durable, operator-visible audit trail of privileged operations
hive-c0re performs on behalf of an agent — the ones that cross the
agent/operator trust boundary. First entry: infra-container restarts via
the infra_admin-gated `restart` tool, which until now were recorded only
as a hive-priv journal trace.
Backend:
- new `audit_log` module: sqlite-backed store (audit_log.sqlite, same dir
as build_logs) with schema (ts/agent/action/target/outcome/detail),
best-effort `record`, `list_recent` (clamped 500), 90-day `vacuum`, and
a process-singleton handle mirroring build_logs.
- Coordinator opens + installs the handle; main spawns the hourly vacuum.
- agent_server::handle_restart_infra records every attempt (ok, error, and
capability-denied) via the global handle — best-effort, never fails the
underlying action.
- dashboard: `GET /api/audit-log` returns recent entries as JSON.
Scope is deliberately agent-initiated privileged actions only (not every
PrivRequest — token writes + nspawn edits are constant lifecycle noise).
Extensible: future agent-initiated priv ops record via the same handle.
Unit tests cover record/list ordering, the 500 clamp, and retention vacuum.
The dashboard *surface* (an AUDIT view consuming /api/audit-log) is a
frontend follow-up coordinated with iris.
The dashboard LOAD tab always showed "no running agent containers":
container_stats::gather() looked for each agent's cgroup at
machine.slice/machine-h\x2d<name>.scope, but that path never exists.
nixos-container runs `systemd-nspawn --keep-unit` with
`Slice = "machine.slice"` (nixpkgs virtualisation/nixos-containers.nix),
so --keep-unit means no separate machined `machine-<name>.scope` is
created — the container's cgroup IS the launching service unit,
`container@<machine>.service`, under machine.slice. systemd-machined
still logs "New machine <name>" (registration), which is what made the
scope-path assumption look plausible, but the cgroup stays on the unit.
Fix scope_dir to machine.slice/container@<machine>.service and drop the
\x2d escaping (the service unit name is used verbatim in the cgroup dir;
the journal shows the literal `container@h-<agent>.service`). Adds a unit
test pinning the path.
Clicking "update & rebuild" (and cancel / respawn / purge) on /core.html
navigated to the bare `ok` response page instead of submitting async. The
C0R3 page (split out of the dashboard) carries `data-async` forms but its
bundle never had the global submit interceptor — that handler lived inline
in tabs.js, so only the dashboard bundle had it. The forms POSTed natively
and the browser followed the response.
Fix: lift the `data-async` submit handler out of tabs.js into a shared
`bindAsyncForms(onSuccess)` in common.js (which already owns the `form`
helper that builds these forms), and call it from both pages:
- tabs.js: `bindAsyncForms(() => refreshState())` — behaviour-preserving
(same handler, now imported).
- core.js: add a `refreshState()` (re-fetch /api/state + re-render) used
for the cold load and as the post-submit refresh, and call
`bindAsyncForms(() => refreshState())` at boot.
Forms with `data-no-refresh` (e.g. meta-update, which gets its update via
the meta SSE events) skip the refresh, same as before.
The per-agent stats page gained a "first-turn ctx" summary chip (the
input tokens of the most recent fresh session's first turn — a
system-prompt/CLAUDE.md sprawl proxy), but agent.md's chip-row
description didn't mention it. Add it next to the reminder chips, noting
the stats.rs derive and the inert-until-capture behaviour (omitted from
JSON until the sessions / turn_stats.session_id capture has rows).
The C0R3 page PR moved the SYST3M tab off the dashboard onto the
standalone /core.html page, leaving the web-ui docs describing a tab that
no longer exists.
- dashboard.md: retitle the "SYST3M tab" section to "C0R3 page
(/core.html)" and reframe its intro — it's now a standalone page off the
H0M3 hub with a createTabStrip sub-tab nav (default R3BU1LD QU3U3), its
own core.js bundle on /dashboard/stream; the dashboard keeps only the
rebuild-queue state for the SW4RM "building…" badges.
- correct the chrome tab-strip list to the actual in-page tabs
(SW4RM / Y3R C4LL / P3RM1SS10NS / SCH3DUL3S) and note SYST3M / ST4TS /
S3TT1NGS / FL0W / L0GS / M4TR1X are all standalone pages off H0M3.
- fix the container-load poll reference (core.js, while the C0NT41N3R
L04D sub-tab is active) and the stray "SYST3M ›" panel references.
- drive-by: the schedules endpoint doc said "SYST3M scheduled-prompts
panel" — it's the SCH3DUL3S panel; corrected.
- web-ui.md index: same tab-list correction + C0R3 page pointer.
The "What the Rust side does" section predated the isolated-container
egress fix and omitted two now-implemented mechanisms:
- HOST_ADDRESS=<bridge-ip> in the nspawn conf — load-bearing for the
container's default route (nixos-container only installs `default via`
when HOST_ADDRESS is non-empty; bridge mode skips host-side route setup).
- the resolver wiring: hive-priv's bridge-DNS marker + the
hyperhive-isolated-dns oneshot rewriting resolv.conf to the bridge
dnsmasq (nixos-container re-copies the host's resolv.conf each start, so
the fix is runtime, ordered before the harness).
Doc-only; brings network.md in line with the merged behavior.
Follow-up to the isolated-container egress fix. The hyperhive-isolated-dns
oneshot (which rewrites resolv.conf to the bridge resolver) was ordered
before network-online.target + tea-login, but NOT relative to the harness
(hive-ag3nt) or the matrix daemon. hive-ag3nt only declares
`after network.target`, so on first boot its first-turn api.anthropic.com
lookup could race ahead of the resolv.conf rewrite and error (self-heals
next turn, but flaps the first one).
Add hive-ag3nt.service + hive-matrix-daemon.service to the oneshot's
`before` so DNS is fixed before any network consumer starts. The matrix
entry is a harmless no-op when matrix is disabled (unit absent).
Caught by damocles in review of the parent PR.
When isolateContainers=true, claude (and all egress) broke in every
container: agents came up with an IP but no way off the bridge subnet.
Two container-side gaps, both confirmed against nixpkgs
nixos-containers.nix:
1. No default route. hive-priv wrote HOST_ADDRESS= empty in the nspawn
conf. nixos-container's container-side setup only installs
`ip route add default via $HOST_ADDRESS` when HOST_ADDRESS is
non-empty, so the container had an address but no gateway -> nothing
off-subnet (incl. api.anthropic.com) was reachable. Fix: write
HOST_ADDRESS=<bridge-ip>. In bridge mode the host-side address/route
setup is skipped, so this only affects the container's default route.
2. No usable resolver. nixos-container copies the host's /etc/resolv.conf
into the container at every start; the host resolver (e.g. 127.0.0.53)
is unreachable from a private netns and isn't authoritative for the
hive's own zones. Fix: hive-priv drops a marker carrying the gateway
IP only when isolated, and a new harness-base oneshot
(hyperhive-isolated-dns) rewrites resolv.conf to point at the bridge
dnsmasq. Inert in shared-netns mode (no marker), so the shared
container toplevel does the right thing in both modes.
The gateway IP is the address part of HIVE_NETWORK_SUBNET (the bridge IP
verbatim, honouring a non-.1 operator override), via a new validated
bridge_gateway_ip() helper with unit tests.
Unblocks defaulting isolation on.
Builds the read/surface half of the per-session first-turn-tokens metric
(the capture — sessions table + turn_stats.session_id — landed separately).
A fresh claude session's first turn pays the full static prefix (system
prompt + tools + CLAUDE.md + first wake) as uncached input, so its
input_tokens is a clean proxy for prompt / CLAUDE.md sprawl — watching it
over time surfaces creep.
- stats.rs: add `Snapshot.first_turn_ctx: Option<u64>` populated by
`read_first_turn_ctx` — the agreed per-session derive (first turn,
`ORDER BY started_at LIMIT 1`, of the most recent session that started
in the window). Inert-until-capture: `.ok()` maps both "no fresh
session yet" and "older db without the sessions table" to None, the
same decoupling as read_bash_breakdown; the field is skipped from the
JSON when None. Pre-capture rows have a NULL session_id and are excluded.
- agent stats.js: add a "first-turn ctx" summary chip, guarded on a
numeric value so it stays hidden until capture has data.
clippy + cargo fmt clean; agent bundle builds.
