Skip to content

🌱 ci: size overlayfs-exec-guard timeout for the cold whisper-cache path - #18604

Merged
hivecommons-hive[bot] merged 8 commits into
v5from
ci-overlayfs-guard-speed
Nov 17, 2026
Merged

hivecommons-hive[bot] merged 8 commits into
v5from
ci-overlayfs-guard-speed

Conversation

@clubanderson

Copy link
Copy Markdown
Member

Why

overlayfs-exec-guard keeps getting cancelled at its 10-minute timeout on PRs that have nothing to do with the image (#18432 three times, #18446, #18500, #18527). Every cancellation I sampled had the same cause: the build ran with a cold whisper-builder cache.

  • The repo's Actions cache holds 12-23 GB against a 10 GB cap (max_cache_size_gb: 10), so GitHub keeps evicting the least recently used entries. The oldest surviving entry is about 2 days old. The whisper-linux-amd64 blobs get dropped often enough that docker.yml's own v5 warm step recompiles whisper several times a day (2.5 min on the self-hosted pool, versus about 3 s on a cache hit).
  • On a miss, the hosted 4-vCPU runner compiles FFmpeg and whisper.cpp from source. That is a serial chain of about 8.5 min (apt about 35 s, FFmpeg 95-150 s, whisper.cpp 210-335 s), which puts the build step at about 10.5 min and the job at about 13 min.
  • Warm runs (12 sampled on 2026-11-17) take 6.5-8.1 min per job, with the buildx step at 4.3-5.8 min. A cold run cannot fit in 10 min.

What

  • timeout-minutes: 10 becomes 20: about 13 min for the cold path plus roughly 50% headroom. The warm path does not change.
  • setup-buildx-action now uses cleanup: false. The runner is ephemeral and hosted, and the post-step builder teardown cost 16-36 s on every run.
  • Regenerated integration-candidate.yml and integration-technical-v2-ci.yml with generate-integration-ci.py (--check and its tests pass).

Not done, and why

  • Self-hosted runner (HIVE_RUNNER_LABELS=["hive-runners-lke"]): the guard is hosted on purpose, because it asserts the systemd containerd plus overlayfs semantics of GitHub runners.
  • Building less than the full image: the assertion targets the final runtime stage, including the in-stage binary rewrite, so a narrower target would need a Dockerfile change and would weaken the guard.
  • Durable whisper cache: a registry cache in GHCR would not be evicted, but it needs a package and visibility decision. Shrinking the repo's gha cache usage under 10 GB would also help docker.yml. Both are follow-ups, not CI-only changes.

CI-only change under .github/, so no changelog fragment is needed.

πŸ€– Generated with Claude Code

The guard was cancelled at its 10-minute timeout whenever the
whisper-linux-amd64 gha cache missed. The repo's Actions cache runs at
12-23 GB against a 10 GB cap, so LRU eviction regularly drops the
whisper-builder blobs; the hosted 4-vCPU runner then compiles FFmpeg and
whisper.cpp from source (~8.5 min serial chain) and the job needs ~13 min.
Warm runs measure 6.5-8 min and are unaffected.

Raise the timeout to 20 min (cold path plus ~50% headroom) and skip the
buildx post-step teardown on the ephemeral runner (16-36s per run).
Regenerated the integration workflows from v2-ci.yml.

Signed-off-by: Andy Anderson <andy@clubanderson.com>
@kubestellar-prow kubestellar-prow Bot added dco-signoff: yes Indicates the PR's author has signed the DCO. size/S Denotes a PR that changes 10-29 lines, ignoring generated files. labels Nov 17, 2026
@github-actions

github-actions Bot commented Nov 17, 2026 •

Copy link
Copy Markdown
Contributor

Open PR overlap check

βœ… Resolved: no conflicting open PRs or same-file overlaps remain against v5.

@hivecommons-hive

hivecommons-hive Bot commented Nov 17, 2026 •

Copy link
Copy Markdown
Contributor

Hive status β€” updated in place; edits do not notify. Full history is in the Activity Monitor.

ℹ️ Sentinel: informational, author is trusted (owner), merge is not blocked. Head c5d2995617bd matched:

  • sensitive_path: changes 3 sensitive path(s) (.github/workflows/integration-candidate.yml, .github/workflows/integration-technical-v2-ci.yml, .github/workflows/v2-ci.yml)

No sentinel-alert label was added. The finding is recorded in the dashboard Security tab and /api/sentinel; set sentinel.trusted_authors_block: true to block trusted authors too.

@clubanderson this PR is held at stage/071-needs-author and the hive will not merge it yet.

Why:

  • the review swarm returned requires_human: it has a question for you, not a fix for the hive
  • the reviewer asked you: hanthor should confirm that a 20-minute timeout and cleanup: false on the hosted overlayfs-exec-guard job in v2-ci.yml are acceptable, since .github/workflows/** is CODEOWNERS-gated.

Next: the hive does not push to your branch. Push a new head (or reply here); CI and the review swarm re-run automatically and the PR returns to the normal path.

hive/awaiting-ci: CI pending β€” sweep will merge when green. (Merged v5 in; the red build-contributor was a cache-blob fetch failure, not the diff.)


🐝 Hive Agent: scanner | Instance: hosted-available-oke-11-placeholder-r05x | SHA: 75901f01

β€” hive: agent=scanner backend=claude model=claude-sonnet-5-5 claude=2.1.325

Evidence bundle `#18604f69cca9` sha256:eefb4acc047115c031121445f7b339e319a0f865878fa7becebfdfafc2ee2198 β€” download from the dashboard or `hivectl review evidence #18604`.

merged by Hive trusted-author auto-merge: author clubanderson holds role owner and GitHub permission admin on repo; CI green at f69cca9.

@hivecommons-hive hivecommons-hive Bot left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Decision needed: @clubanderson should confirm the CI workflow timeout and buildx-cleanup changes in .github/workflows may land.

Routing only because this edits .github/workflows/{v2-ci,integration-candidate,integration-technical-v2-ci}.yml; I found no defect.

What I read (head c5d2995): the diff only raises timeout-minutes 10 -> 20 on the overlayfs-exec-guard job and adds cleanup: false to docker/setup-buildx-action. No permissions, trigger, if: or action pin changes, so the privilege surface is unchanged. The v2-ci.yml comment gives measured cold-cache timings that match the stated cancellations.

Cost to note: a hung job now holds a runner for 20 min instead of 10. cleanup: false is safe on ephemeral hosted runners (runs-on: ubuntu-latest in all three).

β€” hive: agent=reviewer backend=claude model=claude-sonnet-5-5 claude=2.1.325

@hivecommons-hive

hivecommons-hive Bot commented Nov 17, 2026 •

Copy link
Copy Markdown
Contributor

Needs a decision, not a merge:

  • the reviewer swarm returned requires_human @7cdf4b5
  • the review found: the reviewer marked this change high-risk
  • reviewer question: @hanthor should confirm that a 20-minute timeout and cleanup: false on the hosted overlayfs-exec-guard job in v2-ci.yml are acceptable, since .github/workflows/** is CODEOWNERS-gated.
  • the PR edits CI workflow / action files: awaiting owner approval β€” a hive owner / trusted merger's GitHub Approve review on the current head (or a dashboard /hive decision approve) releases the fence and the hive merges it; approval on an earlier head does not count unless the hive's own merge of the base branch is all that moved the head, so re-approve after a push

Head: verdict requires_human @7cdf4b5, 24/51 checks green, 27 pending, DCO yes, soak 0s of 10m0s.

How to release it:

  • A hive owner / trusted merger submits a GitHub Approve review on the current head (or approves from the Hive dashboard, which posts /hive decision approve on their behalf); that explicit approval releases the workflow-path fence and the lane merges on its next pass once CI, soak and the other gates are clear. Approval on an earlier head carries over only the hive's own merge of the base branch (which leaves the diff unchanged): re-approve after any other push.
  • After reading the verdict, a trusted merger can submit a GitHub Approve review for the current head and the lane clears the hold on its next pass; /hive approve replies on a PR are acknowledged but not consumed. An ordinary contributor cannot release their own PR.

Clear the cause and the hive re-evaluates on its next pass; the stage moves to stage/090-ready-for-human-merge, stage/095-pre-merge-soak or the merge queue on its own. Approval does not bypass CI, hold labels or other merge gates.

Updated 2026-11-17 05:39 UTC.

@hivecommons-hive hivecommons-hive Bot left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Decision needed: @clubanderson should confirm that raising the overlayfs-exec-guard timeout to 20 min and disabling buildx builder cleanup in CI workflows is acceptable to merge.

Read the three changed workflow files. The diff is what the body describes: timeout-minutes: 10 to 20 and cleanup: false on setup-buildx-action, hand-edited in v2-ci.yml and mirrored into the two generated integration workflows. No permissions, triggers, secrets, if: conditions or action pins change, so I found no defect.

Routing only because it edits .github/workflows, which I do not approve on my own. Residual note: the 20 min ceiling treats a cold whisper cache as normal; the body already lists the durable-cache follow-up.

β€” hive: agent=reviewer backend=claude model=claude-sonnet-5-5 claude=2.1.325

@hivecommons-hive hivecommons-hive Bot left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Decision needed: @clubanderson should decide whether doubling the overlayfs-exec-guard timeout to 20 min (rather than fixing the cache eviction behind the cold builds) is acceptable, since it edits three .github/workflows files.

Read the three workflow diffs: v2-ci.yml is the source edit, and integration-candidate.yml and integration-technical-v2-ci.yml carry the same two changes, which matches a regenerated copy. I found no defect in the change itself.

  • timeout-minutes: 10 β†’ 20 and setup-buildx-action cleanup: false are applied identically in all three files. The action SHA is unchanged, so no new third-party code is pulled in.
  • cleanup: false is only safe because the job is on ubuntu-latest (hosted, ephemeral), as the PR body says. If this job ever moves to a self-hosted runner, builders would leak. The new comment at v2-ci.yml already notes the ephemerality, which covers it.
  • Routing: this touches workflow files, which I always hand to a person. The trade-off is that a cold-cache run can now hold a hosted runner for up to 20 min before cancelling, versus the underlying cause (Actions cache over its 10 GB cap).

Looks correct to me otherwise. I have not verified the timing measurements in the body.

β€” hive: agent=reviewer backend=claude model=claude-sonnet-5-5 claude=2.1.325

@clubanderson

Copy link
Copy Markdown
Member Author

Answering the reviewer decision question (no changes requested, so nothing was pushed):

  • Why a timeout change and not a cache fix here: the cold builds come from repo-wide Actions cache eviction (12-23 GB stored against a 10 GB cap). This job cannot fix that. The durable fixes are a GHCR registry cache for whisper-builder or cutting overall gha cache usage, and both need an owner decision on packages or docker.yml, which are out of scope for this PR. Until one lands, the cold path happens several times a day, and a 10 min ceiling turns each occurrence into a spurious red.
  • Cost: only a genuinely hung run now holds a hosted runner longer (20 min instead of 10). Normal runs do not. On this head the guard finished in 5m57s (run 44013789071), against 6.5-8.1 min before, because cleanup: false removes the 16-36 s builder teardown.
  • cleanup: false applies only to the ubuntu-latest hosted, ephemeral job, as the reviewer noted.
  • The build-contributor failure on this head is a docker.yml gha cache BlobNotFound, which is the same eviction problem and is unrelated to this diff. I re-ran the failed jobs.

@clubanderson clubanderson added lgtm Indicates that a PR is ready to be merged. approved Indicates a PR has been approved by an approver from all required OWNERS files. labels Nov 17, 2026
@kubestellar-prow

Copy link
Copy Markdown
Contributor

[APPROVALNOTIFIER] This PR is APPROVED

Approval requirements bypassed by manually added approval.

This pull-request has been approved by:

The full list of commands accepted by this bot can be found here.

The pull request process is described here

Details Needs approval from an approver in each of these files:

Approvers can indicate their approval by writing /approve in a comment
Approvers can cancel approval by writing /approve cancel in a comment

Signed-off-by: scanner <280983584+hivecommons-hive[bot]@users.noreply.github.com>
@kubestellar-prow kubestellar-prow Bot removed the lgtm Indicates that a PR is ready to be merged. label Nov 17, 2026
@kubestellar-prow

Copy link
Copy Markdown
Contributor

New changes are detected. LGTM label has been removed.

@hivecommons-hive hivecommons-hive Bot left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Decision needed: @hanthor should a code owner confirm that raising overlayfs-exec-guard's timeout to 20 minutes and setting buildx cleanup: false is acceptable in the workflows?

Routing only because this edits .github/workflows/**, which CODEOWNERS lists as a trust surface. Facts I read in the diff: in v2-ci.yml, integration-candidate.yml and integration-technical-v2-ci.yml, the overlayfs-exec-guard job changes timeout-minutes: 10 to 20 and setup-buildx-action gains cleanup: false. No permissions, secrets, triggers, action pins or if: conditions change, and the action SHAs are unchanged. The body's rationale (cold whisper cache on an ephemeral hosted runner) matches the comment added in v2-ci.yml; I did not verify the quoted timings. The generated files appear to match the source edit.

β€” hive: agent=reviewer-2 backend=claude model=claude-sonnet-5-5 claude=2.1.325

@hivecommons-hive hivecommons-hive Bot left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Decision needed: @hanthor should confirm that a 20-minute timeout and cleanup: false on the hosted overlayfs-exec-guard job in v2-ci.yml are acceptable, since .github/workflows/** is CODEOWNERS-gated.

Routing only: I found no defect, but this edits .github/workflows/** (CODEOWNERS: @clubanderson @hanthor @Danathar), which I never approve on my own.

What I read: the diff touches the overlayfs-exec-guard job in v2-ci.yml and the two generated files integration-candidate.yml and integration-technical-v2-ci.yml. The three edits are identical (timeout-minutes: 10 β†’ 20, plus with: cleanup: false on docker/setup-buildx-action), so the generated copies are consistent with the source. Permissions, triggers, action pins and the guard's assertion are unchanged. cleanup is an existing setup-buildx-action input and the job runs on the ephemeral ubuntu-latest runner, so skipping builder teardown there looks harmless to me.

The one real trade-off is a cold run can now hold a hosted runner for up to 20 minutes before a hang is cancelled, in exchange for fewer spurious cancellations. The PR body documents the timing evidence and the alternatives it rejected. The 6.5–8.1 min warm and ~13 min cold figures are the author's measurements and I did not re-verify them. Several checks were still in progress when I looked, so please wait for the generator --check result.

β€” hive: agent=reviewer-2 backend=claude model=claude-sonnet-5-5 claude=2.1.325

@hivecommons-hive
hivecommons-hive Bot merged commit dfbc34d into v5 Nov 17, 2026
84 of 85 checks passed
@kubestellar-prow
kubestellar-prow Bot deleted the ci-overlayfs-guard-speed branch November 17, 2026 06:07

@hivecommons-hive hivecommons-hive Bot left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Decision needed: @clubanderson should confirm this CI-only .github/workflows timeout/cleanup change may merge, since workflow files are a trust surface the reviewer does not approve.

Read all three workflow hunks. Looks correct to me: the same two edits (timeout-minutes: 10 β†’ 20, docker/setup-buildx-action with cleanup: false) are applied identically in v2-ci.yml and the two generated files, so the generator output and source agree. No permissions, triggers, if: conditions, secrets or action pins change (the buildx SHA is unchanged), and persist-credentials: false is untouched. The body's claims match the diff.

Not verified by me: the cold-cache timing figures (measurement claims from the body) and that cleanup is accepted by the pinned buildx action revision; CI on this head will show the latter.

No code defects found; routing only because the files are under .github/workflows. I have not approved or merged anything.

β€” hive: agent=reviewer-2 backend=claude model=claude-sonnet-5-5 claude=2.1.325

kubestellar-prow Bot pushed a commit that referenced this pull request Nov 17, 2026
* πŸ› meetings: log every failed meetings call and keep a failed approval's reason (#18600)

* πŸ› meetings: log every failed meetings call and keep a failed approval's reason

A Quorum refusal of a Discord join (blocked: no permission to connect)
was answered as 502 with no log line, and the hosted ingress replaced the
body with its generic page, so the owner saw 'hive unavailable' and the
spoke log said nothing. Every meetings call failure is now logged with
method, path, status and error (Warn for 5xx, Info for caller mistakes),
and ApproveRequest records the failure reason on the still-open request.

Refs #18598

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ› meetings: a blocked Discord join names the bot's permissions, and a failed approval shows on the request

The panel read every blocked meeting as 'The meeting refused a guest bot',
which is a Meet/Teams lobby refusal; on Discord blocked means the bot's
role lacks View Channel, Connect or Speak on the voice channel. The join
request row now also renders the reason a failed approval saved.

Refs #18598

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ“– api-reference: re-point meetings route citations after the import shift

Refs #18598

Signed-off-by: Andy Anderson <andy@clubanderson.com>

---------

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* ✨ Remote items: follow up on issues and PRs the hive opened as its App bot (#18610)

* ✨ remote items: follow up on items the hive opened as its App bot

With project.remote_write_identity a write-tier reference repo can write as
the App bot, but the own-item poller listed creator=<owner> under the
owner token only, so feedback on bot-opened issues and PRs never reached
the agent.

- remoteitems: RunTargets takes (repo, login) pairs; each repo is listed and
  its feedback filtered under its own login. Run keeps its signature as a
  one-login wrapper.
- github.RemoteItemsFetcher: optional TokenFor picks the read token per
  repo and is authoritative when set (no fallback to the owner token).
- cmd/hive: every write-tier repo is watched under its CredentialLogin and
  read with the token the proxy injects on its writes (scoped bot token,
  owner token or write_token_file).

Refs #18516

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ› remote items: fix the per-repo login test's use of Store.Items

Store.Items returns a slice; build the tracked set from Item.Key.

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ“– two-lane doc: the own-item follow-up covers bot-opened items too

Signed-off-by: Andy Anderson <andy@clubanderson.com>

---------

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ› fix(ioscan): do not fleet-report an injection-withheld kick as a classifier outage (#18618)

A fail-closed kick block was always logged as "ioscan fail-closed blocked
kick", which the fleet report classes as ioscan_classifier_unavailable. But
the kick is also withheld when untrusted issue/PR text trips the injection
policy (a Critical finding or a classifier block verdict) while the
classifier is answering: the defense working as designed, not an outage.
The #18459 report carried no classifier_error, i.e. exactly that case.

warnBlockedKick now keeps the fleet-reported wording (with classifier_error
and fix) only when the classifier is failing, and otherwise logs "ioscan
fail-closed withheld kick on an injection finding", listed in
fleetreport.IsBenignMessage. The held-PR coordination path goes through the
same helper, so it now also names the classifier error on an outage.

Closes #18459

Hive-Run: #18459
Hive-Plan: selfreport-ioscan_classifier_unavailable
Hive-Spec: fleet-report#ioscan_classifier_unavailable
Signed-off-by: Karibhou <157208370+Karibhou@users.noreply.github.com>

* 🌱 ci: fetch only the histories the squash-attribution check reads (#18609)

`fetch-depth: 0` makes actions/checkout fetch every branch and tag. That
took 17-35s normally and on #18482 (run 43960382195) hung for the whole
5-minute job budget, failing the required "Sign-off survives the squash"
context before the check ran.

Check out the head shallowly, then fetch the head, base and default-branch
histories in full (unshallow), with a 60s per-attempt stall timeout and
three attempts. Picks up the fix a contributor reverted out of #18482 as
unrelated to that PR, plus the stall retry and post-fetch assertions.

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* ✨ Hand workflow-model states to their agent profiles (claim first, dispatch once) (#18571)

* ✨ Hand workflow-model states to their agent profiles

The eval cycle now reads the bound model: an item whose state the model
hands to a profile (TechDocs Accepted -> techdocs-drafter, Drafting on
/verify -> techdocs-verifier) is withheld from every generic lane, claimed
for the agent that fills the profile on the issue-claims ledger, and kicked
once. The hand-off key (item, profile, state, triggering comment) is
written to workflow-dispatch.json on delivery, so neither a later cycle
nor a restart repeats it; a failed or deferred kick is retried.

Part of the drafter hand-off for cncf/techdocs#395.

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ› List a bound repository's drafts at most every five minutes

Signed-off-by: Andy Anderson <andy@clubanderson.com>

---------

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ› Betty closes the loop on deep dives; research agents survive config reloads (#18592)

* πŸ› Betty closes the loop on deep dives; research agents survive reconciles

research-1..N are added to the runtime only (not hive.yaml), and
ReconcileAgents (config reload, reviewer autoscale) removes every agent
not in the enabled config set, so the pool was emptied shortly after boot
and each meeting research request failed with "no research agent
available" while nothing was said in the meeting.

- cmd/hive: the reconcile set carries the research pool's agents.
- research.Pool: a dispatch that finds no agent re-registers them once;
  ErrNoAgents carries a sayable Reason; Status() for the settings page.
- live: a persona deep dive that cannot start, errors, or runs past
  5 minutes is answered with "I couldn't look that up: <reason>." in
  the channel it was asked in; one progress note after 30 s; a spoken
  question's follow-up is spoken.
- Settings, Meeting attendee: Research section with agents, availability
  and a "No research agent available" warning.

Part of #18590

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ› tests: set agents through AgentRuntimeConfig

Part of #18590

Signed-off-by: Andy Anderson <andy@clubanderson.com>

---------

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* ✨ Review follow-ups, trust-only 090 and issue pr-link notes become status-card sections (quiet mode) (#18616)

* ✨ Review follow-ups, trust-only 090 and issue pr-link notes become status-card sections (quiet mode)

In github.notifications.mode: quiet the review Follow-ups checklist is the
collapsible review-followups section of the PR's hive:status card (ticks
kept via the new prstatus.WriteFunc), a 090 ready-for-human-merge entry
whose only objection is the author's trust tier is the stage-ready line,
and the issue-side PR-link stage note is the issue card's issue-pr-link
section. Every 085 and a held or manual-merge 090 still post a comment.

Fixes #18606

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ› status-card follow-ups use their own marker; tests expect neutralised mentions

The card section carried the legacy Follow-ups comment marker, so the next
pass read the card as the legacy comment and EditIssueMarkerComment
rewrote the whole card. Stamp the section with
<!-- hive:review-followups-card --> and skip status comments in the legacy
check. The 090 card line goes through prstatus.Write, which neutralises
'@sha' to a code span; the test now expects that.

Signed-off-by: Andy Anderson <andy@clubanderson.com>

---------

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* 🌱 ci: no mutation comment on error verdicts; no reopen hint for bot or maintainer reporters (#18617)

* 🌱 ci: no mutation comment for an error verdict; no reopen hint for bot or maintainer reporters

mutation-report.sh post no longer creates a <!-- hive-mutation --> comment
for an 'error' verdict (no report, nothing to act on): the report stays in
the job summary and the failing verdict step, and an earlier finding comment
is edited so it is not stale. issue-closed-reopen-hint.yml skips issues
whose reporter is a bot or OWNER/MEMBER/COLLABORATOR. pr-overlap.sh already
posts only when a sibling PR overlaps (#17870); unchanged.

Fixes #18607

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* 🌱 changelog fragment for #18607

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* 🌱 re-pin TRUSTED_SCRIPTS_TREE for the mutation-report.sh change

Signed-off-by: Andy Anderson <andy@clubanderson.com>

---------

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ› Meetings: keep nothing from before the consent notice is delivered (#18621)

Until the consent notice is delivered (consent_posted_at), every
transcript segment, draft and chat line is discarded in the manager:
not stored and not emitted to listeners (Betty, Spektacular, research,
facts, Inbox hand-off). Only a pre_consent_dropped count is kept. A
leave phrase is still honoured and the line itself is not kept.

Lines that arrive after the notice but belong before it are dropped
too: seq at or below Quorum's transcript cursor when the notice landed
(consent_seq), chat stamped earlier, and for two minutes segments that
began before the last speech heard while pending (consent_speech_end).
A reconnect or restart while pending asks Quorum for no replay, and
dropped lines advance the replay cursor.

Refs #18590

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ› fix: stop open-PR guard label retry storms and classify the failure (#18624)

The has-open-pr label write shared the caller's 15s deadline with the
open-PR listing ahead of it, so a slow listing of a large repo left it a
spent context. The write failed with context deadline exceeded, the issue
never got the label, and every later claim pass retried it and warned
again, which filed an unclassified_error_repeated self-report.

- give the label write its own 10s budget, detached from the caller
- remember a failed write per issue: 15m after a transient failure, 6h
  after a refusal (403/404/410/422), and log refusals at INFO
- add the open_pr_guard_label_failed fleet-report class for the
  remaining repeated WARNs

Fixes #18589

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* 🌱 ci: blobless history fetch for the squash-attribution check (#18634)

#18609's unshallow step fetched full blob history and measured 38-42s on
the self-hosted pool, slower than the all-refs checkout it replaced. The
checker reads only commits and trees, so fetch with --filter=blob:none
(measured 84s -> 14s and 277 MB -> 24 MB locally for the same refs).

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ› Deflake TestSessionCreateWithSeedWaitsForRunning (#18632)

The test returned once the instruction was typed, while the queue worker
and seed delivery were still writing (touch, audit, inbox routing). TempDir
cleanup then raced those writes (directory not empty). Wait for the audit
and for the queue worker to exit before the test ends.

Fixes #18545

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ”– release: v5.293.0

Automated release commit. Compiles changelog.d/ fragments and moves
the CHANGELOG.md Unreleased section into a dated v5.293.0 entry.
See src/docs/releases.md.

Signed-off-by: hive-release-bot <actions@github.com>

* ✨ Owner-only Send to <profile> on model-owned stage-tracker rows (#18576)

* ✨ Owner-only Send to <profile> on model-owned stage-tracker rows

A row of a model-bound repository whose state the model hands to an agent
profile offers the owner Send to <profile> in its menu, backed by the
owner-only POST /api/workflows/dispatch: the item is read fresh, the
agent resolved (a profile disabled for automatic dispatch still
resolves), the issue claimed first and the agent kicked; the hand-off is
recorded once the kick lands so the automatic pass does not repeat it.
Audited workflow_dispatch; openapi and api-reference updated.

Part of the drafter hand-off for cncf/techdocs#395.

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ› Register the kicked agent in the dispatch endpoint test runtime

Signed-off-by: Andy Anderson <andy@clubanderson.com>

---------

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ“– docs: Betty on Zoom without a Zoom app, options and recommendation (#18602)

Record the #18596 comparison as a design page: phone dial-in, a desktop
companion that streams the user's own call audio, and Zoom's RTMP live
stream, each against the free-account, bot-rule, consent and cost
constraints. Recommends the desktop companion (macOS first) and records
its proposed streaming contract, per-user companion token, distribution
plan and prototype PR set. Indexed in docs/design/README.md and linked
from meetings.md, which now notes that Zoom refuses web-client bots.

Hive-Run: #18596
Hive-Plan: 18596 recommendation and prototype PR set
Hive-Spec: #18596#recommendation

Signed-off-by: Karibhou <157208370+Karibhou@users.noreply.github.com>
Co-authored-by: clubanderson <407614+clubanderson@users.noreply.github.com>

* πŸ› hub: converge cluster host_aliases on every hosted spoke, push-reported clusters included (#18639)

The host_aliases reconcile sweep (#17805) patches Deployments over kubectl,
but vllm-d and a-ks-wec2 are push-reported (pull_only) clusters the hub
cannot reach, so the sweep skipped every hive there silently. Only spokes
rendered from the template after #17805 carried the aliases.

The hub now sends the SaaS record cluster's host_aliases on every heartbeat
response and the spoke converges its own Deployment with a JSON patch
(exact for the managed hostnames, unrelated aliases kept, throttled to one
check per 15 minutes). The hub sweep reports push-reported clusters and
WARNs on unreadable or unreachable Deployments instead of logging at Debug.

Fixes #18637

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* ✨ meetings: desktop companion meetings (/api/companion/*) (#18611)

* ✨ meetings: desktop companion meetings (/api/companion/*)

A companion meeting is fed by the desktop companion: a person's own meeting
client audio (call) and microphone (mic), captured on their computer after a
versioned consent confirmation, relayed as 16 kHz PCM to Quorum's companion
platform (quorum 0.3.0). Betty's replies go to the companion window through
Quorum's outbox. Owner or read-write; a companion meeting is usable only by
its starter (or an owner).

Refs #18596

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ“– docs: rewrite api-reference citations shifted by the companion routes

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ› meetings: companion handlers use the logging meetingsError (#18600)

Signed-off-by: Andy Anderson <andy@clubanderson.com>

---------

Signed-off-by: Andy Anderson <andy@clubanderson.com>
Co-authored-by: scanner <280983584+hivecommons-hive[bot]@users.noreply.github.com>

* 🌱 deps: quorum v0.3.0 (desktop companion platform) (#18622)

quorum.mjs sha256 from the v0.3.0 release's quorum.mjs.sha256; davey stays
0.1.12 (quorum's package-lock.json).

Refs #18596

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* ✨ Betty: parallel answers with attribution, sources and timings, voice queue, delete finished meetings, quiet Spektacular restarts (#18635)

* πŸ› Meetings: don't restart-loop the Spektacular stream child when live mode isn't available

The pinned spektacular (0.31.0) has no `meeting` command
(hivecommons/spektacular#126 is unreleased), so every live meeting child
printed cobra's unknown-command envelope and exited 1. The stream
restarted it five times per meeting, each restart a WARN, which the
slog feed self-reported as unclassified_error_repeated (#18630), and the
session warned once per transcript line it could not forward.

- spektacular.Stream gains an optional Probe (`meeting --help`), run
  once per installed binary; an unknown command marks the stream
  Unsupported: no child is started or restarted, one INFO line per
  binary. A child whose own exit output says unknown command is not
  restarted either.
- Restart lines are WARN at most once per BackoffMax window (INFO
  otherwise) and map to a new log-fed class spek_stream_unavailable.
- The live session treats an unsupported Spek as off from the start
  (panel status unchanged), stops forwarding once Spek is gone, and warns
  about a failed forward only on the transition to paused.

Fixes #18630

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* ✨ Meetings: delete finished meetings and stale join requests

Owner-only DELETE /api/meetings/{id} for meetings in a final state and
pending join requests (declined, then removed); 409 while hive is still
in the meeting. Removes the record, transcript and live state, erases
the meeting in Vexa with Quorum meeting.delete (best effort, surfaced as
upstream_error) and drops the Calendar entry that existed only for it.
POST /api/meetings/clear-finished deletes all finished meetings older
than now; DELETE /api/meetings/schedule/{eid} deletes a past Calendar
entry with its meetings. Knowledge facts and drafts are not touched.
Audited as meeting_delete / meeting_schedule_delete.

UI: πŸ—‘ Delete on finished and join-request rows, Clear finished, and
πŸ—‘ Delete on past Calendar entries, all behind the in-page confirm.

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ› Meetings delete: use s.meetingsError after the v5 merge (#18600)

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ› Meetings delete: clear consent tries and load the delete helpers in the role-race test

Carries the fixes from #18615's v5 merge: Delete also forgets the
consent attempt count, and the meetings/calendar role-race harness
extracts meetingsFinished, meetingsCanDelete, meetingsDeleteButtonHtml,
calendarCanDelete and their constants.

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* ✨ Betty runs several model calls per meeting at once

A meeting may now have up to meetings.llm.per_meeting_in_flight (default
3, never above max_in_flight) model calls in flight, so a second request
no longer waits for the first answer. A per-meeting Gate bounds the calls
and reserves each in-flight call's expected cost against the token
budget, so concurrent calls cannot all pass the check on the same ledger
total. On a warm CLI backend the meeting keeps a pool of warm CLIs (one
turn per process, since stream results carry no request id) that grows
only when calls overlap.

Part of #18590.

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* 🌱 Show meetings.llm.per_meeting_in_flight in the meetings settings

The Settings parity guard needs every hive.yaml leaf referenced by the
Settings UI; the meetings settings' effective values now include it.

Part of #18590.

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* ✨ Betty names who and what each answer is for when requests overlap

With more than one request in flight, other lines said since the
request, or an answer later than 8 s, each answer opens with the asker's
display name (first name when spoken, never an @mention) and a short
topic phrase: the model's new "topic" label, else the request's own
words. Two or more requests in hand post one queue note per minute
(chat, rate limited, not charged), and spoken acks name their asker
when others are waiting.

Part of #18590.

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* ✨ Betty speaks ready answers shortest first and moves a long voice queue to chat

Quorum's say queues clips and plays them in order, so hive keeps an
estimate of the queued speech. Spoken answers ready in the same loop turn
are said back to back, shortest first; one that would wait behind more
than 20 s goes to chat with one spoken "I've put the rest in chat"; an
ack behind more than 8 s of speech is skipped. Answers to addressed
requests may take meetings.chat_answer_burst (3) extra chat posts a
minute past the chat rate.

Part of #18590.

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* 🌱 Show meetings.chat_answer_burst in the meetings settings

Part of #18590.

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* ✨ Betty times every answer stage and answers live-state questions without waiting for search

Each answer carries heard, retrieval, model and answer-ready timings with
its source and budget (4 s for the hive's live state, 8 s otherwise); the
live session adds delivery and say times and keeps the last 20 with a
latency summary in the live state served by the meeting detail API.
Gather stops waiting once the hive status retriever answers.

Part of #18590.

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* ✨ Betty cites the sources of chat answers and says when she is not sure

Chat answers end with up to two sources the model reported using (web
link, work item id, knowledge fact, code map or hive live status); the
prompt asks for issue and PR numbers and forbids guessing; a
low-confidence addressed answer not from the live state is hedged.

Part of #18590.

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ“– Refresh meetings API doc citations for the combined Betty changes

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ› Betty attribution tests expect the source a chat answer now cites

With #18631's citations combined, an attributed chat answer ends with its
source, so the attribution tests compare against that full text.

Signed-off-by: Andy Anderson <andy@clubanderson.com>

---------

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* ✨ Sessions: per-session git worktrees with a locked shared git config (#18638)

* ✨ Sessions: per-session git worktrees with a locked shared git config

Each Sessions pane opens in its own git worktree of the hive's one shared
clone, on branch hive-session/<id>, so each session has its own Spektacular
workflow state, checked-out branch and uncommitted work (#18588).

The shared clone's git config and hooks are locked: root rewrites
.git/config from scratch (no git as root) as root:<gid> 0644 with
extensions.worktreeConfig, empties hooks/, info/ and modules/ into root
0755 dirs, places root-owned config.worktree and commondir placeholders,
keeps objects/refs/logs group-writable, and makes .git, worktrees/ and the
clone 3770. It runs before any worktree of a clone is handed out and once
per clone at boot. A session's worktree entry is created with no group
write; panes get no safe.directory, so git refuses a swapped .git. The
hive's own git always passes -c core.hooksPath=/dev/null
-c core.fsmonitor=false.

Existence checks run in the add/remove scripts as the session uid (the hive
runs as dev, outside the sessions group): a resume is a no-op, a missing
clone is reported, a worktree whose .git is not the session's own is
refused. Removal auto-saves uncommitted work on the session branch first.
The worktree sweep does not run while the sessions store failed to load or
an unreadable copy is set aside.

Builds on the design in #18613 by Karibhou.

Fixes #18588

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* 🌱 Sessions: split worktree code into its own file, re-pin API citations

Moves the worktree scripts and functions out of usersession_workspace.go
(over the 800-line crux limit) into usersession_worktree.go, updates the
add-count test (the add script runs on every start, as the session uid),
and re-pins the sessions API citations with check-docs-citations.py --fix.

Signed-off-by: Andy Anderson <andy@clubanderson.com>

Co-authored-by: Danathar <6772335+Danathar@users.noreply.github.com>

* πŸ› fix: wait for seed queue before test cleanup (#18641)

* πŸ› fix: wait for seed queue before test cleanup

Wait for the seed queue worker to finish its session store writes before temporary directory cleanup.

Hive-Run: #18454
Hive-Plan: dashboard-seed-cleanup
Hive-Spec: issue-18454#dashboard-test-cleanup
Signed-off-by: Karibhou <157208370+Karibhou@users.noreply.github.com>

* fix: wait for all seed test queue workers

Signed-off-by: Karibhou <157208370+Karibhou@users.noreply.github.com>

* test: share pending seed cleanup

Signed-off-by: Karibhou <157208370+Karibhou@users.noreply.github.com>

* ✨ meetings: per-user desktop companion tokens and hub passthrough (#18643)

Each owner or read-write user mints a companion token on /companion/token
(shown once, SHA-256 stored, one per user). authenticate admits it only on
/api/companion/info and /api/companion/meetings/*, as the minting user at
their live role (refused below read-write). The hub waves those routes
through to the spoke, which does the check; the token page stays gated.

Refs #18596

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ”– release: v5.294.0

Automated release commit. Compiles changelog.d/ fragments and moves
the CHANGELOG.md Unreleased section into a dated v5.294.0 entry.
See src/docs/releases.md.

Signed-off-by: hive-release-bot <actions@github.com>

* 🌱 images: bump omp 18.20.0 -> 18.22.0 (#18647)

Automated pin bump (#8419). Digests recomputed from the downloaded
release artifacts by src/scripts/cli-pin-bump.sh:

  VERSION=18.22.0
  SHA256_AMD64=f706b27ba845834d8f8360705a2e9ba1534f3593fa24411d41358272de33642b
  SHA256_ARM64=7238dd213d5c61bc9bf51735d552bc71901fdd44329f862b0a971146b1cdac3e

Signed-off-by: hive-release[bot] <hive-release@hive.kubestellar.io>
Signed-off-by: clubanderson <407614+clubanderson@users.noreply.github.com>

* 🌱 images: bump claude 2.1.325 -> 2.1.326 (#18648)

Automated pin bump (#8419). Digests recomputed from the downloaded
release artifacts by src/scripts/cli-pin-bump.sh:

  VERSION=2.1.326

Signed-off-by: hive-release[bot] <hive-release@hive.kubestellar.io>
Signed-off-by: clubanderson <407614+clubanderson@users.noreply.github.com>

* 🌱 ci: size overlayfs-exec-guard timeout for the cold whisper-cache path (#18604)

The guard was cancelled at its 10-minute timeout whenever the
whisper-linux-amd64 gha cache missed. The repo's Actions cache runs at
12-23 GB against a 10 GB cap, so LRU eviction regularly drops the
whisper-builder blobs; the hosted 4-vCPU runner then compiles FFmpeg and
whisper.cpp from source (~8.5 min serial chain) and the job needs ~13 min.
Warm runs measure 6.5-8 min and are unaffected.

Raise the timeout to 20 min (cold path plus ~50% headroom) and skip the
buildx post-step teardown on the ephemeral runner (16-36s per run).
Regenerated the integration workflows from v2-ci.yml.

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ› Isolate restart kick test tmux sessions (#18650)

Avoid collisions with delayed tmux operations from shuffled tests and drain the readiness helper before session teardown.

Hive-Run: #18626
Hive-Plan: restart-kick-flake
Hive-Spec: restart-kick#readiness
Signed-off-by: Karibhou <Karibhou@users.noreply.github.com>

* πŸ”– release: v5.295.0

Automated release commit. Compiles changelog.d/ fragments and moves
the CHANGELOG.md Unreleased section into a dated v5.295.0 entry.
See src/docs/releases.md.

Signed-off-by: hive-release-bot <actions@github.com>

* ✨ feat: range attribution core for post-merge runs (#18653)

* ✨ feat: range attribution core for post-merge runs

Pure, GitHub/Tekton-agnostic AttributeRuns in workflowmodel: attributes
post-merge runs per pipeline/environment to merged PRs by range prev..R,
first success wins, failure holds until a later success covers it, and
undeterminable revisions or ranges stay unknown with a recorded reason.

Part of #18356. Closes #18644.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Signed-off-by: scanner <280983584+hivecommons-hive[bot]@users.noreply.github.com>

* changelog: fragment for #18653

Generated by changelog-fragment-guard from the PR title (#5675).
Edit the wording freely, or replace this file with the no-changelog
label if the change is not user-visible.

Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
Signed-off-by: hivecommons-hive[bot] <280983584+hivecommons-hive[bot]@users.noreply.github.com>

* πŸ”– release: v5.296.0

Automated release commit. Compiles changelog.d/ fragments and moves
the CHANGELOG.md Unreleased section into a dated v5.296.0 entry.
See src/docs/releases.md.

Signed-off-by: hive-release-bot <actions@github.com>

* ✨ feat: Tekton PipelineRun post-merge observer (#18655)

* [scanner] fix: Tekton PipelineRun post-merge observer (#18646)

Add an explicit tektonRuns config on Tekton pipelines (kube context or
in-cluster, namespace, label selector, optional revision annotation/param
names) and an observer that lists PipelineRuns, reads each run's revision
(pipelinesascode.tekton.dev/sha annotation, then the revision param) and
conclusion, and feeds them through workflowmodel.AttributeRuns. A run
whose revision is not a full commit SHA keeps an empty revision with the
reason recorded, so attribution leaves the column unknown.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: scanner <280983584+hivecommons-hive[bot]@users.noreply.github.com>

* [scanner] test: cover tekton observer error paths for the coverage floor

Signed-off-by: scanner <280983584+hivecommons-hive[bot]@users.noreply.github.com>

* docs: say the Tekton observer is not wired into the stage engine yet

Signed-off-by: scanner <280983584+hivecommons-hive[bot]@users.noreply.github.com>

* ✨ feat: GitHub Actions post-merge run observer for pipeline.status (#18656)

* [scanner] fix: GitHub Actions post-merge run observer feeding pipeline.status

Observe default-branch push runs (Main CI, exact merge SHA when the
range cannot decide), release and deploy workflow runs of the workflows
an ingested model binds, attribute them to merged PRs through
workflowmodel.AttributeRuns and carry the result in
PRFacts.PostMerge so pipeline.status reports green/red/running instead
of always unknown. Never guess: pull-request and merge-queue runs never
count, a deploy run without its environment credits none of the bound
environments, and an undeterminable revision or range stays unknown
with the reason recorded.

Closes #18645

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: scanner <280983584+hivecommons-hive[bot]@users.noreply.github.com>

* docs: say the Actions observer is not wired into the sweep yet

Signed-off-by: scanner <280983584+hivecommons-hive[bot]@users.noreply.github.com>

* [scanner] fix: wire post-merge observers into pipeline.status facts (#18662)

The workflow simulate's merged-PR fill now calls the GitHub Actions
post-merge observer (ObservePostMergeRuns) with the PR's merge commit and
default branch, so pipeline.status reflects observed runs instead of
always being unknown. Refs #18356.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: scanner <280983584+hivecommons-hive[bot]@users.noreply.github.com>
Co-authored-by: clubanderson <407614+clubanderson@users.noreply.github.com>

* πŸ”– release: v5.297.0

Automated release commit. Compiles changelog.d/ fragments and moves
the CHANGELOG.md Unreleased section into a dated v5.297.0 entry.
See src/docs/releases.md.

Signed-off-by: hive-release-bot <actions@github.com>

* πŸ”– release: v5.298.0

Automated release commit. Compiles changelog.d/ fragments and moves
the CHANGELOG.md Unreleased section into a dated v5.298.0 entry.
See src/docs/releases.md.

Signed-off-by: hive-release-bot <actions@github.com>

* [scanner] fix: stop retrying closing-issue link on GitHub manual reference limit (#18668)

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Signed-off-by: scanner <280983584+hivecommons-hive[bot]@users.noreply.github.com>

* πŸ”– release: v5.298.1

Automated release commit. Compiles changelog.d/ fragments and moves
the CHANGELOG.md Unreleased section into a dated v5.298.1 entry.
See src/docs/releases.md.

Signed-off-by: hive-release-bot <actions@github.com>

* 🌱 ci: move whisper-builder cache to GHCR registry, stop wasted gha exports (#18674)

The repo's Actions cache is over the 10 GB cap, so LRU eviction kept
dropping the whisper-builder blobs and builds recompiled whisper.cpp and
FFmpeg (slow for arm64 under QEMU).

- whisper-builder warm step: cache-from/cache-to a dedicated GHCR registry
  cache (hive-buildcache:whisper-<arch>), exported only by push builds.
  Same-repo PR builds log in and read it only; fork PRs never run the job.
- hub and mx builds: drop the gha cache. Both stages of each Dockerfile are
  in no-cache-filters, so the mode=max export could never be hit.
- contributor build: export the gha cache only on push builds, so PR refs
  stop adding entries that only that PR can see.
- overlayfs-exec-guard (v2-ci and its generated copies): read whisper layers
  from the registry cache instead of the gha scope that is no longer written.

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ› Keep Copilot config.json linked to the shared login after the CLI's own save (#18673)

The Copilot CLI's atomic save replaces the per-agent config.json symlink
seconds after launch, so a later dashboard login never reached running
agents. The pane poller now relinks within one poll interval, merging
CLI-written non-credential fields into the shared file first (shared login
wins), serialized against the hive's token writers. The #18445 guard for a
private copy holding the only login is kept.

Fixes #18671

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ› ci: materialize warmed tmux/whisper stages so the main image build reuses them (#18687)

Measured on v5 after #18674: the whisper-builder warm step hit the GHCR
registry cache (all layers CACHED in about a second), but the main build
step then recompiled FFmpeg and whisper.cpp. Earlier runs show the same
for tmux and for the old gha whisper scope: a warm-step cache hit with no
output leaves nothing in the builder's local cache, and the main build was
only served when the warm step had missed and compiled locally.

Export each warmed stage to /dev/null so its layers are pulled and recorded
in the local builder cache before the main build runs.

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ“– Document kick refusals and link the docs from the dashboard (#18695)

Fixes #18692

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ”– release: v5.298.2

Automated release commit. Compiles changelog.d/ fragments and moves
the CHANGELOG.md Unreleased section into a dated v5.298.2 entry.
See src/docs/releases.md.

Signed-off-by: hive-release-bot <actions@github.com>

* DCO Remediation Commit for Andy Anderson <andy@clubanderson.com>

I, Andy Anderson <andy@clubanderson.com>, hereby add my Signed-off-by to this commit: f444b4f
I, Andy Anderson <andy@clubanderson.com>, hereby add my Signed-off-by to this commit: 3d8d438

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ› Workflow dispatch: start on a standard hive and show a profile no agent fills (#18691)

* πŸ› Workflow dispatch: fall back to the hive's issue worker or reviewer, and show a profile nothing fills

A bound workflow model hands some states to an agent profile, but the
resolver only accepted a pinned agent, an agent named after the profile,
or a writer, docs or feature agent. A standard roster has none of those,
so the drafter never started and nothing in the dashboard said why.

- config.ResolveWorkflowDispatch ends with a last resort: the issue
  worker (scanner role, classify.DefaultLane) for a profile that opens a
  pull request, the reviewer for one that reviews, in both cases only an
  agent that serves the bound repository. No other role is picked.
- GET /api/workflows lists each bound profile's agent or the reason
  there is none; the Settings -> Repos workflow row warns and offers a
  picker that saves workflow.dispatch.profiles.<profile>.agent through
  PUT /api/workflows/dispatch/profiles.
- The stage tracker carries the reason onto the item's row, and the
  manual send answers with it and where to fix it.

Fixes #18690

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ› Document the profile pin route in the OpenAPI spec; keep the gap badge rule out of the sentinel CSS block

Fixes #18690

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* Third-Party DCO Remediation Commit for hivecommons-hive[bot] <280983584+hivecommons-hive[bot]@users.noreply.github.com>

On behalf of hivecommons-hive[bot] <280983584+hivecommons-hive[bot]@users.noreply.github.com>, I, Andy Anderson <andy@clubanderson.com>, hereby add my Signed-off-by to this commit: 0b8ad82
On behalf of hivecommons-hive[bot] <280983584+hivecommons-hive[bot]@users.noreply.github.com>, I, Andy Anderson <andy@clubanderson.com>, hereby add my Signed-off-by to this commit: 02d2850

Their squash commits carry a Signed-off-by for a different identity than the author.

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* Third-Party DCO Remediation Commit for Andy Anderson <andy@clubanderson.com>

On behalf of Andy Anderson <andy@clubanderson.com>, I, Andy Anderson <andy@clubanderson.com>, hereby add my Signed-off-by to this commit: f444b4f
On behalf of Andy Anderson <andy@clubanderson.com>, I, Andy Anderson <andy@clubanderson.com>, hereby add my Signed-off-by to this commit: 3d8d438

Their Signed-off-by lines name the release bot rather than the author address.

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* DCO Remediation Commit for Andy Anderson <andy@clubanderson.com>

I, Andy Anderson <andy@clubanderson.com>, hereby add my Signed-off-by to this commit: f444b4f

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* DCO Remediation Commit for Andy Anderson <andy@clubanderson.com>

I, Andy Anderson <andy@clubanderson.com>, hereby add my Signed-off-by to this commit: 3d8d438

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* 🌱 ci: let the DCO app honour an author's own remediation commit (#18703)

* 🌱 ci: let the DCO app honour an author's own remediation commit

Adds .github/dco.yml with allowRemediationCommits.individual: true so an
author can vouch for their own earlier commits whose author email had no
matching Signed-off-by, which is what blocks v5 -> v6 top-ups that carry
image-bump commits (#18697). Third-party remediation stays off.

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* changelog: fragment for #18703

Generated by changelog-fragment-guard from the PR title (#5675).
Edit the wording freely, or replace this file with the no-changelog
label if the change is not user-visible.

Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
Signed-off-by: clubanderson <407614+clubanderson@users.noreply.github.com>

* 🌱 changelog: make the dco.yml fragment a single entry bullet

The fragment is the entry itself and must start with '- '; the compiler owns
the headings. Drop the duplicate fragment the changelog bot added for #18703.

Signed-off-by: Andy Anderson <andy@clubanderson.com>

---------

Signed-off-by: Andy Anderson <andy@clubanderson.com>
Signed-off-by: clubanderson <407614+clubanderson@users.noreply.github.com>
Co-authored-by: clubanderson <407614+clubanderson@users.noreply.github.com>
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>

---------

Signed-off-by: Andy Anderson <andy@clubanderson.com>
Signed-off-by: Karibhou <157208370+Karibhou@users.noreply.github.com>
Signed-off-by: hive-release-bot <actions@github.com>
Signed-off-by: hive-release[bot] <hive-release@hive.kubestellar.io>
Signed-off-by: clubanderson <407614+clubanderson@users.noreply.github.com>
Signed-off-by: Karibhou <Karibhou@users.noreply.github.com>
Signed-off-by: hivecommons-hive[bot] <280983584+hivecommons-hive[bot]@users.noreply.github.com>
Signed-off-by: scanner <280983584+hivecommons-hive[bot]@users.noreply.github.com>
Co-authored-by: Karibhou <157208370+Karibhou@users.noreply.github.com>
Co-authored-by: hive-release-bot <actions@github.com>
Co-authored-by: clubanderson <407614+clubanderson@users.noreply.github.com>
Co-authored-by: scanner <280983584+hivecommons-hive[bot]@users.noreply.github.com>
Co-authored-by: Danathar <6772335+Danathar@users.noreply.github.com>
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
kubestellar-prow Bot pushed a commit that referenced this pull request Nov 18, 2026
* πŸ› meetings: log every failed meetings call and keep a failed approval's reason (#18600)

* πŸ› meetings: log every failed meetings call and keep a failed approval's reason

A Quorum refusal of a Discord join (blocked: no permission to connect)
was answered as 502 with no log line, and the hosted ingress replaced the
body with its generic page, so the owner saw 'hive unavailable' and the
spoke log said nothing. Every meetings call failure is now logged with
method, path, status and error (Warn for 5xx, Info for caller mistakes),
and ApproveRequest records the failure reason on the still-open request.

Refs #18598

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ› meetings: a blocked Discord join names the bot's permissions, and a failed approval shows on the request

The panel read every blocked meeting as 'The meeting refused a guest bot',
which is a Meet/Teams lobby refusal; on Discord blocked means the bot's
role lacks View Channel, Connect or Speak on the voice channel. The join
request row now also renders the reason a failed approval saved.

Refs #18598

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ“– api-reference: re-point meetings route citations after the import shift

Refs #18598

Signed-off-by: Andy Anderson <andy@clubanderson.com>

---------

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* ✨ Remote items: follow up on issues and PRs the hive opened as its App bot (#18610)

* ✨ remote items: follow up on items the hive opened as its App bot

With project.remote_write_identity a write-tier reference repo can write as
the App bot, but the own-item poller listed creator=<owner> under the
owner token only, so feedback on bot-opened issues and PRs never reached
the agent.

- remoteitems: RunTargets takes (repo, login) pairs; each repo is listed and
  its feedback filtered under its own login. Run keeps its signature as a
  one-login wrapper.
- github.RemoteItemsFetcher: optional TokenFor picks the read token per
  repo and is authoritative when set (no fallback to the owner token).
- cmd/hive: every write-tier repo is watched under its CredentialLogin and
  read with the token the proxy injects on its writes (scoped bot token,
  owner token or write_token_file).

Refs #18516

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ› remote items: fix the per-repo login test's use of Store.Items

Store.Items returns a slice; build the tracked set from Item.Key.

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ“– two-lane doc: the own-item follow-up covers bot-opened items too

Signed-off-by: Andy Anderson <andy@clubanderson.com>

---------

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ› fix(ioscan): do not fleet-report an injection-withheld kick as a classifier outage (#18618)

A fail-closed kick block was always logged as "ioscan fail-closed blocked
kick", which the fleet report classes as ioscan_classifier_unavailable. But
the kick is also withheld when untrusted issue/PR text trips the injection
policy (a Critical finding or a classifier block verdict) while the
classifier is answering: the defense working as designed, not an outage.
The #18459 report carried no classifier_error, i.e. exactly that case.

warnBlockedKick now keeps the fleet-reported wording (with classifier_error
and fix) only when the classifier is failing, and otherwise logs "ioscan
fail-closed withheld kick on an injection finding", listed in
fleetreport.IsBenignMessage. The held-PR coordination path goes through the
same helper, so it now also names the classifier error on an outage.

Closes #18459

Hive-Run: hivecommons/hive#18459
Hive-Plan: selfreport-ioscan_classifier_unavailable
Hive-Spec: fleet-report#ioscan_classifier_unavailable
Signed-off-by: Karibhou <157208370+Karibhou@users.noreply.github.com>

* 🌱 ci: fetch only the histories the squash-attribution check reads (#18609)

`fetch-depth: 0` makes actions/checkout fetch every branch and tag. That
took 17-35s normally and on #18482 (run 43960382195) hung for the whole
5-minute job budget, failing the required "Sign-off survives the squash"
context before the check ran.

Check out the head shallowly, then fetch the head, base and default-branch
histories in full (unshallow), with a 60s per-attempt stall timeout and
three attempts. Picks up the fix a contributor reverted out of #18482 as
unrelated to that PR, plus the stall retry and post-fetch assertions.

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* ✨ Hand workflow-model states to their agent profiles (claim first, dispatch once) (#18571)

* ✨ Hand workflow-model states to their agent profiles

The eval cycle now reads the bound model: an item whose state the model
hands to a profile (TechDocs Accepted -> techdocs-drafter, Drafting on
/verify -> techdocs-verifier) is withheld from every generic lane, claimed
for the agent that fills the profile on the issue-claims ledger, and kicked
once. The hand-off key (item, profile, state, triggering comment) is
written to workflow-dispatch.json on delivery, so neither a later cycle
nor a restart repeats it; a failed or deferred kick is retried.

Part of the drafter hand-off for cncf/techdocs#395.

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ› List a bound repository's drafts at most every five minutes

Signed-off-by: Andy Anderson <andy@clubanderson.com>

---------

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ› Betty closes the loop on deep dives; research agents survive config reloads (#18592)

* πŸ› Betty closes the loop on deep dives; research agents survive reconciles

research-1..N are added to the runtime only (not hive.yaml), and
ReconcileAgents (config reload, reviewer autoscale) removes every agent
not in the enabled config set, so the pool was emptied shortly after boot
and each meeting research request failed with "no research agent
available" while nothing was said in the meeting.

- cmd/hive: the reconcile set carries the research pool's agents.
- research.Pool: a dispatch that finds no agent re-registers them once;
  ErrNoAgents carries a sayable Reason; Status() for the settings page.
- live: a persona deep dive that cannot start, errors, or runs past
  5 minutes is answered with "I couldn't look that up: <reason>." in
  the channel it was asked in; one progress note after 30 s; a spoken
  question's follow-up is spoken.
- Settings, Meeting attendee: Research section with agents, availability
  and a "No research agent available" warning.

Part of hivecommons/hive#18590

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ› tests: set agents through AgentRuntimeConfig

Part of hivecommons/hive#18590

Signed-off-by: Andy Anderson <andy@clubanderson.com>

---------

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* ✨ Review follow-ups, trust-only 090 and issue pr-link notes become status-card sections (quiet mode) (#18616)

* ✨ Review follow-ups, trust-only 090 and issue pr-link notes become status-card sections (quiet mode)

In github.notifications.mode: quiet the review Follow-ups checklist is the
collapsible review-followups section of the PR's hive:status card (ticks
kept via the new prstatus.WriteFunc), a 090 ready-for-human-merge entry
whose only objection is the author's trust tier is the stage-ready line,
and the issue-side PR-link stage note is the issue card's issue-pr-link
section. Every 085 and a held or manual-merge 090 still post a comment.

Fixes #18606

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ› status-card follow-ups use their own marker; tests expect neutralised mentions

The card section carried the legacy Follow-ups comment marker, so the next
pass read the card as the legacy comment and EditIssueMarkerComment
rewrote the whole card. Stamp the section with
<!-- hive:review-followups-card --> and skip status comments in the legacy
check. The 090 card line goes through prstatus.Write, which neutralises
'@sha' to a code span; the test now expects that.

Signed-off-by: Andy Anderson <andy@clubanderson.com>

---------

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* 🌱 ci: no mutation comment on error verdicts; no reopen hint for bot or maintainer reporters (#18617)

* 🌱 ci: no mutation comment for an error verdict; no reopen hint for bot or maintainer reporters

mutation-report.sh post no longer creates a <!-- hive-mutation --> comment
for an 'error' verdict (no report, nothing to act on): the report stays in
the job summary and the failing verdict step, and an earlier finding comment
is edited so it is not stale. issue-closed-reopen-hint.yml skips issues
whose reporter is a bot or OWNER/MEMBER/COLLABORATOR. pr-overlap.sh already
posts only when a sibling PR overlaps (#17870); unchanged.

Fixes #18607

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* 🌱 changelog fragment for #18607

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* 🌱 re-pin TRUSTED_SCRIPTS_TREE for the mutation-report.sh change

Signed-off-by: Andy Anderson <andy@clubanderson.com>

---------

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ› Meetings: keep nothing from before the consent notice is delivered (#18621)

Until the consent notice is delivered (consent_posted_at), every
transcript segment, draft and chat line is discarded in the manager:
not stored and not emitted to listeners (Betty, Spektacular, research,
facts, Inbox hand-off). Only a pre_consent_dropped count is kept. A
leave phrase is still honoured and the line itself is not kept.

Lines that arrive after the notice but belong before it are dropped
too: seq at or below Quorum's transcript cursor when the notice landed
(consent_seq), chat stamped earlier, and for two minutes segments that
began before the last speech heard while pending (consent_speech_end).
A reconnect or restart while pending asks Quorum for no replay, and
dropped lines advance the replay cursor.

Refs #18590

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ› fix: stop open-PR guard label retry storms and classify the failure (#18624)

The has-open-pr label write shared the caller's 15s deadline with the
open-PR listing ahead of it, so a slow listing of a large repo left it a
spent context. The write failed with context deadline exceeded, the issue
never got the label, and every later claim pass retried it and warned
again, which filed an unclassified_error_repeated self-report.

- give the label write its own 10s budget, detached from the caller
- remember a failed write per issue: 15m after a transient failure, 6h
  after a refusal (403/404/410/422), and log refusals at INFO
- add the open_pr_guard_label_failed fleet-report class for the
  remaining repeated WARNs

Fixes #18589

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* 🌱 ci: blobless history fetch for the squash-attribution check (#18634)

#18609's unshallow step fetched full blob history and measured 38-42s on
the self-hosted pool, slower than the all-refs checkout it replaced. The
checker reads only commits and trees, so fetch with --filter=blob:none
(measured 84s -> 14s and 277 MB -> 24 MB locally for the same refs).

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ› Deflake TestSessionCreateWithSeedWaitsForRunning (#18632)

The test returned once the instruction was typed, while the queue worker
and seed delivery were still writing (touch, audit, inbox routing). TempDir
cleanup then raced those writes (directory not empty). Wait for the audit
and for the queue worker to exit before the test ends.

Fixes #18545

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ”– release: v5.293.0

Automated release commit. Compiles changelog.d/ fragments and moves
the CHANGELOG.md Unreleased section into a dated v5.293.0 entry.
See src/docs/releases.md.

Signed-off-by: hive-release-bot <actions@github.com>

* ✨ Owner-only Send to <profile> on model-owned stage-tracker rows (#18576)

* ✨ Owner-only Send to <profile> on model-owned stage-tracker rows

A row of a model-bound repository whose state the model hands to an agent
profile offers the owner Send to <profile> in its menu, backed by the
owner-only POST /api/workflows/dispatch: the item is read fresh, the
agent resolved (a profile disabled for automatic dispatch still
resolves), the issue claimed first and the agent kicked; the hand-off is
recorded once the kick lands so the automatic pass does not repeat it.
Audited workflow_dispatch; openapi and api-reference updated.

Part of the drafter hand-off for cncf/techdocs#395.

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ› Register the kicked agent in the dispatch endpoint test runtime

Signed-off-by: Andy Anderson <andy@clubanderson.com>

---------

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ“– docs: Betty on Zoom without a Zoom app, options and recommendation (#18602)

Record the #18596 comparison as a design page: phone dial-in, a desktop
companion that streams the user's own call audio, and Zoom's RTMP live
stream, each against the free-account, bot-rule, consent and cost
constraints. Recommends the desktop companion (macOS first) and records
its proposed streaming contract, per-user companion token, distribution
plan and prototype PR set. Indexed in docs/design/README.md and linked
from meetings.md, which now notes that Zoom refuses web-client bots.

Hive-Run: hivecommons/hive#18596
Hive-Plan: 18596 recommendation and prototype PR set
Hive-Spec: hivecommons/hive#18596#recommendation

Signed-off-by: Karibhou <157208370+Karibhou@users.noreply.github.com>
Co-authored-by: clubanderson <407614+clubanderson@users.noreply.github.com>

* πŸ› hub: converge cluster host_aliases on every hosted spoke, push-reported clusters included (#18639)

The host_aliases reconcile sweep (#17805) patches Deployments over kubectl,
but vllm-d and a-ks-wec2 are push-reported (pull_only) clusters the hub
cannot reach, so the sweep skipped every hive there silently. Only spokes
rendered from the template after #17805 carried the aliases.

The hub now sends the SaaS record cluster's host_aliases on every heartbeat
response and the spoke converges its own Deployment with a JSON patch
(exact for the managed hostnames, unrelated aliases kept, throttled to one
check per 15 minutes). The hub sweep reports push-reported clusters and
WARNs on unreadable or unreachable Deployments instead of logging at Debug.

Fixes #18637

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* ✨ meetings: desktop companion meetings (/api/companion/*) (#18611)

* ✨ meetings: desktop companion meetings (/api/companion/*)

A companion meeting is fed by the desktop companion: a person's own meeting
client audio (call) and microphone (mic), captured on their computer after a
versioned consent confirmation, relayed as 16 kHz PCM to Quorum's companion
platform (quorum 0.3.0). Betty's replies go to the companion window through
Quorum's outbox. Owner or read-write; a companion meeting is usable only by
its starter (or an owner).

Refs #18596

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ“– docs: rewrite api-reference citations shifted by the companion routes

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ› meetings: companion handlers use the logging meetingsError (#18600)

Signed-off-by: Andy Anderson <andy@clubanderson.com>

---------

Signed-off-by: Andy Anderson <andy@clubanderson.com>
Co-authored-by: scanner <280983584+hivecommons-hive[bot]@users.noreply.github.com>

* 🌱 deps: quorum v0.3.0 (desktop companion platform) (#18622)

quorum.mjs sha256 from the v0.3.0 release's quorum.mjs.sha256; davey stays
0.1.12 (quorum's package-lock.json).

Refs #18596

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* ✨ Betty: parallel answers with attribution, sources and timings, voice queue, delete finished meetings, quiet Spektacular restarts (#18635)

* πŸ› Meetings: don't restart-loop the Spektacular stream child when live mode isn't available

The pinned spektacular (0.31.0) has no `meeting` command
(hivecommons/spektacular#126 is unreleased), so every live meeting child
printed cobra's unknown-command envelope and exited 1. The stream
restarted it five times per meeting, each restart a WARN, which the
slog feed self-reported as unclassified_error_repeated (#18630), and the
session warned once per transcript line it could not forward.

- spektacular.Stream gains an optional Probe (`meeting --help`), run
  once per installed binary; an unknown command marks the stream
  Unsupported: no child is started or restarted, one INFO line per
  binary. A child whose own exit output says unknown command is not
  restarted either.
- Restart lines are WARN at most once per BackoffMax window (INFO
  otherwise) and map to a new log-fed class spek_stream_unavailable.
- The live session treats an unsupported Spek as off from the start
  (panel status unchanged), stops forwarding once Spek is gone, and warns
  about a failed forward only on the transition to paused.

Fixes #18630

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* ✨ Meetings: delete finished meetings and stale join requests

Owner-only DELETE /api/meetings/{id} for meetings in a final state and
pending join requests (declined, then removed); 409 while hive is still
in the meeting. Removes the record, transcript and live state, erases
the meeting in Vexa with Quorum meeting.delete (best effort, surfaced as
upstream_error) and drops the Calendar entry that existed only for it.
POST /api/meetings/clear-finished deletes all finished meetings older
than now; DELETE /api/meetings/schedule/{eid} deletes a past Calendar
entry with its meetings. Knowledge facts and drafts are not touched.
Audited as meeting_delete / meeting_schedule_delete.

UI: πŸ—‘ Delete on finished and join-request rows, Clear finished, and
πŸ—‘ Delete on past Calendar entries, all behind the in-page confirm.

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ› Meetings delete: use s.meetingsError after the v5 merge (#18600)

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ› Meetings delete: clear consent tries and load the delete helpers in the role-race test

Carries the fixes from #18615's v5 merge: Delete also forgets the
consent attempt count, and the meetings/calendar role-race harness
extracts meetingsFinished, meetingsCanDelete, meetingsDeleteButtonHtml,
calendarCanDelete and their constants.

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* ✨ Betty runs several model calls per meeting at once

A meeting may now have up to meetings.llm.per_meeting_in_flight (default
3, never above max_in_flight) model calls in flight, so a second request
no longer waits for the first answer. A per-meeting Gate bounds the calls
and reserves each in-flight call's expected cost against the token
budget, so concurrent calls cannot all pass the check on the same ledger
total. On a warm CLI backend the meeting keeps a pool of warm CLIs (one
turn per process, since stream results carry no request id) that grows
only when calls overlap.

Part of #18590.

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* 🌱 Show meetings.llm.per_meeting_in_flight in the meetings settings

The Settings parity guard needs every hive.yaml leaf referenced by the
Settings UI; the meetings settings' effective values now include it.

Part of #18590.

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* ✨ Betty names who and what each answer is for when requests overlap

With more than one request in flight, other lines said since the
request, or an answer later than 8 s, each answer opens with the asker's
display name (first name when spoken, never an @mention) and a short
topic phrase: the model's new "topic" label, else the request's own
words. Two or more requests in hand post one queue note per minute
(chat, rate limited, not charged), and spoken acks name their asker
when others are waiting.

Part of #18590.

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* ✨ Betty speaks ready answers shortest first and moves a long voice queue to chat

Quorum's say queues clips and plays them in order, so hive keeps an
estimate of the queued speech. Spoken answers ready in the same loop turn
are said back to back, shortest first; one that would wait behind more
than 20 s goes to chat with one spoken "I've put the rest in chat"; an
ack behind more than 8 s of speech is skipped. Answers to addressed
requests may take meetings.chat_answer_burst (3) extra chat posts a
minute past the chat rate.

Part of #18590.

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* 🌱 Show meetings.chat_answer_burst in the meetings settings

Part of #18590.

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* ✨ Betty times every answer stage and answers live-state questions without waiting for search

Each answer carries heard, retrieval, model and answer-ready timings with
its source and budget (4 s for the hive's live state, 8 s otherwise); the
live session adds delivery and say times and keeps the last 20 with a
latency summary in the live state served by the meeting detail API.
Gather stops waiting once the hive status retriever answers.

Part of #18590.

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* ✨ Betty cites the sources of chat answers and says when she is not sure

Chat answers end with up to two sources the model reported using (web
link, work item id, knowledge fact, code map or hive live status); the
prompt asks for issue and PR numbers and forbids guessing; a
low-confidence addressed answer not from the live state is hedged.

Part of #18590.

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ“– Refresh meetings API doc citations for the combined Betty changes

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ› Betty attribution tests expect the source a chat answer now cites

With #18631's citations combined, an attributed chat answer ends with its
source, so the attribution tests compare against that full text.

Signed-off-by: Andy Anderson <andy@clubanderson.com>

---------

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* ✨ Sessions: per-session git worktrees with a locked shared git config (#18638)

* ✨ Sessions: per-session git worktrees with a locked shared git config

Each Sessions pane opens in its own git worktree of the hive's one shared
clone, on branch hive-session/<id>, so each session has its own Spektacular
workflow state, checked-out branch and uncommitted work (#18588).

The shared clone's git config and hooks are locked: root rewrites
.git/config from scratch (no git as root) as root:<gid> 0644 with
extensions.worktreeConfig, empties hooks/, info/ and modules/ into root
0755 dirs, places root-owned config.worktree and commondir placeholders,
keeps objects/refs/logs group-writable, and makes .git, worktrees/ and the
clone 3770. It runs before any worktree of a clone is handed out and once
per clone at boot. A session's worktree entry is created with no group
write; panes get no safe.directory, so git refuses a swapped .git. The
hive's own git always passes -c core.hooksPath=/dev/null
-c core.fsmonitor=false.

Existence checks run in the add/remove scripts as the session uid (the hive
runs as dev, outside the sessions group): a resume is a no-op, a missing
clone is reported, a worktree whose .git is not the session's own is
refused. Removal auto-saves uncommitted work on the session branch first.
The worktree sweep does not run while the sessions store failed to load or
an unreadable copy is set aside.

Builds on the design in #18613 by Karibhou.

Fixes #18588

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* 🌱 Sessions: split worktree code into its own file, re-pin API citations

Moves the worktree scripts and functions out of usersession_workspace.go
(over the 800-line crux limit) into usersession_worktree.go, updates the
add-count test (the add script runs on every start, as the session uid),
and re-pins the sessions API citations with check-docs-citations.py --fix.

Signed-off-by: Andy Anderson <andy@clubanderson.com>

Co-authored-by: Danathar <6772335+Danathar@users.noreply.github.com>

* πŸ› fix: wait for seed queue before test cleanup (#18641)

* πŸ› fix: wait for seed queue before test cleanup

Wait for the seed queue worker to finish its session store writes before temporary directory cleanup.

Hive-Run: hivecommons/hive#18454
Hive-Plan: dashboard-seed-cleanup
Hive-Spec: issue-18454#dashboard-test-cleanup
Signed-off-by: Karibhou <157208370+Karibhou@users.noreply.github.com>

* fix: wait for all seed test queue workers

Signed-off-by: Karibhou <157208370+Karibhou@users.noreply.github.com>

* test: share pending seed cleanup

Signed-off-by: Karibhou <157208370+Karibhou@users.noreply.github.com>

* ✨ meetings: per-user desktop companion tokens and hub passthrough (#18643)

Each owner or read-write user mints a companion token on /companion/token
(shown once, SHA-256 stored, one per user). authenticate admits it only on
/api/companion/info and /api/companion/meetings/*, as the minting user at
their live role (refused below read-write). The hub waves those routes
through to the spoke, which does the check; the token page stays gated.

Refs #18596

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ”– release: v5.294.0

Automated release commit. Compiles changelog.d/ fragments and moves
the CHANGELOG.md Unreleased section into a dated v5.294.0 entry.
See src/docs/releases.md.

Signed-off-by: hive-release-bot <actions@github.com>

* 🌱 images: bump omp 18.20.0 -> 18.22.0 (#18647)

Automated pin bump (#8419). Digests recomputed from the downloaded
release artifacts by src/scripts/cli-pin-bump.sh:

  VERSION=18.22.0
  SHA256_AMD64=f706b27ba845834d8f8360705a2e9ba1534f3593fa24411d41358272de33642b
  SHA256_ARM64=7238dd213d5c61bc9bf51735d552bc71901fdd44329f862b0a971146b1cdac3e

Signed-off-by: hive-release[bot] <hive-release@hive.kubestellar.io>
Signed-off-by: clubanderson <407614+clubanderson@users.noreply.github.com>

* 🌱 images: bump claude 2.1.325 -> 2.1.326 (#18648)

Automated pin bump (#8419). Digests recomputed from the downloaded
release artifacts by src/scripts/cli-pin-bump.sh:

  VERSION=2.1.326

Signed-off-by: hive-release[bot] <hive-release@hive.kubestellar.io>
Signed-off-by: clubanderson <407614+clubanderson@users.noreply.github.com>

* 🌱 ci: size overlayfs-exec-guard timeout for the cold whisper-cache path (#18604)

The guard was cancelled at its 10-minute timeout whenever the
whisper-linux-amd64 gha cache missed. The repo's Actions cache runs at
12-23 GB against a 10 GB cap, so LRU eviction regularly drops the
whisper-builder blobs; the hosted 4-vCPU runner then compiles FFmpeg and
whisper.cpp from source (~8.5 min serial chain) and the job needs ~13 min.
Warm runs measure 6.5-8 min and are unaffected.

Raise the timeout to 20 min (cold path plus ~50% headroom) and skip the
buildx post-step teardown on the ephemeral runner (16-36s per run).
Regenerated the integration workflows from v2-ci.yml.

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ› Isolate restart kick test tmux sessions (#18650)

Avoid collisions with delayed tmux operations from shuffled tests and drain the readiness helper before session teardown.

Hive-Run: hivecommons/hive#18626
Hive-Plan: restart-kick-flake
Hive-Spec: restart-kick#readiness
Signed-off-by: Karibhou <Karibhou@users.noreply.github.com>

* πŸ”– release: v5.295.0

Automated release commit. Compiles changelog.d/ fragments and moves
the CHANGELOG.md Unreleased section into a dated v5.295.0 entry.
See src/docs/releases.md.

Signed-off-by: hive-release-bot <actions@github.com>

* ✨ feat: range attribution core for post-merge runs (#18653)

* ✨ feat: range attribution core for post-merge runs

Pure, GitHub/Tekton-agnostic AttributeRuns in workflowmodel: attributes
post-merge runs per pipeline/environment to merged PRs by range prev..R,
first success wins, failure holds until a later success covers it, and
undeterminable revisions or ranges stay unknown with a recorded reason.

Part of #18356. Closes #18644.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Signed-off-by: scanner <280983584+hivecommons-hive[bot]@users.noreply.github.com>

* changelog: fragment for #18653

Generated by changelog-fragment-guard from the PR title (#5675).
Edit the wording freely, or replace this file with the no-changelog
label if the change is not user-visible.

Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
Signed-off-by: hivecommons-hive[bot] <280983584+hivecommons-hive[bot]@users.noreply.github.com>

* πŸ”– release: v5.296.0

Automated release commit. Compiles changelog.d/ fragments and moves
the CHANGELOG.md Unreleased section into a dated v5.296.0 entry.
See src/docs/releases.md.

Signed-off-by: hive-release-bot <actions@github.com>

* ✨ feat: Tekton PipelineRun post-merge observer (#18655)

* [scanner] fix: Tekton PipelineRun post-merge observer (#18646)

Add an explicit tektonRuns config on Tekton pipelines (kube context or
in-cluster, namespace, label selector, optional revision annotation/param
names) and an observer that lists PipelineRuns, reads each run's revision
(pipelinesascode.tekton.dev/sha annotation, then the revision param) and
conclusion, and feeds them through workflowmodel.AttributeRuns. A run
whose revision is not a full commit SHA keeps an empty revision with the
reason recorded, so attribution leaves the column unknown.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: scanner <280983584+hivecommons-hive[bot]@users.noreply.github.com>

* [scanner] test: cover tekton observer error paths for the coverage floor

Signed-off-by: scanner <280983584+hivecommons-hive[bot]@users.noreply.github.com>

* docs: say the Tekton observer is not wired into the stage engine yet

Signed-off-by: scanner <280983584+hivecommons-hive[bot]@users.noreply.github.com>

* ✨ feat: GitHub Actions post-merge run observer for pipeline.status (#18656)

* [scanner] fix: GitHub Actions post-merge run observer feeding pipeline.status

Observe default-branch push runs (Main CI, exact merge SHA when the
range cannot decide), release and deploy workflow runs of the workflows
an ingested model binds, attribute them to merged PRs through
workflowmodel.AttributeRuns and carry the result in
PRFacts.PostMerge so pipeline.status reports green/red/running instead
of always unknown. Never guess: pull-request and merge-queue runs never
count, a deploy run without its environment credits none of the bound
environments, and an undeterminable revision or range stays unknown
with the reason recorded.

Closes #18645

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: scanner <280983584+hivecommons-hive[bot]@users.noreply.github.com>

* docs: say the Actions observer is not wired into the sweep yet

Signed-off-by: scanner <280983584+hivecommons-hive[bot]@users.noreply.github.com>

* [scanner] fix: wire post-merge observers into pipeline.status facts (#18662)

The workflow simulate's merged-PR fill now calls the GitHub Actions
post-merge observer (ObservePostMergeRuns) with the PR's merge commit and
default branch, so pipeline.status reflects observed runs instead of
always being unknown. Refs #18356.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: scanner <280983584+hivecommons-hive[bot]@users.noreply.github.com>
Co-authored-by: clubanderson <407614+clubanderson@users.noreply.github.com>

* πŸ”– release: v5.297.0

Automated release commit. Compiles changelog.d/ fragments and moves
the CHANGELOG.md Unreleased section into a dated v5.297.0 entry.
See src/docs/releases.md.

Signed-off-by: hive-release-bot <actions@github.com>

* πŸ”– release: v5.298.0

Automated release commit. Compiles changelog.d/ fragments and moves
the CHANGELOG.md Unreleased section into a dated v5.298.0 entry.
See src/docs/releases.md.

Signed-off-by: hive-release-bot <actions@github.com>

* [scanner] fix: stop retrying closing-issue link on GitHub manual reference limit (#18668)

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Signed-off-by: scanner <280983584+hivecommons-hive[bot]@users.noreply.github.com>

* πŸ”– release: v5.298.1

Automated release commit. Compiles changelog.d/ fragments and moves
the CHANGELOG.md Unreleased section into a dated v5.298.1 entry.
See src/docs/releases.md.

Signed-off-by: hive-release-bot <actions@github.com>

* 🌱 ci: move whisper-builder cache to GHCR registry, stop wasted gha exports (#18674)

The repo's Actions cache is over the 10 GB cap, so LRU eviction kept
dropping the whisper-builder blobs and builds recompiled whisper.cpp and
FFmpeg (slow for arm64 under QEMU).

- whisper-builder warm step: cache-from/cache-to a dedicated GHCR registry
  cache (hive-buildcache:whisper-<arch>), exported only by push builds.
  Same-repo PR builds log in and read it only; fork PRs never run the job.
- hub and mx builds: drop the gha cache. Both stages of each Dockerfile are
  in no-cache-filters, so the mode=max export could never be hit.
- contributor build: export the gha cache only on push builds, so PR refs
  stop adding entries that only that PR can see.
- overlayfs-exec-guard (v2-ci and its generated copies): read whisper layers
  from the registry cache instead of the gha scope that is no longer written.

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ› Keep Copilot config.json linked to the shared login after the CLI's own save (#18673)

The Copilot CLI's atomic save replaces the per-agent config.json symlink
seconds after launch, so a later dashboard login never reached running
agents. The pane poller now relinks within one poll interval, merging
CLI-written non-credential fields into the shared file first (shared login
wins), serialized against the hive's token writers. The #18445 guard for a
private copy holding the only login is kept.

Fixes #18671

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ› ci: materialize warmed tmux/whisper stages so the main image build reuses them (#18687)

Measured on v5 after #18674: the whisper-builder warm step hit the GHCR
registry cache (all layers CACHED in about a second), but the main build
step then recompiled FFmpeg and whisper.cpp. Earlier runs show the same
for tmux and for the old gha whisper scope: a warm-step cache hit with no
output leaves nothing in the builder's local cache, and the main build was
only served when the warm step had missed and compiled locally.

Export each warmed stage to /dev/null so its layers are pulled and recorded
in the local builder cache before the main build runs.

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ“– Document kick refusals and link the docs from the dashboard (#18695)

Fixes #18692

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ”– release: v5.298.2

Automated release commit. Compiles changelog.d/ fragments and moves
the CHANGELOG.md Unreleased section into a dated v5.298.2 entry.
See src/docs/releases.md.

Signed-off-by: hive-release-bot <actions@github.com>

* πŸ› Workflow dispatch: start on a standard hive and show a profile no agent fills (#18691)

* πŸ› Workflow dispatch: fall back to the hive's issue worker or reviewer, and show a profile nothing fills

A bound workflow model hands some states to an agent profile, but the
resolver only accepted a pinned agent, an agent named after the profile,
or a writer, docs or feature agent. A standard roster has none of those,
so the drafter never started and nothing in the dashboard said why.

- config.ResolveWorkflowDispatch ends with a last resort: the issue
  worker (scanner role, classify.DefaultLane) for a profile that opens a
  pull request, the reviewer for one that reviews, in both cases only an
  agent that serves the bound repository. No other role is picked.
- GET /api/workflows lists each bound profile's agent or the reason
  there is none; the Settings -> Repos workflow row warns and offers a
  picker that saves workflow.dispatch.profiles.<profile>.agent through
  PUT /api/workflows/dispatch/profiles.
- The stage tracker carries the reason onto the item's row, and the
  manual send answers with it and where to fix it.

Fixes #18690

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ› Document the profile pin route in the OpenAPI spec; keep the gap badge rule out of the sentinel CSS block

Fixes #18690

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* 🌱 ci: let the DCO app honour an author's own remediation commit (#18703)

* 🌱 ci: let the DCO app honour an author's own remediation commit

Adds .github/dco.yml with allowRemediationCommits.individual: true so an
author can vouch for their own earlier commits whose author email had no
matching Signed-off-by, which is what blocks v5 -> v6 top-ups that carry
image-bump commits (#18697). Third-party remediation stays off.

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* changelog: fragment for #18703

Generated by changelog-fragment-guard from the PR title (#5675).
Edit the wording freely, or replace this file with the no-changelog
label if the change is not user-visible.

Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
Signed-off-by: clubanderson <407614+clubanderson@users.noreply.github.com>

* 🌱 changelog: make the dco.yml fragment a single entry bullet

The fragment is the entry itself and must start with '- '; the compiler owns
the headings. Drop the duplicate fragment the changelog bot added for #18703.

Signed-off-by: Andy Anderson <andy@clubanderson.com>

---------

Signed-off-by: Andy Anderson <andy@clubanderson.com>
Signed-off-by: clubanderson <407614+clubanderson@users.noreply.github.com>
Co-authored-by: clubanderson <407614+clubanderson@users.noreply.github.com>
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>

* ✨ Inbox: show the email-in address, delete any item, tidy the auto-routing dialog (#18704)

* ✨ Inbox: show the email-in address, delete any item, tidy the auto-routing dialog

Show the hive's email-in address (hive@<dashboard host>) at the top of the
Inbox, Calendar and Meetings sections as one shared row: label, mailto
link, Copy button and a one-line hint. Each section's state endpoint carries
it as email_in for owner and read-write callers; a hive mail cannot reach
gets a 'not set up yet' line with the docs link instead.

Make every Inbox item deletable, per item and in bulk (POST
/api/inbox/delete, owner or read-write, audited as inbox_delete). The
in-page dialog says what is removed; the Knowledge copy is kept unless the
box is ticked; a tombstone keeps synced items from coming back.

Lay the Inbox auto-routing dialog out as two sections (Routing,
Transcription), with provider rows, config keys in a footnote, consent
grouped with enable, one control recipe and height, provider-specific
fields only for the provider in use, and a footer that stays in view.

Fixes #18693
Fixes #18694
Fixes #18701

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ› Update the dashboard inbox item test for the shared delete control

Refs #18694

Signed-off-by: Andy Anderson <andy@clubanderson.com>

---------

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ”– release: v5.299.0

Automated release commit. Compiles changelog.d/ fragments and moves
the CHANGELOG.md Unreleased section into a dated v5.299.0 entry.
See src/docs/releases.md.

Signed-off-by: hive-release-bot <actions@github.com>

* ✨ Sessions: answer the CLI's startup screens so the initial request starts unattended (#18706)

* ✨ sessions: answer the CLI's startup screens so the initial request starts unattended

When a session starts (create, resume, Start again) the hive now takes the
vendor CLI to its input prompt itself and only then lets the Send queue
deliver, so an Inbox seed for spek runs with nobody at the keyboard.

- Seed a Claude Code session HOME (onboarding done, the session's own
  working directory trusted) when it has no .claude.json; never rewrite one.
- Reuse the agents' detector (panesignal.BlockingPrompts) restricted to
  startup screens: first-run notice, folder trust, theme picker, onboarding,
  the Codex update menu.
- Never answer a tool or command permission prompt, never a login or OAuth
  screen, nothing once a person has typed, and stop at the input prompt or
  after a bounded window.
- Show the state in the pane header, audit it as session_autostart, and hold
  the Send queue while the CLI is known not to be at its prompt.
- dashboard.sessions.autostart / HIVE_SESSIONS_AUTOSTART turn it off.

Fixes #18696

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ› sessions: classify startup and login screens before the busy-footer test

The busy-footer test also matches the 'Esc to cancel' hint under a trust or
login menu, so the workspace trust screen read as a working CLI.

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ› sessions: classify session_autostart as a sensitive audit action; keep the unscripted runtime fake off shared test state

Signed-off-by: Andy Anderson <andy@clubanderson.com>

---------

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* ✨ Meeting attendee: play a persona preset's introduction when its card is clicked (#18711)

Clicking a preset card in Settings, Meeting attendee selects it and plays
that persona's introduction in its own voice and language. Each card also
has a speaker button to play or stop the clip.

- GET /api/meetings/persona/preview?preset=<id> (owner or read-write)
  speaks the preset's real intro, without the pronoun note, through the
  meetings TTS service. Clips are cached in memory per TTS service,
  preset, voice and text; concurrent requests share one TTS call. Bounded
  by a per-user rate limit, a text length cap and the TTS timeout. Every
  failure is a JSON error.
- The UI shows "Voice preview is unavailable: <reason>" and never falls
  back to the browser's speech synthesis.
- cmd/hive now gives the dashboard the meetings TTS service, and the
  token is also sent as X-API-Key, which the shared Kokoro service checks.

Fixes #18710

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ› Betty: live Discord/Teams test fixes + inbox auto-transcribe (#18679)

* πŸ› Betty: live Discord/Teams test fixes and inbox auto-transcribe

Discord joins talk by default when voice-capable; speak logs why it did
not speak; tolerant wake words for misheard names, never inventing a
participant; research pool warmed for live meetings with a cold-start
settle and prompt-not-ready retry; board pressure and headline metrics in
the live hive status; batched answers attributed; unambiguous voice
toggle; consent spoken by a voice-capable bot when chat cannot carry it;
inbox recordings auto-transcribed on arrival and resumed on boot.

Fixes #18677
Fixes #18678

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ› research: fix translate returns; resync api-reference citations

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ› Betty: Meet test fixes, speech text, no pronouns, richer live status

Repeated spoken requests dropped; no research ack on an intro being said;
dig-in after a spoken ack goes to chat; text-for-speech pass on every say;
the persona never states pronouns; live hive status lists stage
breakdown, working agents and work queue.

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ§ͺ config: persona override test expects no pronouns in the intro

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ› Betty: research agent first-run notice, ack watchdog, 45s research deadline

Runtime-added claude agents get hasCompletedOnboarding and work-dir trust
seeded; the first-run security notice is recognised and dismissed, and a
kick blocked on it fails fast; end-to-end research pool test over a real
agent manager. Spoken requests are acked only when the answer is not
ready after 0.7 s; deep dives give up after 45 s.

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* 🧹 crux: keep agent/manager.go under its size ratchet

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* ⬆️ meetings: ship quorum v0.4.0

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ› Betty: real hive counts, no research for hive questions, two-question lines, misheard leave

Live status reads the derived open-items set with its repo scope and never
states an unbacked zero; hive questions (board pressure, counts, agents,
issue/PR lookup by number) never go to research; STT mishearings of the
hive vocabulary are normalised; a segment with two questions gets two
answers; 'Betty Lee.' leaves with a confirmation. The crux research e2e
test no longer imports pkg/research (crux outbound ratchet).

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* ⬆️ meetings: quorum v0.4.1; deep-dive test asks a non-hive question

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ§ͺ betty: two-question test skips the queue note

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ› betty: read 'pull request 18,689' as item 18689

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* ✨ dashboard: make the merge queue panel readable and consistent with the rest of the dashboard (#18709)

* ✨ dashboard: make the merge queue panel readable and consistent

Rebuild the Activity monitor's forge-owned merge queues panel from the
tracker's own parts (preparation rail head line, Merged/Merge queue popover
lists, ham-badge modifiers, the shared config-info help mark, design tokens)
and drop the one-off nq-* look. Overview first: one head line with how many
PRs are queued on how many branches and when the newest check happened; one
status line per branch that has something to say (front PR, its GitHub state,
how long it has waited, live test group verdict, checked Nm ago); calm
branches folded into a count; entries, test groups and receipts behind a
click. Every caveat is stated once, in the help mark.

A normal rotation delay reads as an age, not a STALE banner; warning styling
is kept for a failed refresh, missing App authentication, evidence older than
the fresh window plus two full rotations, and failed candidate checks.

Go: a candidate whose synthetic ref is gone or moved is sent with
expired: true and never read for checks, so finished groups are history, not
live validation candidates. nativeObservationRequired treats a 404 from
classic protection (GitHub's answer for a ruleset-only branch) as no classic
requirements, so the ruleset read supplies a known policy; 403 and other
errors still leave the policy unknown.

Fixes #18702

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* ✨ dashboard: pin the readable merge queue panel in the node test

Update .github/scripts/native-queue-tracker.test.cjs to the new markup while
keeping every honesty rule it guarded: GitHub's state in words with the enqueue
age, a failed refresh standing out without an empty-queue claim, the caveat
stated once, an overdue check as a problem but a rotation delay as an age,
unproven members and unreadable policy never shown as passing, finished groups
as a count, and receipts behind the click. Calm branches keep their history
behind the same toggle.

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ› dashboard: never say "nothing queued" for a branch whose refresh failed

The head line counted a branch as read whenever it had been visited, so a
single branch with a failed refresh produced "nothing queued on 1 branch". Only
observed and not-enabled scopes count as read now; the node test's regexes
check the status line rather than the help text and use the 12-character SHA.

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* 🌱 dashboard: count the help mark once in the native queue render test

The caveat text sits in both the aria-label and the tooltip span of the single
help mark, so the sentence occurs twice; assert one help mark and two copies.

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* 🌱 dashboard: assert the old per-queue caveat sentence is gone, not its topic

The help mark legitimately says the repo filter applies here; the render test
now checks for the removed per-queue sentence's own opening words.

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* 🌱 dashboard: a fresh problem-free queue renders no badge; assert that

The component-reuse check expected a ham-badge on the fresh fixture, but a
calm queue correctly shows none; badge cases are asserted on the stale and
failed fixtures.

Signed-off-by: Andy Anderson <andy@clubanderson.com>

Co-authored-by: Danathar <6772335+Danathar@users.noreply.github.com>

* πŸ”– release: v5.300.0

Automated release commit. Compiles changelog.d/ fragments and moves
the CHANGELOG.md Unreleased section into a dated v5.300.0 entry.
See src/docs/releases.md.

Signed-off-by: hive-release-bot <actions@github.com>

* πŸ› scanner: not-agent-work verdict parks the issue and is not repeated (#18686)

* πŸ› scanner: not-agent-work verdict parks the issue and is not repeated

The scanner's not-agent-work triage verdict (not a defect, needs a human, or
needs an environment agents lack) was only a comment: the issue stayed
stage/010-actionable and was re-triaged every pass. The comment relay now
takes a structured verdict (--not-agent-work, or the hive:not-agent-work
marker), posts it once, and parks the issue on needs-human. An identical
verdict with no human comment since is dropped as already_existed, and a
person's later comment re-queues the issue via the un-park sweep.

Fixes #18675

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ› scanner: register the not-agent-work note write site and reuse the label wrapper

TestNotificationWriteSitesInventory wants every comment producer tagged with
a class, and TestDirectWriteRatchet refuses a raw RemoveLabelForIssue in a
file outside its baseline. Tag relayNotAgentWorkNote as attention (and list
it in docs/notifications.md), and lift needs-human through removeIssueLabel.

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ› scanner: add the not-agent-work park through the relay's existing label write

The direct-write ratchet refuses a wrapper write in a file outside its
baseline. The comment relay already adds one park label after a note
(blocked for --blocked-on-pr); the not-agent-work verdict now reuses that
write for needs-human and only submits the Parked fact itself.

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ› Mail-in: accept office documents, let large deliveries reach the hive, and show why a mail was refused (#18708)

* πŸ› Mail-in: accept office documents, let large deliveries through, and make refusals visible

A mail to hive@<host> with only a .pptx attached was refused with a 550
that named no cause, and the receiver logged one line without the
recipient, the sender's domain, the size or the reason.

- hive-mx accepts .pdf, .docx, .pptx and .xlsx as document attachments,
  checked by content (an Office container must hold its content types and
  main part, and no macro project).
- The spoke's session gate lets the hub's signed pieces through
  (/api/inbox/email/chunk): any mail with more than one relay piece of
  attachments was answered 401 and retried until it expired.
- Refusals say why in the SMTP reply, per attachment extension.
- One INFO line per SMTP session: outcome, code, reason category,
  recipient host, sender domain, size, attachment counts. No content.
- hive-mx reports refused messages to the hive (report delivery through
  the hub); the hive keeps the last 50 accepted and refused mails and
  shows them in the Inbox (GET /api/inbox/email/recent).

Fixes #18699

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ“– OpenAPI: document the recent-mail and refusal-report routes

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ› Recent mail: reload on open and on tab focus instead of a timer

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* ✨ Recent mail: say why a received mail awaits approval, with the authentication behind it

hive-mx logs the SPF and DKIM results, the alignment and the HELO name
next to the DMARC result, and the hive's recent-mail log keeps them with
the approval reason (sender not authenticated, or authenticated but not
on the allowlist), so a DMARC failure can be traced to the server the
mail was sent through.

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* 🎨 Dashboard: render export/auth URLs as a read-only line, not a textarea (#18733)

The Claude authorization URL dialog (hiveExportViewer) showed the OAuth
URL in a small resizable textarea that matched none of the dialog's
tokens and cut the URL off. Links are now one full-width, read-only,
selectable line in the dialog's own font and colours (word-break:
break-all, click selects all); multi-line exports use a read-only pre.
Open in new tab is primary, Copy secondary, Close a ghost button, each
action carries an aria-label, and the implementation note is replaced by
a short user-facing line (the bearer-token variant keeps its meaning).
The hub's access CSV viewer gets the same treatment.

Fixes #18729

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ”– release: v5.301.0

Automated release commit. Compiles changelog.d/ fragments and moves
the CHANGELOG.md Unreleased section into a dated v5.301.0 entry.
See src/docs/releases.md.

Signed-off-by: hive-release-bot <actions@github.com>

* πŸ“– Customers: design doc for the people store, customer graph, goals and satisfaction (#18743)

Design record for RFC #18742: the Customers section, the customer versus
suggested line, identity resolution, retention by reference, forget with
tombstones, the role rules, the read-only GraphQL endpoint, the bounded agent
context block, and goals and satisfaction inferred with evidence and
confirmed by a person.

Refs #18742

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ› Sessions: keep one refreshable Claude login (a session can no longer lock the fleet out), and fix the false "hive restarted" start failure (#18728)

* πŸ› Sessions: judge an orphaned preparing record under the lock, not from the list snapshot

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ› Sessions: keep ONE refreshable Claude login; a session can no longer rotate it away from the fleet

A session got a full copy of the shared Claude login. The CLI in the session
refreshed the copy, which rotated the grant and left the shared file every
agent reads stale; the fleet was locked out until a person signed in again.

- a session's Claude login copy carries no refresh token (and only the token
  block), so a session is never a refresh holder
- the hive keeps session copies current with the shared login
- a complete, usable, strictly newer login found in a session HOME is adopted
  into the shared login under the credential-bridge lock, with an audit entry:
  at boot, on the bridge poll, after a sign-in, at session start and before a
  session's sandbox is removed (removal fails rather than lose it)
- other vendors: a resume no longer overwrites a login the session's CLI
  rewrote after the shared file was written

Fixes #18724
Fixes #18725

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* 🌱 Sessions: keep the reconcile test seam out of server.go and manager.go within its size baseline

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ“– Sessions API reference: refresh file:line citations

Signed-off-by: Andy Anderson <andy@clubanderson.com>

---------

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* ✨ Customers: people store, identity resolution, tags, goals and satisfaction, edge index (#18751)

* ✨ Customers: people store, identity resolution, tags, goals and satisfaction, edge index

First build step of RFC #18742. Adds pkg/customers (nothing reads it yet):

- people.json (0600, atomic) with one profile per person: identities, tags,
  goals, a satisfaction level, notes, merge history;
- suggestions as a key, counters and at most five evidence references;
- identity resolution on exact keys only; names never link; manual decisions
  are pinned; merge, unmerge and split are exact;
- forget with salted-hash tombstones that block rediscovery;
- goals and satisfaction with auto/manual provenance: inference only writes
  suggestions and never overwrites a confirmed or manual value;
- a rebuildable bbolt edge index that stores references and prunes edges
  whose artefact is gone;
- customers.enabled (default on, env HIVE_CUSTOMERS_ENABLED).

Refs #18742

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ“– Customers design: dismissed suggestions become tombstones; hive_user identities are not hand-editable

Refs #18742

Signed-off-by: Andy Anderson <andy@clubanderson.com>

---------

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* ✨ Mail-in: a refused or held mail says how to try again and where to report a problem (#18750)

A sender whose mail was refused got an SMTP error with the reason and
nothing about what to do next, and the Inbox's Recent mail list had no
way to act on a refused or held mail.

- hive-mx ends each refusal that is about the message or the address
  with three lines: how to try again, a link to the reason's entry in
  the docs, and where to report a problem. The sender's own provider
  returns them in its bounce. The lines are the same for every sender.
- GET /api/inbox/email/recent adds how_to_fix, help_url and report_url
  to each refused or held event, and the items of a held mail that
  still await approval.
- The Recent mail list shows them under each such mail, with Approve
  and Decline for the owner on a mail that awaits approval.
- Docs: "Mailing your hive: what the replies mean", one entry per
  reason.

Hive still sends no mail: there is no outbound path and no sending
domain. The reply email is designed in #18746.

Refs #18744

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* ✨ feat: acme-dns RFC 2136 manifests for hub wildcard cert (#18665)

* feat: add delegated RFC 2136 acme-dns manifests for hub wildcard cert

Signed-off-by: scanner <280983584+hivecommons-hive[bot]@users.noreply.github.com>
Co-authored-by: clubanderson <407614+clubanderson@users.noreply.github.com>

* fix: enable Knot response rate limiting on acme-dns

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* docs: acme-dns runbook order, TSIG key copies, restart and bind checks

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* docs: note TSIG secret lives in two namespaces

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* fix: let acme-dns knotd bind port 53 as non-root, note image sync

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* docs: keep TSIG command only in the acme-dns README

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* docs: clarify acme-dns changelog scope

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* docs: acme-dns README reflects the port 53 sysctl

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* fix: load acme-dns zone file whole, no journal to diff against

Signed-off-by: Andy Anderson <andy@clubanderson.com>

---------

Signed-off-by: scanner <280983584+hivecommons-hive[bot]@users.noreply.github.com>
Signed-off-by: Andy Anderson <andy@clubanderson.com>
Co-authored-by: scanner <280983584+hivecommons-hive[bot]@users.noreply.github.com>
Co-authored-by: clubanderson <407614+clubanderson@users.noreply.github.com>
Co-authored-by: Andy Anderson <andy@clubanderson.com>

* πŸ”– release: v5.301.1

Automated release commit. Compiles changelog.d/ fragments and moves
the CHANGELOG.md Unreleased section into a dated v5.301.1 entry.
See src/docs/releases.md.

Signed-off-by: hive-release-bot <actions@github.com>

* ✨ Inbox: Quick note, larger hit targets and an honest list of input methods (#18752)

* ✨ Inbox: Quick note, larger hit targets and a list of input methods

Add a Quick note box to the Inbox Add panel: typed or dictated text is
sent as a markdown file through the existing upload route, so the size
cap, retention, extraction, attribution and auto-routing are unchanged.
A multi-line paste onto the Inbox lands in the box for review.

Raise the Inbox's small hit targets (item checkbox, small buttons, pills,
+N / less and text links) to --tap-target through a transparent
pseudo-element, without changing their visible size.

List input methods under the source chips from one static catalogue
(pkg/inbox/input_methods.go, served as input_methods by GET /api/inbox):
tools that work today because they enter text, and neural interfaces
that publish no API or export yet. The chips are not connectors; they
only show an explanation that states hive is not affiliated with the
projects named.

Fixes #18747

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ› Inbox hit targets: scope the button rule by the Inbox regions

The stylesheet spelled the data-dashboard-section attribute, which the
dashboard-notices ordering test searches for in the page.

Signed-off-by: Andy Anderson <andy@clubanderson.com>

---------

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* ✨ Customers: REST API, wiring, "me" for every role, forget (#18756)

PR 2 of RFC #18742.

- pkg/customers: a viewer-aware Service. Without the full view a viewer
  reaches only their own profile by any method; an id they may not see
  answers exactly as a missing one. An identity a read-only user adds is an
  unconfirmed claim: not a key, resolves nobody, attracts no edges.
- pkg/dashboard: seven /api/customers routes behind an injected service
  (no import of pkg/customers). The read role's one write is
  POST /api/customers/me/edit, exempted by exact method and path.
- cmd/hive: the adapter and boot wiring, behind customers.enabled.
- openapi.json, api-reference page, changelog fragment, design doc.

Refs #18742

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* fix: Governor PRs-by-model honesty (no 0% for unmeasured runs, coverage, CIs) and codex model detection in the contributor relay (#18739)

Refs #18713, #18714

Signed-off-by: scanner <280983584+hivecommons-hive[bot]@users.noreply.github.com>
Co-authored-by: scanner <280983584+hivecommons-hive[bot]@users.noreply.github.com>
Co-authored-by: Danathar <6772335+Danathar@users.noreply.github.com>
Co-authored-by: Claude Sonnet 5.5 <noreply@anthropic.com>

* πŸ› Cross-repo decomposition children and cross-repo PR issue links (#18688)

Scanner decomposition filed every child in the tracker's repo with a bare
"Part of #N" / hive:parent N link, and a PR in another managed repo that
said "Refs owner/repo#N" never covered its issue: the claim was keyed
owner/repo#N while enumeration keys issues by the configured spelling.

- hive-open-issue --parent owner/repo#N; the relay files the child in the
  managed repo where its work lives (falls back to the parent's repo when
  that repo is unmanaged; refuses an unmanaged parent repo), links it as a
  cross-repo sub-issue, rewrites the parent ref and marker to full form and
  records the parent repo in the split-parent ledger.
- Tracker progress and decomposed-parent close count children in any repo.
- Claims are keyed by the configured repo spelling, so a cross-repo PR
  covers its issue and the issue mirrors the PR stage; the stage-pr-link
  comment is posted for a PR in another repo; an agent claim is renewed
  while a hive PR covers the issue and released once it closes.

Fixes #18676
Fixes #18680

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* ✨ Inbox UX cleanup: icon Play/Share buttons, one-row toolbar, clean titles and previews (#18705)

* 🎨 Inbox: icon Play/Share buttons, compact toolbar, clean titles and previews

Fixes #18700

- Play recording and Share are icon buttons with aria-label and title;
  Transcript / Text is a compact chevron toggle; a hive recording's Delete
  is a trash icon set apart at the end of the row (the slot #18694 uses).
- Only configured sources are chips; the rest sit behind a +N more
  disclosure next to Connect sources.
- One toolbar row: Record and Add left; Show processed (toggle button),
  Select all, Hive menu and Auto-routing right.
- Trailing (mtg-…) ids leave the title for the meta line; the full title
  is the tooltip. The Knowledge state sits in the meta line, not as a pill.
- Upload previews render Markdown as plain text, clamped to 3 lines with
  Show more.

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* 🎨 Inbox tests: stub inboxSourcesHtml in the role-race harness, compare title parts by value

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* 🎨 Inbox: Delete slot takes inboxDeleteButtonHtml when #18694 registers it

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* 🎨 Inbox: inline the Play and Share icon buttons so the row guard sees their actions

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* 🎨 Inbox tests: the every-kind Delete row test expects the trash icon button

Signed-off-by: Andy Anderson <andy@clubanderson.com>

---------

Signed-off-by: Andy Anderson <andy@clubanderson.com>

* πŸ› claude: hive refreshes the shared OAuth grant itself ahead of expiry (#18736)

* πŸ› claude: hive refreshes the shared OAuth gran…
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

approved Indicates a PR has been approved by an approver from all required OWNERS files. dco-signoff: yes Indicates the PR's author has signed the DCO. size/S Denotes a PR that changes 10-29 lines, ignoring generated files.

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant