mirror of
https://github.com/rennf93/roboco.git
synced 2026-08-03 07:23:24 +02:00
master
568
Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
 Renzo FandGitHub
|
666f261a1a | fix(kimi): spawn's own prepared instance no longer counts against the concurrency cap (#714) | ||
|
Renn F
|
95e7d5df7c |
fix(kimi): cap concurrent Kimi agents to protect the shared auth chain
Every Kimi container redeems the same rotating refresh-token chain; Moonshot rotates with a short reuse grace, so two containers refreshing near-simultaneously fork the chain and a later stale redemption revokes the whole family - fleet-wide re-login (observed twice in production, each after paired spawns). With one consumer at a time refreshes are strictly sequential and the chain stays coherent, so the spawn gate now skips-and-retries Kimi spawns past ROBOCO_KIMI_MAX_CONCURRENT (default 1), sharing the provider-parked bail path. The compose files also gain the four Kimi tunables their environment blocks silently dropped - documented .env overrides never reached the orchestrator container. |
||
|
Renn F
|
1c3c313c1e |
fix(coroner): playbook-kind postmortems stop rendering dead buttons
A kind=playbook process change drafts into the playbook queue at propose time and both approve/reject refuse it - but legacy rows carry no marker status, so the response defaulted to proposed and the panel rendered approve/dismiss buttons that bounce forever. The list route now derives not_applicable for playbook-kind changes regardless of stored status, which the panel already renders as its Drafted-as-playbook badge. |
||
|
|
6374bbbed0 |
feat(kimi): Kimi K3 provider on the official kimi-code CLI (#713)
* feat(kimi): Kimi K3 provider on the official kimi-code CLI (Wave 1) ModelProvider.KIMI routes through KimiCliProvider driving Moonshot's kimi CLI on a Kimi subscription (OAuth device-code, no metered key). One-shot delivery roles only (V1), interactive ban wired in both guard lists. Auth: one shared RW auth mount; containers symlink credentials/ and oauth/ (the CLI's cross-process refresh-lock dir) into a container-local KIMI_CODE_HOME so every container and the host redeem the SAME rotating refresh chain - live-verified that per-copy chains cross-invalidate after the reuse-grace window. No orchestrator refresh daemon; an expires_at preflight exits 78. Config renderer mirrors the login-managed provider/model blocks field-for-field (live-captured; the model value is the CLI-side name, never the raw API id), plus per-role deny rules and the bash-guard as a PreToolUse hook via a wrapper script (an env key on a hooks entry makes the CLI silently drop ALL hooks - live-verified). Usage capture sums wire.jsonl usage.record 4-bucket events; sniff classifies rate-limit/auth from structured error text only, mapped to the shared 75/78 park contract. Image installs the CLI latest-at-build (no version pin, by policy) with the resolved version stamped as provenance, binary split to /usr/local away from mutable state. Migrations 090 (enum) + 091 (provider seed); catalog, pricing, routing mode, and orchestrator park/usage wiring mirror the codex integration. * feat(kimi): surface sweep + fleet-wide pin drop (Wave 2) Compose x3 gain the agent-kimi-image service and the orchestrator's read-write ~/.kimi-code mount + kimi-usage dir; .env.example documents the Kimi block. Panel mirrors ModelProvider.KIMI and adds the kimi routing mode (catalog filter, mode button, mix-picker group, badge) with tests; provider routes gain the kimi remediation entry. CLAUDE.md and docs/map document the runtime. Per the no-pins policy, agent-grok/ gemini/codex Dockerfiles drop their version pins for latest-at-build with resolved-version provenance stamps (grok resolves 0.2.112 vs the old 0.2.56 pin - verified by real builds of all four images). --------- Co-authored-by: Renn F <rennf93@users.noreply.github.com> |
||
|
Renn F
|
eb470dfb33 |
fix(x): Barfly drafts become standalone link-posts
X 403s programmatic replies into conversations that don't mention the account (every non-Enterprise tier), which is Barfly's entire discovery surface - so its drafts now carry commentary plus the conversation's /i/web/status/ URL instead of a reply target. Handles are stripped (mentions in plain posts are rejected too), the 280 budget accounts for the URL, the exploration prompt asks for standalone commentary, and the reject->redraft path re-appends a dropped link. |
||
|
Renn F
|
98ee4d79e5 |
fix(x): thread mention replies and audit every post outcome
x_reply drafts (the one case X's 2026-02-23 reply policy allows - the author summoned us) now post as real threaded replies via the carried mention id; barfly drafts stop sending in_reply_to_tweet_id entirely. Failed posts log a server-side warning and write an x_post.post_failed audit row; successes write x_post.posted - at the _post chokepoint so both approve routes are covered. |
||
|
Renn F
|
48feb033d6 |
fix(api): carry orchestration_markers in the task response
TaskResponse declares the field, but task_to_response never assigned it,
so every GET /tasks response served null. The orchestrator fetches its
work over that endpoint and builds prompts from the returned dict, so
all twelve marker reads in orchestrator.py resolved to `{}`.
Live symptom: a Barfly explorer was told "SCREENED CANDIDATES: (none)"
and correctly refused to invent a tweet — while BarflyEngine only opens
an exploration task when it HAS candidates, so the ones it gathered were
sitting on the row the whole time. The gateway path reads markers off the
ORM object directly, which is why propose_conversation_replies could see
what the prompt could not.
Same blindness applied to every other marker consumer on the dispatch
path: the spotlight's seen-features dedup, pest control's evidence
snapshot, coroner's autopsy subject, and the dev-redispatch's
original_developer lookup.
|
||
|
|
42f8a5d18e |
fix(board): nothing_to_propose exit for Board Program explorers + PR checks on slave-based PRs (#712)
* fix(board): give Board Program explorers a nothing_to_propose exit Every propose_* verb requires at least one item, so an explorer that legitimately found nothing — Barfly with no worthwhile X conversations, Coroner with no autopsy subject — had no way to close its exploration task. It declined, called i_am_idle(), and the task stayed PENDING forever: the dispatcher re-matched it every tick and respawned the board agent (~$0.61 a spawn, ~3 per 5-minute respawn-breaker cooldown window, indefinitely), and BoardProgramEngine's one-open-cycle dedup wedged that whole program shut, since the ledger row only closes once its exploration task goes terminal. nothing_to_propose(task_id, reason) is the explicit exit. task_id is required rather than inferred: one explorer role owns several independently-cadenced programs (head_marketing owns six) and each assigns its exploration task to the same agent, so several are open at once by design and guessing "the caller's oldest" completes the WRONG cycle — stamping its reason onto an unrelated program's ledger while the task actually being worked stays wedged. Resolution validates the named task exists, carries a registered program source, is assigned to the caller, and is non-terminal, then gates on the program's declared explorer role from the registry, so a program registered later needs no edit here. The reason lands on board_program_cycles (migration 089) and renders into the next cycle's LEARN context, replacing a bare "proposed 0, approved 0" with why. That write runs in its own savepoint: it flushes on the same session as the completion, and a bare try/except around a same-session flush leaves the transaction pending-rollback, so a DB blip there would discard the completion at the post-response commit while the verb reported success. All fourteen exploration prompts offer the exit, pinned by a registry-parametrized test that fails when a future program is unwired. * ci: fire PR checks on slave-based PRs, not master alone All five gating workflows declared `pull_request: branches: [master]`, but every fleet PR targets slave — cell->root, root->slave, and the CEO's own. So `pull_request` never fired for any of them, and their only coverage was the `push` trigger, which is gated on branch PREFIX (feature/bug/chore/docs/hotfix). A branch named anything else got zero checks — not a red run, an absent one — and a PR with no required check present merges on a false green. PR #711 shipped that way on a `fix/` branch. Basing on the branch a PR merges INTO rather than what its head is named makes coverage independent of branch naming, so a non-conforming prefix can only ever cost the redundant push run, never the whole gate. The same five also omitted slave from `push` (ci.yml aside, which added it for the release gate's fail-closed CI read), so the panel suite, both CodeQL analyses, and the e2e smoke never ran on the trunk master is cut from. --------- Co-authored-by: Renn F <rennf93@users.noreply.github.com> |
||
|
|
a7b970a3b2 |
feat(board): materialize program items as Main-PM roots, make reports actionable (#711)
Two coupled gaps in the Board Program output path. Approved items were created unowned and in BACKLOG. Nothing dispatches BACKLOG, and once activated a cell PM claimed the parentless task as a root, where _cell_pm_complete resolves its merge target through resolve_parent_branch — which for a parentless task falls through to the project head rung. The result was a cell branch merging straight into the trunk, bypassing the Main-PM root, the root->master PR and the CEO gate (live: PRs #703 and #704 both targeted slave directly). All eight materializers now create a PENDING, main-pm-assigned root with team=Team.MAIN_PM, matching what approve_and_start does for an intake draft. The team is load-bearing, not cosmetic: _next_hint_pr_fail, _deliver_pr_fail_to_owner, delegate's wave-chain dispatch and the PR layer label all key on it, and a cell-teamed root drops the 'do NOT re-submit the root' steer that exists because of PR #138's infinite pr_fail loop. The item's own cell survives as a delegation hint in the description, which is what the Main PM's briefing renders. Periscope, Sentinel and Coroner produced artifacts with no way to act on them — three panel surfaces carried explicit 'no approve/reject UI' comments while each item already held a machine-readable suggested action. They now have per-item approve and dismiss, modelled on the roadmap queue: idempotent per item, CEO-gated, deep-copy-before-mutate so SQLAlchemy's dirty check still fires, and every decision recorded through record_decision so it reaches the next cycle's prompt. Approving materializes through the same corrected Main-PM-owned path. Target project resolves to each engine's own existing anchor — RoboCo's project for Periscope and Sentinel, the incident's project for Coroner — and fails with a clean invalid_state naming what is unresolvable rather than guessing at a repo. Co-authored-by: Renn F <rennf93@users.noreply.github.com> |
||
|
roboco-app[bot]GitHubroboco-app[bot] <302741806+roboco-app[bot]@users.noreply.github.com>Backend Developer 1Backend Documenterroboco-app[bot] <302741806+roboco-app[bot]@users.noreply.github.com>Backend PM |
66f0287d11 |
[3e7a5064] Eval harness: wire the real-spawn path so python -m roboco.eval run works end-to-end (#703)
* [431e73b7] Wire the real-spawn path: OrchestratorStageSpawner + disposable MCP config (#701)
* [431e73b7] Wire the eval harness real-spawn path: OrchestratorStageSpawner + disposable MCP config
_generate_mcp_config now prefers settings.api_url when set (both
PROJECT_HOST_PATH branches), so spawned MCP servers resolve to the
harness's disposable orchestrator URL instead of the real production
hostname or 127.0.0.1:port. OrchestratorStageSpawner.__init__ replaces
the NotImplementedError with a real AgentOrchestrator() constructed the
same way the production dispatcher builds it. The runner module
docstring + __main__.py docstring/run-subparser help drop the
NOT-YET-FUNCTIONAL wording. A new unit test pins the no-production-reach
guarantee: with settings.api_url patched, the MCP config's
ROBOCO_API_URL/ROBOCO_ORCHESTRATOR_URL point at the disposable URL (not
production), and the agent UUID is the real fixed UUID from
foundation.identity.AGENTS.
* [431e73b7] docs(eval): reflect the wired real-spawn path in tests map, CLAUDE.md, and CHANGELOG
---------
Co-authored-by: Backend Developer 1 <be-dev-1@roboco.tech>
Co-authored-by: Backend Documenter <be-doc@roboco.tech>
* [5cc75f71] Fix disposable orchestrator container-reachability + document real-UUID isolation design (#705)
* [5cc75f71] Fix disposable orchestrator container-reachability + document real-UUID isolation
* [5cc75f71] docs(eval-harness): document container-reachability fix + real-UUID isolation in map docs
---------
Co-authored-by: Backend Developer 1 <be-dev-1@roboco.tech>
Co-authored-by: Backend Documenter <be-doc@roboco.tech>
* [d96ec059] Rewrite stale eval-spawner pinning test to assert wired behavior (test-only, CI-green for PR #703) (#707) (#708)
* [d96ec059] test(eval): assert OrchestratorStageSpawner constructs a real AgentOrchestrator
Rewrite the stale pinning test that asserted the PRE-wiring
NotImplementedError (removed by commit
|
||
|
|
80ccf415cb |
feat(cockpit): expose first_pass_yield and a real escaped-defects metric (#709)
The Company Scorecard renders three charter objectives but the cockpit summary only ever carried one of the metrics, so two cards read "No data yet" permanently. first_pass_yield is a pass-through — MetricsService.get_org_scorecard() already computes it on the same 30d/org scope the rest of the delivery block uses, and CockpitService.summary simply never forwarded it. escaped_defects is new. The obvious definition — a blocker finding opened on a task that already reached a terminal state — is unimplementable: every producer of a task_review_findings row fires as part of a bounce whose transition requires a non-terminal task, so it would read zero forever, and a permanently-green card is the same fabrication the panel change removes. What it counts instead: a blocker still at 'addressed', never 'verified', on a task that has since completed. That is reachable because stamp_addressed_verified only bulk-verifies rows matching its OWN origin, so a blocker raised by one origin and never re-confirmed by that origin survives to completion on the developer's word alone. docs/map/metrics-observability.md documents what a zero actually means: the one reachable trigger is a PM-origin blocker on a task escalated to the CEO rather than completed by the PM, since escalate_to_ceo carries no findings-resolved precondition and ceo_approve verifies only ceo-origin rows. It also records that the count is per-finding over a rolling 30-day window, which is not the same unit as the charter's "per release". Co-authored-by: Renn F <rennf93@users.noreply.github.com> |
||
|
|
a36a180b73 |
fix(orchestrator): let a failed Board Program exploration retry (#706)
`_board_dispatched` is an in-memory, never-expiring set of (agent_slug, task_id). The solo exploration dispatchers consulted it, so an exploration whose propose verb rejected was never retried for the life of the process — Periscope, Sentinel, Scales and Barfly each spawned once on 2026-07-25, failed, and sat PENDING until the stack restarted. The guard was written for the two-reviewer board REVIEW pass, where it is correct: a reviewer has no verb to advance the task, so a respawn can only loop. An explorer is the opposite — `propose_*` is exactly such a verb, so a respawn can and should advance it. Drop it from the 15 `_dispatch_*_exploration` functions. Bounding falls to `_pm_respawn_should_gate`, which is what actually bounds a loop: DB-persisted, reset by a status change, and cooled down so a deploy that fixes the cause lets the work resume. `_dispatch_board_reviewer` keeps the set (its `_board_review_complete` reads it as a has-run signal) as does vault curation (same-process race guard behind its own durable marker). The 14 `*_dispatch_is_one_shot` tests asserted the old contract on the false rationale that board roles have no progression verb; they now assert a second tick re-attempts. Co-authored-by: Renn F <rennf93@users.noreply.github.com> |
||
|
|
879afc14a4 |
Board Program LEARN context, ruff 0.16, and verb-rejection observability (#700)
* fix(board): LEARN decisions name the item, not its per-cycle index A cycle's reject reasons are rendered into the NEXT cycle's exploration prompt, but the ref recorded alongside each reason was the item's stored id (item-0/item-1) — a per-cycle index that means something different every cycle and appears nowhere the explorer can resolve. The reason survived the loop; what it was about did not. Record the item's title instead, via a shared learn_ref() helper (falls back to the id when title-less, and reads target_task_title for Scales, whose items name the live task they mutate). * chore(lint): satisfy ruff 0.16 — keyword-only signatures and markdown formatting The dev toolchain resolved ruff 0.16.0, which stabilises PLR0917 (too many positional arguments) and formats python code blocks inside markdown. Both fired repo-wide and neither had anything to do with the code they flagged. - 36 signatures gain a `*` so their tail arguments are keyword-only, and the 104 call sites that passed them positionally are converted. mypy was the safety net for the static ones; the full suite caught nine more that only bind at runtime (the MCP tool functions, whose real callers already pass named JSON arguments). - 28 markdown files reformatted by 0.16's code-block formatter. - One RUF036 (`None` mid-union) autofixed in the GitLab provider. * fix(gateway): log the reason when a verb rejects A rejected envelope rides an HTTP 200, its body is never logged, and there is no trace table — so in the access log a verb an agent could not satisfy looks identical to one that worked. On 2026-07-25 four Board Programs (Periscope, Sentinel, Scales, Barfly) each POSTed their propose verb three or four times, persisted nothing, and left their exploration tasks PENDING; the reason was unrecoverable afterwards, from the logs or from the agents' own transcripts. Log error/message/remediate/missing plus the calling agent at envelope_to_response — the one chokepoint every v1 flow and do route returns through. Success envelopes stay silent. --------- Co-authored-by: Renn F <rennf93@users.noreply.github.com> |
||
|
|
401f8a2cc9 |
feat(board): Board Programs — the complete twelve-program catalog (Phases 1-3) (#699)
* feat(board): Pest Control — the first project-scoped Board Program The Product Owner hunts latent defects (what the org records but nobody reads): a weekly cycle — accelerated off-schedule when the trailing-7-day rework rate crosses pest_rework_threshold, with the cheap dedup/scope gates evaluated before the metrics queries — opens one held exploration task against the least-recently-explored opted-in project (deterministic round-robin; opted_in_projects gains a stable ORDER BY), with server- assembled evidence in the spawn prompt (rework hotspots, recurring-findings and waived-minor ledger aggregates, all capped) plus prior-cycle LEARN context. The PO calls the new PO-only propose_bug_hunt verb once: ≤5 items, evidence required per item, targets validated against pest_control participation. CEO decides per item — approve materializes a BACKLOG task (source pest_control, never auto-starts), reject records the reason; both feed the LEARN ledger by exploration task id; all-terminal completes the cycle. Telegram queue pushes carry working Approve/Reject handlers mirroring the roadmap kind. Doctrine: board.md Pest Control section + product-owner verb entry + regenerated verb tables. * feat(panel): Pest Control review queue Command Center gains the pest review queue (per-item approve/reject with reason, mirroring the roadmap queue); the Programs card and the project settings participates-in checkboxes pick the new program up registry-driven — the settings section renders for the first time now that a project-scoped program exists. * feat(board): Periscope — HoM market-research brief program Weekly org-scoped cycle: a solo HoM spawn researches the market (web research with mandatory source URLs — uncited findings are rejected) and files one structured brief via the new HoM-only propose_market_brief verb: headline, cited findings, threats/opportunities, positioning note, all soup-checked and screened through the injection guard at persist time (web-derived text later reaches prompts; flags recorded, content never dropped). A brief is a report, not a proposal: the verb completes the exploration in the same call (the x_feature asymmetry), the cycle ledger auto-closes, and the CEO gets a best-effort notification with no approve/reject surface (periscope deliberately never joins Telegram's action kinds). The latest brief is injected into the roadmap exploration prompt — Periscope feeds Printer, the first cross-role program input. * feat(panel): Market Briefs tab (read-only) Business page gains a Market Briefs tab listing Periscope briefs — headline, cited findings, threats/opportunities — read-only by design; a report has nothing to approve. * feat(board): Coroner — event-triggered Auditor postmortems The first EVENT program: no cron — three best-effort hooks open an autopsy when a task bounces to its 3rd revision (the audit chokepoint), is cancelled after work started, or is budget-blocked; all gated on arming + one-open-autopsy dedup, none can fail the underlying transition. A solo Auditor spawn reads the incident (server-assembled findings + transition context) and files one propose_postmortem: incident summary, root cause, failed stage (validated against the real status vocabulary), and ONE process change — a playbook-kind change drafts via PlaybookService directly into the normal pending-curation queue; the briefed draft_playbook manifest grant was deliberately NOT added, preserving the existing 'auditor curates but never drafts' invariant test. Complete-at-propose (report asymmetry), cycle ledger auto-closes, CEO notified link-only. Integrated as a union with Periscope across the shared program surfaces. * feat(panel): Coroner postmortems card Read-only postmortems list under Business → Programs — incident, root cause, failed stage, process change; nothing to approve, the process-change artifact (a draft playbook) rides the existing curation queue. * feat(board): Sentinel — Auditor drift-watch quality reports Weekly org-scoped cycle: a solo Auditor spawn receives a server-assembled drift context (waived-findings trend, open findings by severity, conventions-violation hotspots, top spend — all capped, pure ORM) and files one propose_quality_report: headline, 1-7 area-validated items with evidence and suggested actions, overall assessment. Report semantics — complete-at-propose, cycle auto-closes, CEO notified display-only (never on Telegram's approve/reject surface); items are structured so a later convert-to-task control is cheap. Integration adopts Sentinel's module- level dict-dispatch for board-program routing (xenon-driven), folding all prior programs in; app router mounting extracted to a helper for the same budget. * feat(panel): Quality Reports tab (read-only) Business page gains the Sentinel quality-reports tab — headline, per-area observations with evidence and suggested actions; read-only, a report has nothing to approve. * feat(board): Spackle — gap-fill audit program Biweekly project-scoped PO cycle over the half-shipped surface area: API routes without panel surfaces (and vice versa), armed flags without docs, docs promises the code doesn't keep, dead-end tabs — the inventory diffing is the PO's own read-tool work, ordered by the spawn prompt with file:line citations required; the server injects only prior-cycle LEARN and the rotation target. Rotation is now a shared module-level helper (pick_rotation_target, parameterized by source) both project-scoped engines use — pest_control delegates to it, behavior-identical, with a cross-pollution test proving the two programs' rotations stay independent. propose_gap_fill mirrors the bug-hunt verb (≤5 items, two-sided evidence required, participation gate); per-item CEO decide materializes BACKLOG source=spackle tasks; full Telegram kind incl. approve/reject handlers. All seven program routers now mount from one helper. * feat(panel): Spackle gap-fill review queue Command Center gains the gap-fill queue mirroring the pest-control one — per-item approve/reject with the two-sided gap evidence rendered. * feat(board): Scales — monthly portfolio rebalance Org-scoped PO cycle over the stale backlog: the spawn receives a capped stale-task snapshot (BACKLOG/PENDING unclaimed >30 days) plus the charter and prior-cycle LEARN, and files one propose_rebalance — 1-7 items, each a resolvable task_ref with action reprioritize (validated new priority) or cancel, rationale required. Per-item CEO decide: approve EXECUTES the action (audited priority update, or the normal cancel path) — the first program whose materializer mutates existing tasks instead of creating them; reject records the reason; LEARN by exploration task id; all-terminal completes the cycle. Full Telegram decide-kind wiring. Integrated as the eight-program union (registry, dict dispatch, routers helper, teardown enumerations). * feat(panel): Scales rebalance review queue Command Center gains the rebalance queue — per-item approve/reject with the action, target task, and rationale rendered. * feat(board): Mirror — quarterly positioning audit Project-scoped HoM cycle over messaging surfaces: README claims vs shipped reality, docs-site promises vs code, charter alignment — the audit is the HoM's own read-tool work with citations required; the server injects the charter, prior-cycle LEARN, and the shared rotation target. propose_ messaging_fixes mirrors the gap-fill verb (≤5 items, drift evidence naming claim + contradicting reality, participation gate); per-item CEO decide materializes BACKLOG source=mirror documentation tasks; full Telegram decide-kind wiring. Nine-program union across the shared surfaces. * feat(panel): Mirror messaging-fixes review queue * feat(board): Megaphone — HoM standing editorial calendar Cron cycle (3 days, org-scoped, gated on X credentials — drafting content nobody can post is pointless): the HoM receives a shipped-this-week digest plus Unreleased changelog bullets and files one propose_editorial_post (angle-validated, ≤280, brand voice) that materializes a held x_editorial draft through the SAME X-queue origination chokepoint release posts use — zero new approval surface, notifications and CEO decide for free. Complete-at-propose; cycle auto-closes. Ten-program union. * feat(panel): x_editorial source labels in the X queue surfaces * feat(board): Librarian — proactive playbook mining Biweekly org-scoped Auditor cycle: mines recurring non-private learning journals (≥2-count grouping with a recency fallback) against the existing playbook-title inventory and files one propose_playbook_drafts — 1-3 drafts, each with the repeated-pattern evidence that justifies it, duplicate titles rejected in-batch and against the live store. Drafts are created via PlaybookService directly (the Coroner precedent — the 'auditor curates but never drafts' do-verb invariant stays intact and tested) and land in the normal pending-curation queue the Auditor's own triage already surfaces; no new panel surface. Complete-at-propose; display-only CEO notification. Eleven-program union. * feat(board): War Room — release campaign planning EVENT program with a REAL originator (unlike coroner's stub): a release publish hooks a campaign brief beside the release-post seam, and the CEO's run-now originates on demand — the cron loop never fires it. The HoM designs a 2-6 post arc (teaser → launch → follow-up → spotlight; 280-cap, future strictly-ascending publish_after, stage vocabulary) and one propose_campaign call materializes each post as a held x_campaign draft through the X-queue chokepoint. V1 is manual-cadence by design: publish_ after renders as queue guidance and the CEO approves each post at its moment — nothing auto-posts, ever; the auto-schedule upgrade is a documented ceiling. Twelve-program union: full registry complete. * feat(panel): x_campaign labels + publish-after guidance in the X queue * feat(board): Barfly — adjacent-conversation replies Cron cycle (2 days, org-scoped, X-credentials gated): the engine searches X for conversations where RoboCo is relevant but unmentioned (new OAuth- signed search_recent on the client; queries + candidate cap configurable), screens every fetched tweet through the injection guard (stored unclamped — a clamp was truncating the candidate under the envelope, caught by the dev's own tests), dedupes via the existing x_seen_mentions ledger (no migration; also prevents double-drafting against the mentions poll), and opens one held HoM exploration carrying the screened candidates. propose_ conversation_replies enforces candidate-id-only replies (≤5, 280-cap); each materializes a held x_barfly draft through the X-queue chokepoint, threaded via a new in_reply_to seam on post_tweet that only x_barfly drafts use. The X redraft machinery is now dict-dispatch over per-source extractors with reply-ref carry for x_barfly. Thirteen-program registry. War Room's test fakes gained the new abstract search_recent stub. * feat(board): Dogfood — the PO walks the product The fourteenth and final registry entry, completing the catalog. EVENT program (release-publish hook beside the war-room hook + CEO run-now, both through the same real originator; the cron loop never fires it), project- scoped with shared rotation. The permission surface is the careful part: the PO's dogfood spawn — and ONLY that spawn — gets the Playwright MCP mounted, via a task-scoped fail-closed probe mirroring the video-authoring precedent (a PO spawned for roadmap/pest/scales never sees browser tools; tested both ways); the PM agent image bakes chromium unconditionally like the ux image, the mount stays task-gated in code. The walk targets the rotation target's live surfaces (panel_base_url only when the target is the org's own project, honest degradation otherwise); propose_friction_ fixes files ≤5 walked-path-evidenced items; per-item CEO decide materializes BACKLOG source=dogfood tasks; full Telegram decide kind. Also: megaphone/librarian/war_room arming keys restored to the settings validator — their panel toggles would have been rejected (dropped in earlier unions; the same silent-arming class the drill killed once already). * feat(panel): Dogfood friction review queue * chore(board): final whole-branch sweep fixes The night's closing adversarial pass over the integrated fourteen-program registry found ONE functional defect — the war-room test fakes' post_tweet predated Barfly's in_reply_to_tweet_id kwarg (LSP violation, the only red in an otherwise fully green gate) — plus doc/test drift, all fixed: the source-parity test completes to fourteen (spackle/mirror were silently absent while its neighboring comment claimed full coverage), the PO identity doc gains its missing Dogfood verb, the auditor quick-list gains propose_postmortem, three stale comments corrected (rotation docstring, panel registry header, X source enumerations), the dogfood release-hook gains the exception-swallow test its four sibling hooks already had, and the CHANGELOG's Unreleased section documents the whole Board Programs train. Full make quality: exit 0, all gates green. * docs: full documentation sweep for the Board Programs train CLAUDE.md's roadmap-engine entry superseded by the Board Program registry entry (all fourteen programs, arming, scoping, LEARN, guardrails) with the role verb tables and playwright row refreshed; docs/rag gains the agent- facing architecture doc plus full propose_* call-shape sections in the three board role docs, and corrects the strategy-engine section to shipped reality (only idle→roadmap is wired); docs/map covers the registry + all twelve engines with flags, gotchas, and drift notes. The 0.27.0 reference inventory confirmed only the release-executor's canonical set carries the version — left for the 0.28.0 cut. * feat(board): human titles + descriptions on every program surface Raw registry keys rendered as bare panel labels — an operator reading x_feature had no idea what enabling or running it does. The registry dataclass gains title/description (test-enforced non-empty for every entry, unique titles), the API passes them through, and every surface renders title-with-description-tooltip instead of the key: the Programs card (label, toggle hint, run-now toast), and the project settings participates-in/excluded-from checkboxes. --------- Co-authored-by: Renn F <rennf93@users.noreply.github.com> |
||
|
|
e77c3b7a63 |
feat(board): Board Program registry — Phase 1 (engine, LEARN ledger, per-project scoping, panel) (#689)
* feat(board): Board Program registry — generic trigger/dedup/originate/LEARN engine
One registry (foundation/policy/board_programs.py) + one BoardProgramEngine +
one orchestrator loop replace the bespoke roadmap/spotlight loops, behavior-
preserved: same sources, dispatch routing, one-open-cycle dedup (ledger rows
auto-close when their exploration task goes terminal, so x_feature's
complete-at-propose flow can't wedge), and live per-program interval
overrides with the tick capped at 1h.
program_armed() is the single arming chokepoint: the settings-store
board_program.<key>.enabled override when present, else the legacy flag —
routed through BoardProgramEngine, RoadmapEngine.run_cycle, and XEngine's
spotlight gate, so the panel toggle can never be a silent no-op against a
legacy boot flag.
LEARN: board_program_cycles (migration 087) accrues per-item CEO decisions
(exact attribution by exploration_task_id where the caller holds it) and
feeds the last closed cycles back into both exploration prompts. The
strategy engine's idle signal now opens a roadmap cycle (enabled+dedup
respected) instead of only nudging.
Per-project scoping (migration 088, projects.board_programs, dual polarity):
plain keys opt a project INTO project-scoped programs; "!key" opts it OUT
of an org-scoped program's outputs (default eligible — parity). Enforced at
propose_roadmap (names the excluded project) and defensively at materialize;
validation rejects unknown keys and meaningless polarity both directions.
API: GET /api/board-programs + POST /api/board-programs/{key}/run-now
(CEO-gated); settings keys for both migrated programs.
* feat(panel): Board Programs card + per-project program controls
Business page gains a Programs tab: per-program rows (role, trigger, scope,
open-cycle badge), enabled switch on the settings-store key, Run now
(disabled while a cycle is open). The edit-project dialog gains the
program controls next to the CI-watch/video toggles: participates-in
checkboxes for project-scoped programs, excluded-from checkboxes for
org-scoped outputs.
* test(board): full-gate hermeticity — mypy casts + shared-DB purge fixtures
make quality runs one pytest process over all suites against the shared
persistent DB: integration collects before unit, so the board-programs API
test's committed run-now state (settings-store overrides, an open cycle row,
its board_roadmap task) poisoned 13 downstream unit tests that pass in
isolation. The polluter now purges its own committed state in fixture
teardown, and the four consumer files get an autouse per-test purge
(board_program.% settings keys, ledger rows, open exploration tasks) so
they are hermetic regardless of collection order. Also the four
cast("UUID", ...) sites the tests-scope mypy run requires.
* feat(panel): re-home per-project program controls onto the settings page
Wave C deleted the edit-project dialog these controls originally landed in;
they now live on the project settings page's budget/ops card next to the
CI-watch/video toggles — participates-in switches for project-scoped
programs, excluded-from switches for org-scoped outputs, dual-polarity
tooltips, order-independent dirty tracking. Nine makeProject test fixtures
gain the required board_programs field the rebase left behind.
---------
Co-authored-by: Renn F <rennf93@users.noreply.github.com>
|
||
|
Renn F
|
037114338a |
fix(x): release-post drafts stop parroting changelog bullets
Highlights reorder to marketing order (Added/Changed before Fixed/Security — the drafting model anchors on highlight #1 and the changelog opens with Security), the prompt bans verbatim highlight copying and internal plumbing jargon, and the deterministic fallback becomes a generic announcement that can never quote a raw bullet. |
||
|
Renn F
|
902fc9a78c |
chore(models): finish the opus-5 replacement — no stale opus-4 fixtures remain
Every fixture, comment example, and panel mock that presented an Opus 4.x id as current now carries claude-opus-5 (or the current sonnet/haiku ids in the panel usage mocks). The only deliberate claude-opus-4 survivors are the pricing-table fragment and its tests — they price the historical usage rows, which would otherwise re-read as $0. |
||
|
Renn F
|
3c87596685 |
chore(models): move the opus alias to claude-opus-5
Claude Opus 5 released today — same $5/$25 sticker, 1M context; Opus 4.8 moves to legacy. The pricing table gains a dedicated claude-opus-5 fragment (the claude-opus-4 substring doesn't cover it, so without the row the fleet's Opus usage would silently cost-track as $0 — exactly what test_opus_is_priced now guards). |
||
|
Renn F
|
803638e8e9 |
chore(models): upgrade the opus alias to claude-opus-4-8
MODEL_MAP["opus"] moves off claude-opus-4-6 to the newest Opus tier at the same price; pricing already matched via the claude-opus-4 fragment, and a new test_opus_is_priced guard keeps the alias priced. Fixtures, panel mocks, and docs follow. |
||
|
|
987eb09c78 |
fix(workspace): role-aware worktree refresh at every spawn (#692)
* fix(workspace): refresh a present per-task worktree at every respawn ensure_worktree_self_heal treated an already-present worktree as a pure no-op (venv-link + chown only), so a worktree created once at first claim or first claim_review stayed frozen at that commit across every later respawn even as new commits landed on origin — the root mechanism behind a live multi-round QA/PR-gate bounce loop, where the reviewer kept re-examining its own stale round-1 checkout. _ensure_worktree_before_spawn now classifies the caller's role (WORKTREE_AUTHOR_ROLES: developer/documenter, mirroring the gateway commit tool's RBAC) and _refresh_present_worktree compares local HEAD against origin/<branch>: behind-or-equal fast-forwards for every role (never discarding an author's uncommitted edits to do it); strictly ahead is always left alone; diverged only resets for a pure reader, whose local history can never be anything but a stale prior-round checkout. conventions_check_for_task's list-vs-content gap (list from git objects, content from the physical worktree) is closed as a side effect: the reviewer's worktree is now current as of spawn, and the branch under review gains no further commits while it sits in awaiting_pr_review. * fix(workspace): refresh re-added worktrees; fail the dirty guard toward preservation - A pruned worktree re-added from a surviving local ref now runs the same fetch-and-classify refresh as a present one, so an evicted reviewer worktree cannot resurrect a stale checkout. - A failing git status reads as dirty, never clean: the guard that protects an author's uncommitted edits fails toward preservation. - The hard reset verifies the worktree is actually on the task branch first; a detached or drifted worktree is left alone with a warning. - The conventions-check docstring states the remaining second-claim ceiling instead of claiming full closure. --------- Co-authored-by: Renn F <rennf93@users.noreply.github.com> |
||
|
|
e97f46af6e |
fix(git): reviewer reads must prefer origin over a diverged local ref (#690)
* fix(git): reviewer reads must prefer origin over a diverged local ref _resolve_head_ref (GitService.diff/list_changed_files/read_file_at_branch) kept local priority on ANY divergence from origin, real or rewritten. A reviewer's clone parked on pre-rebase history after the branch's routine force-push sync stayed frozen there across every subsequent review round, while origin held every fix commit — QA repeatedly bounced work that had already landed. Every caller here is a reader, never the branch's own author mid-write, so origin now wins whenever it carries anything the local ref lacks; local keeps priority only when it strictly contains origin (unpushed commits, or equal). The read-only git MCP surface (roboco_git_log) hit the same staleness through a separate path: /api/git/log resolved the requested branch as a bare name straight off whatever the caller's own clone had on disk, with no fetch at all. It now routes through the same fixed _resolve_head_ref. * test(e2e): give the armed flow-verb timeout real headroom The armed value is also verb-2's entire execution budget (claim + every claim guard + set_plan + start + tracing gate), which grows as guards land; 1s flaked on loaded CI runners while passing locally. The cancel-and-release semantics only need the timeout far below the hang. --------- Co-authored-by: Renn F <rennf93@users.noreply.github.com> |
||
|
|
eb0dcb6ecb |
fix(orchestrator): task-scoped oscillation breaker for escalate/unblock ping-pong (#685)
* fix(orchestrator): task-scoped oscillation breaker for escalate/unblock ping-pong An escalation ping-pong oscillates a task between two agents (cell PM escalate_up -> BLOCKED -> main PM unblock -> restored -> respawn -> escalate again). The per-(agent, task) respawn gate never trips on it: the restored side is dispatched by _dispatch_claimed_without_agent, which consults no respawn counter at all, so one side of the round trip always has fuel regardless of the other's strikes — and even a tripped main-PM counter only stalls the task silently at blocked instead of surfacing the oscillation. - Strikes are counted task-scoped at the unblock() chokepoint (agent-agnostic; legitimate needs_revision rework never calls unblock, so it structurally cannot trip this), durable in the existing orchestration_markers column — no migration. - Progress between round-trips (commits / revision_count advancing) resets the count: real forward motion is not an oscillation. - On trip: the task is blocked with a HUMAN resolver (the budget-breach posture), both dispatchers stop respawning onto it, further unblock() refuses until an admin override clears the marker, and the CEO notification names both agents and the cycle count. - _notification_has_live_work now treats a HITL-blocked related task as no live work, closing the same loop for the admin-route escalation path. * fix(orchestrator): wire the oscillation trip to the dispatchers and make recovery reachable - TaskResponse serializes blocker_resolver_type: the dispatchers' HITL-blocked skip and the notification-path live-work check now actually fire over the wire instead of only against in-process rows. - The oscillation marker clears on every human transition out of BLOCKED (snapshot or not), and the human unblock route treats a tripped task as the requested intervention: clears the marker and proceeds, while the agent gateway verb keeps refusing. - The progress fingerprint includes the terminal-children count, so a coordination root whose children advanced between escalations resets instead of accruing toward a false trip. --------- Co-authored-by: Renn F <rennf93@users.noreply.github.com> |
||
|
|
23ae0ca217 |
fix(gateway): covers_parent_criteria hint that teaches the shape; CEO pause/resume (#686)
* fix(gateway): teach the delegate remediate + PM prompt the covers_parent_criteria shape; allow CEO through the plain pause route - A child draft rejected for missing covers_parent_criteria now gets a copy-pasteable corrected skeleton with the parent's real criteria inlined, and the PM delegation guidance shows the field as part of every child draft — a PM no longer loops on a rejection that named the field but never showed the shape. - The plain pause route now authorizes the CEO tier like its sibling lifecycle routes; agent-side pause restrictions are unchanged. * fix(gateway): delegate-coverage hint heals and degrades on legacy parents - The coverage-reject path self-heals a criteria-bearing parent whose ids are empty or out of length before rendering the hint, so the skeleton always shows real references; the renderer itself also falls back to quoted criterion texts for any criterion without an id instead of emitting a placeholder or truncating the listing. - The remediate names both legal reference forms (id or exact text) again. - Route comments state the pause/resume check as deliberately CEO-only instead of claiming a precedent whose role set is wider. * test(gateway): real TaskTable rows in the remediation hint round-trips mypy over tests/ rejects a SimpleNamespace where unknown_ac_refs takes a TaskTable; instantiating the ORM row directly needs no session and types cleanly. --------- Co-authored-by: Renn F <rennf93@users.noreply.github.com> |
||
|
|
4b2546ae19 |
fix(findings): path-shaped file refs + per-round collapsible findings (#687)
* fix(findings): enforce path-shaped file refs; group panel findings by round - The findings chokepoint rejects a file that is not a repo-relative path shape (prose like a PR reference validated before, and the panel then rendered a doomed file-content fetch for it) — narrative belongs in evidence, the remediate says so. - The task-detail Findings tab groups findings into per-round collapsible sections (newest expanded) and only attempts a code snippet for a path-shaped file ref, so historical prose refs render as plain metadata instead of a broken loader. * fix(findings): admit client-repo path conventions; teach the file-less option - The shape gate reviews arbitrary client projects, not just this repo: plus and at-sign join the character class so SvelteKit route files, @types dirs, and @2x assets stay citable. Spaces stay excluded — they are the prose signal. - The file-rejection remediate names the file-less option for cross-cutting findings. - The client mirror notes its deliberate non-ASCII divergence from the server gate (unicode server-pass renders snippetless, fail-open). --------- Co-authored-by: Renn F <rennf93@users.noreply.github.com> |
||
|
|
21910d75ea |
chore(board): revive dormant board wiring — research key, pitch flow, auditor playbooks (#684)
* chore(compose): pass research key/provider + provisioning token/org through to the orchestrator ROBOCO_RESEARCH_API_KEY / ROBOCO_RESEARCH_PROVIDER and ROBOCO_PROVISIONING_TOKEN / ROBOCO_PROVISIONING_ORG were absent from every compose environment stanza, so .env values never reached the container: research silently ran on the NullProvider (empty results forever) and any approved pitch died on ProvisioningDisabledError. .env.example also falsely claimed the provisioning creds are panel-managed. * feat(board): pitch CEO notification + auditor playbook-draft surfacing A proposed pitch now nudges the CEO (APPROVAL notification + Telegram link to the Pitches tab, best-effort — a send failure never fails the verb). auditor_triage surfaces the oldest pending playbook draft once anomalies are clear — the curation verbs were granted but nothing ever pointed the Auditor at the review queue; the scheduled audit prompt names the discovery path. * docs(prompts): pitch doctrine section + auditor reply-only-dm drift fix board.md never mentioned the pitch verb, so no board agent ever had a reason to call it — it gets a dedicated section mirroring the roadmap/spotlight ones, plus a roadmap-exploration escape hatch (needs-its-own-repo ideas pitch instead). product-owner.md gains its missing propose_roadmap + pitch entries. The flat 'Auditor has no dm' claims are corrected to the real grant: never initiates, reply-only in a CEO-opened thread. Doctrine guarded by a prompt-content test. --------- Co-authored-by: Renn F <rennf93@users.noreply.github.com> |
||
|
|
71f5426e40 |
fix(git): never discard committed local work in rebase_onto_base (#683)
The shared rebase primitive (dev sync_branch verb, PM/CEO rebase path,
submit-freshen, merge-conflict resolver) opened with fetch -> checkout ->
unconditional reset --hard origin/<head_branch>. The dirty-tree gate
protects uncommitted edits only; the reset silently rewound past every
committed-but-unpushed commit — and the commit do-verb never pushes, so
mid-rework a dev routinely has exactly that. The force-with-lease push
then republished the truncated branch as authoritative (the lease
matched the freshly-fetched, never-moved origin ref).
rebase_onto_base now classifies local vs origin/<head> post-fetch:
- behind/equal: reset --hard origin as before (origin loses nothing)
- strictly ahead: reset skipped — the rebase runs from the local tip and
the lease'd push publishes the previously-doomed commits
- diverged: a patch-equivalence probe (rev-list --right-only
--cherry-pick) first rescues the self-inflicted residue of a prior
rebase whose force-push failed (treated as ahead, self-heals on
retry); only genuine two-sided divergence returns a new
{status: diverged, local_only, origin_only} — no reset, no rebase,
no push, neither side silently discarded
- an absent local ref is recovered from origin (branch + checkout,
never reset)
Callers: the sync_branch verb maps diverged to an actionable envelope
steering to i_am_blocked (stash-preserved note included); the
submit-freshen hard-rejects it like conflicts; the merge-conflict
resolver already escalates any non-rebased/superseded status to the
CEO and degrades gracefully (pinned by test, no code change).
New real-git suite (bare origin + clone, no subprocess mocking)
asserts origin-side outcomes: ahead-publishes, behind-adopts,
diverged-refuses-untouched, absent-ref recovery, superseded,
conflicts, and wedge self-heal on retry via a rejecting pre-receive
hook.
Co-authored-by: Renn F <rennf93@users.noreply.github.com>
|
||
|
|
3516d925fe |
fix(tasks): reconcile acceptance_criteria_ids at the update chokepoint (#682)
Every post-create rewrite of acceptance_criteria (task PATCH route, prompter update_live_draft / _patch_batch_child / update_live_batch) routes through TaskService.update()'s generic field loop, which overwrote the criteria without touching acceptance_criteria_ids — leaving ids mismatched or empty, and an empty id list silently disabled the parent-coverage gate entirely. - New pure _reconcile_ac_ids: one id per new criterion; text-unchanged criteria keep their id (children and findings reference criteria by id or exact text — a blanket re-mint would orphan every live reference), new/reworded text mints fresh, dropped criteria drop theirs. create() now stamps through the same helper (explicitly supplied ids still win). - update() derives acceptance_criteria_ids whenever acceptance_criteria is rewritten without an explicit id list. - The parent-coverage gate self-heals a criteria-bearing row whose ids are empty/out-of-length (re-stamp in place) instead of returning early and silently waiving coverage for the whole subtree. Co-authored-by: Renn F <rennf93@users.noreply.github.com> |
||
|
|
c4ba351ae0 |
fix(sequencing): reachability-aware claim bar + sequence_held surfacing (#681)
Three coupled claim-path bugs from the 2026-07-24 live incident, fixed at the shared root: - The edge-agnostic sequence bar phantom-held a task behind an unrelated, never-connected same-parent sibling that coincidentally shared a lower raw sequence (stamp_wave_sequence stamps from a partial per-task view). _claim_blocked_by_sequence now branches on is_batch_root_subtask: a MegaTask root-subtask (globally-computed Kahn wave, a deliberate staged-release barrier) keeps the strict rule unchanged; every other same-parent context routes through the pure sequence_blocker_id, which only blocks on a real transitive predecessor via dependency_ids UNIONED with completed_dependency_ids. A task with no same-parent dependency edge at all falls back to the raw bar unchanged (#452 preserved). - The hold surfaced as claim()'s bare None and was misdiagnosed by the verb runner as a concurrent-transition invalid_state. New sequence_hold_reason + a proactive _sequencing_claim_guard return a dedicated Envelope.sequence_held naming the blocker, on both the PENDING and NEEDS_REVISION reclaim paths. - give_me_work offered tasks the claim gate then rejected: both offer paths (list_pending_for_agent, _drop_dependency_held) now consult the bar via the exact claim predicate (is_pending_claim_blocked, extended to NEEDS_REVISION). Co-authored-by: Renn F <rennf93@users.noreply.github.com> |
||
|
|
a3f84f3165 |
chore(routing): retire haiku from delivery-lifecycle roles (#680)
Haiku can't reliably emit the structured envelopes the lifecycle now runs on — pass_review's per-AC criteria_verified, delegate's covers_parent_criteria, the findings ledger. A haiku QA/PM claims, gets validation-rejected, idles, respawns, and loops without advancing a task (2026-07-24 live: fe-qa on haiku looped four awaiting_qa tasks to zero progress). The per-token savings (~2x under the Sonnet-5 promo, 3x after) are dwarfed by the cost of a review that never completes. Three coordinated changes: ROLE_MODEL_MAP's qa/documenter defaults move haiku -> sonnet (the actual source of the live incident); the cost_tiered developer:low -> haiku seed retires to empty (the floor would upgrade it anyway); and a structured-verb capability floor upgrades any below-floor Anthropic assignment to sonnet at resolution — from a pin, a ROLE row, or a future map edit — in both the assignment and legacy paths. Non-Anthropic providers are untouched (an Anthropic-tier floor, not a provider policy). pr_reviewer/auditor stay on opus. Co-authored-by: Renn F <rennf93@users.noreply.github.com> |
||
|
|
f8b4a6755c |
fix(pr-review): fleet PRs are ours by branch ownership, not author identity (#668)
With a GitHub App bound, fleet PRs are authored by <app-slug>[bot] whose author_association is NONE — the inbound classifier's author heuristics read that as an outsider and ingested the org's own dev-stream PR as external_pr for adversarial review (2026-07-23 live: PR #667). The repo-owner author check only ever covered the PAT era. _ingest_pr_if_reviewable now skips any same-repo PR whose head branch an active task owns BEFORE the author-based classification, and active_task_owns_branch widens from the single polled project to every project sharing its git_url (the poll collapses a monorepo's cell-projects to one canonical project, so a sibling cell's ownership must count — the same sibling scope external_review_task_exists already uses, now shared via _repo_sibling_project_ids). A deleted-fork head (GitHub sends head.repo=null) now classifies as fork, failing closed to review instead of risking a silent ownership skip on a branch-name collision. Residual, documented: an org PR whose task went terminal with the PR left open falls through to the author heuristics. Co-authored-by: Renn F <rennf93@users.noreply.github.com> |
||
|
|
a036c97985 |
fix(tg): cockpit data correctness — real GLM pricing, display timezone, agent activity tracking (#666)
* fix(tg): cockpit data correctness — real GLM pricing, display timezone, agent activity tracking
Three root causes behind the Mini App/bot showing wrong numbers:
Pricing: glm-5.2 gets a grounded per-token rate (z.ai published pricing,
$1.40/$4.40/$0.26 per 1M, source+date in the table comment) so a GLM
fleet day stops reporting $0.00 for half a million tokens; ungrounded
Ollama-Cloud models render "subscription (untracked)" instead of a bare
zero (is_ollama_cloud_model, consumed directly by the cockpit). Side
effect, intended and documented: honestly-priced GLM now trips the
downgrade-only comparator for new qa/documenter complexity pins.
Display timezone: the cockpit bucketed days in UTC for a GMT+2 operator.
New pure foundation module display_time (resolve_zone/local_date/
trailing_dates/day_bounds_utc, DST-correct with tests for the 23h/25h
days) + ROBOCO_DISPLAY_TIMEZONE (IANA-validated, default UTC); the
cockpit's spend/velocity series bucket raw session/completion rows by
the display zone. The UTC-keyed rollup table and the main dashboard are
deliberately untouched.
Agent activity: AgentTable.status was never set to ACTIVE and
current_task_id was never written anywhere — "active: 0, working: []"
was structurally permanent. Every claim path now marks the claimant
ACTIVE with rollback symmetry (_finalize_claim for dev/PM claims,
_qa_or_doc_claim for QA/doc/PR-gate claims, pr_review_claim for external
review) and every release path clears it (pass/fail QA, pr_pass/pr_fail,
complete_review, advance-to-PM-review, reaper unclaim, voluntary
unclaim, reassign retarget, pool divert, admin transitions, unblock
restore-to-in-progress). The bot's /status shares the cockpit's fleet
derivation so the two surfaces can't disagree. Known ceiling, commented:
one current_task_id column shows a multi-root coordinator PM's most
recent claim only.
Drill: sonnet develop -> sonnet adversarial (refuted the original
chokepoint coverage claim; QA/doc/reviewer paths were unwired) ->
correction round (wired them all + restored a dropped assertion, deleted
a dead helper and the dead subscription_billed field) -> review.
* fix(db): post_update on AgentTable.current_task breaks the flush cycle
agents.current_task_id and tasks.assigned_to reference each other, so a
flush touching both rows — every claim now marks its agent ACTIVE — is
an instance-level circular dependency SQLAlchemy cannot topologically
sort. The e2e smoke's full verb paths (12 tests) hit it; the unit and
integration suites never flush both dirty rows with relationships
loaded. post_update emits the FK as a second UPDATE, the canonical fix
for mutually-referencing rows.
* fix(budgets): enforce only explicitly-set budgets — no per-TaskType defaults
The per-TaskType default cap table blocked an unbudgeted coordination
root one opus planning turn in ($1.50 PLANNING default vs. real
coordination spend) — a false positive by design the moment the fleet
runs a priced model. Budgets are now explicit-input only:
effective_task_budget_usd returns None for an unset budget_usd, the
budget sweep skips enforcement (and never prices spend) on None, and
the unblock re-check passes on None so clearing the budget field is
itself a valid resolution. The project monthly cap stays as the
explicit-input fleet-wide backstop. Panel copy tells the truth
("No cap" placeholder; empty = uncapped), and the TaskType default
table plus its resolver are deleted.
---------
Co-authored-by: Renn F <rennf93@users.noreply.github.com>
|
||
|
|
0c1d450a05 |
fix(intake): emit the complete reply when no text deltas streamed (#665)
The intake/secretary chat driver treated StreamEvent text deltas as the ONLY text channel: AssistantMessage TextBlocks were always swallowed as already-streamed (the double-render guard). The CLI's partial-message emission turned out to be remotely gated — on 2026-07-23 the NAS containers got zero stream_event lines from the identical binary, flags, SDK, model, and settings that stream fine elsewhere — so the guard became a total blackout: replies were generated, the relay carried only init/status/turn_end, and the CEO saw nothing. SdkIntakeSession.send now tracks whether any text delta arrived during the turn; normalize() emits an AssistantMessage's complete text only when none did. Streaming mode is byte-identical (deltas render live, completes stay suppressed); gated mode delivers the reply as one block instead of nothing. Covers Secretary (same machinery). Regression tests: fallback emission, default suppression, and both modes end-to-end through the session layer. Co-authored-by: Renn F <rennf93@users.noreply.github.com> |
||
|
Renzo FandGitHub
|
226e1b586a | fix(routing): escape hatch out of mix mode — clear-all overrides + pins warning (#663) | ||
|
Renzo FandGitHub
|
d4b7e1e7b8 | fix: post-finale completeness sweep — routing surface, provider config, budgets, compose env, interactive exemption (#661) | ||
|
|
21d6730400 |
feat(providers): Gemini CLI provider — ModelProvider.GEMINI (#660)
* feat(providers): Gemini CLI provider — ModelProvider.GEMINI Mirrors the grok blueprint with source-verified divergences (all facts pinned against google-gemini/gemini-cli @ 9681621c): no refresher daemon — Google's refresh tokens are reusable, so the RO host mount is COPIED to a writable container-local ~/.gemini and each container refreshes in-process independently (the write-back crash risk on RO never triggers); settings.json renders security.auth.selectedType 'oauth-personal', experimental.enableAgents=false (subagent ban), autoConfigureMemory=false with a bounded heap; tool scoping rides the tiered TOML Policy Engine (deny-only rules that yolo mode structurally cannot beat); gemini -p with --output-format stream-json; usage parsed from the run's own stdout stats — the adversarial pass caught the parser reading the json-mode nested shape while the entrypoint runs stream-json's FLAT shape (every real run would have priced $0 forever, hidden by fixtures sharing the assumption) — now flat-primary with the nested shape as cited fallback; rate-limit classified from structured error.type only (model-echo immune), native exit 41 auth passthrough; per-model pricing for the three GA models; migrations 084 (enum) + 085 (seed) complete the 082-085 finale chain. V1 excludes interactive intake/secretary. Stack-merge required two behavior-preserving complexity refactors in the shared park/usage plumbing (a park-pair loop; a usage-reader dispatch dict). * fix(providers): route gemini usage read through the containment barrier Mirrors the codex/grok fix — _gemini_usage_json now delegates to _read_usage_json_contained, so CodeQL's path-injection alert on the gemini read is resolved by the same resolve-and-contain guard. --------- Co-authored-by: Renn F <rennf93@users.noreply.github.com> |
||
|
|
c70ff3cf9a |
feat(providers): Codex CLI provider — OpenAI via ModelProvider.OPENAI (#659)
* feat(providers): Codex CLI provider — OpenAI via ModelProvider.OPENAI Mirrors the grok blueprint end to end: CodexCliProvider (RO ~/.codex mount, ANTHROPIC_* blanked), an orchestrator-side codex_auth.py refresher (JWT-exp staleness, atomic rewrite, lock-serialized single-use rotation, --check backstop; the CLI's own in-process refresh write no-ops on the RO mount by design — margins keep the orchestrator ahead of the CLI's 5-minute window), config.toml rendering with required=true gateway MCP servers, execpolicy deny rules (forbidden-only), per-role --sandbox (developer=workspace-write, review/doc roles read-only), codex exec --json with pinned ROBOCO_CODEX_CLI_MODEL (gpt-5.3-codex), usage summed from typed turn.completed events priced via the real 4-bucket split, dedicated image + entrypoint, registry/park/finalize/ compose/release wiring. V1 excludes interactive intake/secretary. Per adversarial review: migration 083 seeds the openai provider row enabled=True (without it every routing path 404'd — the whole feature was operationally dead code; grok needed the same seed in 039), the panel picker gained the OpenAI catalog group it silently lacked, and exit classification is structural — only stderr and error.message fields from error events are sniffed (word-boundaried patterns, exact auth phrases, bare 'login' dropped), so the model echoing on-topic words can never false-park the provider fleet-wide, proven by a benign-transcript test. Known open risk flagged, not claimed: whether codex's workspace-write OS sandbox excludes /app is unverified, and no hook mechanism exists to port the bash-guard defense-in-depth. * fix(providers): containment barrier on usage.json reads (code scanning) CodeQL flagged the codex usage read as path injection — correctly: os.path.basename does not neutralize '..', and the upstream segment validator isn't in CodeQL's taint model. The grok/codex reads collapse into one _read_usage_json_contained helper that resolves the built path and refuses anything outside the resolved usage root — a hostile id can never escape regardless of upstream drift. Traversal + containment regression tests added; a stray noqa in the test file replaced with a named constant per repo rule. * fix(providers): use realpath+startswith containment CodeQL recognizes The is_relative_to() guard was a real barrier but not in CodeQL's py/path-injection sanitizer model, so the alert persisted. Switch to the canonical os.path.realpath + startswith(root + os.sep) form, which CodeQL recognizes as a path-traversal barrier; behavior is identical (refuse any candidate resolving outside the usage root). * fix(providers): regexp-allowlist the usage-id segment (CodeQL barrier) Neither is_relative_to nor realpath+startswith was recognized by CodeQL's py/path-injection sanitizer model across the str->Path->open flow. Sanitize the tainted component at the source instead: the id must fullmatch a strict slug token ([A-Za-z0-9][A-Za-z0-9._-]*, no separators, no '..'), which CodeQL recognizes as a path-injection barrier; the realpath+startswith containment stays as defense-in-depth. * fix(providers): standalone regexp guard so CodeQL recognizes the barrier The sanitizer was one disjunct of a compound 'or' condition, which CodeQL's guard analysis does not trace as a barrier. Split the regexp fullmatch into its own single-condition guard (the redundant '..' check is dropped — the required alphanumeric first char already excludes it). --------- Co-authored-by: Renn F <rennf93@users.noreply.github.com> |
||
|
|
165892dc62 |
feat(routing): cost-tiered complexity routing + saved presets (#656)
The 08-31 lever: model_assignments gains one compound rung —
AGENT_SLUG > ROLE('{role}:{complexity}') > ROLE > GLOBAL — so a
low-complexity task can route to a cheaper tier while coordinators stay
pinned. Structurally opt-in: zero rows means byte-identical routing
(pinned by a named test across every precedence case), the cost_tiered
apply-mode (seeds developer:low→haiku) is reachable only from the
explicit PM-gated endpoint — verified no startup path can apply it.
Overrides are downgrade-only (input-price comparator), allowlisted to
{developer, qa, documenter} — cell_pm excluded per the org's own
coordinator definition and its documented weak-model incidents — and
validated at write time (disabled/unconfigured provider rejected with
remediation; cross-provider-family overrides warn explicitly).
Per adversarial review: the four mode-switch applies now spare compound
rows exactly like agent pins (the 2026-07-17 unscoped-wipe class, new
victim, same fix extended via one shared wipe helper) with panel cache
invalidation + truthful confirm dialogs; preset apply validates the
entire payload BEFORE the wipe (validate-all-first), with a savepoint
crash test proving rollback.
Presets (CEO request): routing_presets table (migration 082) snapshots
the full mix — mode, per-agent overrides, complexity rows — with
save/apply/delete endpoints and a panel preset bar; applying skips
since-removed models with per-entry notes, never silently.
Task complexity threads task_id through _resolve_agent_route at both
call sites; taskless spawns unchanged. 235 backend + 23 panel tests.
Co-authored-by: Renn F <rennf93@users.noreply.github.com>
|
||
|
|
10f039c36f |
feat(eval): golden-task eval harness + doctrine cohort stamp (#655)
* fix(notifications): exponential backoff + CAS claim for expired-unacked re-escalation The sweep re-escalated every expired unacked ack-required notification on every ~60s tick, forever — the live incident: 3 fresh blocker escalations + Telegram DMs per minute from a static stale pile. Now each notification carries reescalation_count / last_reescalated_at / reescalation_delivered_count (migration 079): first fire at expiry, then doubling intervals from 1h capped at 24h, hard stop after ROBOCO_NOTIFICATION_MAX_REESCALATIONS (default 5) with one permanent log carrying attempts-vs-delivered so 'seen and ignored' is distinguishable from 'route never worked'. The due/wait/capped decision is a pure function in foundation/policy/communications.py. Per adversarial review, the attempt slot is claimed by compare-and-set (UPDATE ... WHERE reescalation_count = :n) BEFORE delivery — the previous draft leaned on the 60s dedup window, which never engages for BLOCKER_ESCALATION (_LOOP_PRONE_TYPES excludes it), so concurrent sweeps would have double-delivered. A lost claim skips delivery outright. Legacy rows read as count=0 and keep today's first-fire semantics. 61 tests incl. a two-session CAS race and a real alembic upgrade/downgrade round trip. * feat(budgets): per-task and per-project cost budgets (flag-gated) tasks.budget_usd + projects.monthly_budget_usd (migration 080, chained on 079; adds ix_agent_spawn_sessions_task_id since both enforcement seams filter on bare task_id). Behind ROBOCO_TASK_BUDGETS_ENABLED (default off, feature-flags card) — verifiably inert when off. Claim-time: a project-month-spend guard applies to WORK-STARTING claims only (i_will_work_on / i_will_plan) — per adversarial review, review/ doc/gate/inbound-PR claims are exempt so in-flight work can always finish reviewing and merging at cap. Spend counts closed sessions' estimated_cost_usd PLUS open sessions priced live from token snapshots (the original closed-only sum read parallel long sessions as $0). Sweep-side: the existing budget sweep also prices the active task's spend vs budget_usd (TaskType defaults when null); on breach the task is BLOCKED (HUMAN resolver, budget marker) BEFORE the graceful stop so the unclaim no-ops and the dispatcher never respawns onto it, and the CEO notification names both recovery steps. unblock on a budget-blocked task re-checks live spend and refuses while still over — no silent re-breach loop. Panel: budget inputs in both dialogs (0 rejected — a zero budget silently blocks everything), spend logic consolidated in TaskService.task_spend_usd. 42 new tests incl. a real-DB spend-query suite and a two-tick non-refire sweep test. * feat(eval): golden-task eval harness + doctrine cohort stamp roboco/eval: 6 BenchTaskSpec fixtures run through the real lifecycle in a disposable environment (the e2e_smoke harness's fake GitHub + local git origin + throwaway DB catalog — real isolation, not convention), scored deterministically (terminal status, revision_count, cycle time, tokens/cost via the agent_spawn_sessions task_id join) plus a local- model judge whose output is nested under a non_deterministic-marked object so cohort diffs don't read judge noise as regression. CLI: python -m roboco.eval run --role <slug> --cohort <name>. Source- checkout-only by declared posture (deptry-scoped ignore + a hard ImportError guard naming why; tests/ never ships in images or wheels). agent_spawn_sessions.doctrine_version (migration 081, chained on 080) is stamped at spawn-session finalize from the composed prompt layers — with the session's model column it identifies a cohort durably. Per adversarial review: bench runs patch the vault flags off (they were writing real markdown into the operator's vault), and the real-spawn OrchestratorStageSpawner is deliberately cut to NotImplementedError — spawned containers' MCP wiring resolves to the production orchestrator under real agent UUIDs, so real spawns wait for a dedicated follow-up; the injectable scripted spawner is the working path. Full suite 13852 passed / 94% coverage in the source worktree; deptry/mypy/xenon clean. --------- Co-authored-by: Renn F <rennf93@users.noreply.github.com> |
||
|
|
7c8453e210 |
feat(budgets): per-task and per-project cost budgets (flag-gated) (#654)
* fix(notifications): exponential backoff + CAS claim for expired-unacked re-escalation The sweep re-escalated every expired unacked ack-required notification on every ~60s tick, forever — the live incident: 3 fresh blocker escalations + Telegram DMs per minute from a static stale pile. Now each notification carries reescalation_count / last_reescalated_at / reescalation_delivered_count (migration 079): first fire at expiry, then doubling intervals from 1h capped at 24h, hard stop after ROBOCO_NOTIFICATION_MAX_REESCALATIONS (default 5) with one permanent log carrying attempts-vs-delivered so 'seen and ignored' is distinguishable from 'route never worked'. The due/wait/capped decision is a pure function in foundation/policy/communications.py. Per adversarial review, the attempt slot is claimed by compare-and-set (UPDATE ... WHERE reescalation_count = :n) BEFORE delivery — the previous draft leaned on the 60s dedup window, which never engages for BLOCKER_ESCALATION (_LOOP_PRONE_TYPES excludes it), so concurrent sweeps would have double-delivered. A lost claim skips delivery outright. Legacy rows read as count=0 and keep today's first-fire semantics. 61 tests incl. a two-session CAS race and a real alembic upgrade/downgrade round trip. * feat(budgets): per-task and per-project cost budgets (flag-gated) tasks.budget_usd + projects.monthly_budget_usd (migration 080, chained on 079; adds ix_agent_spawn_sessions_task_id since both enforcement seams filter on bare task_id). Behind ROBOCO_TASK_BUDGETS_ENABLED (default off, feature-flags card) — verifiably inert when off. Claim-time: a project-month-spend guard applies to WORK-STARTING claims only (i_will_work_on / i_will_plan) — per adversarial review, review/ doc/gate/inbound-PR claims are exempt so in-flight work can always finish reviewing and merging at cap. Spend counts closed sessions' estimated_cost_usd PLUS open sessions priced live from token snapshots (the original closed-only sum read parallel long sessions as $0). Sweep-side: the existing budget sweep also prices the active task's spend vs budget_usd (TaskType defaults when null); on breach the task is BLOCKED (HUMAN resolver, budget marker) BEFORE the graceful stop so the unclaim no-ops and the dispatcher never respawns onto it, and the CEO notification names both recovery steps. unblock on a budget-blocked task re-checks live spend and refuses while still over — no silent re-breach loop. Panel: budget inputs in both dialogs (0 rejected — a zero budget silently blocks everything), spend logic consolidated in TaskService.task_spend_usd. 42 new tests incl. a real-DB spend-query suite and a two-tick non-refire sweep test. --------- Co-authored-by: Renn F <rennf93@users.noreply.github.com> |
||
|
|
1d5a8e846f |
fix(notifications): exponential backoff + CAS claim for expired-unacked re-escalation (#652)
The sweep re-escalated every expired unacked ack-required notification on every ~60s tick, forever — the live incident: 3 fresh blocker escalations + Telegram DMs per minute from a static stale pile. Now each notification carries reescalation_count / last_reescalated_at / reescalation_delivered_count (migration 079): first fire at expiry, then doubling intervals from 1h capped at 24h, hard stop after ROBOCO_NOTIFICATION_MAX_REESCALATIONS (default 5) with one permanent log carrying attempts-vs-delivered so 'seen and ignored' is distinguishable from 'route never worked'. The due/wait/capped decision is a pure function in foundation/policy/communications.py. Per adversarial review, the attempt slot is claimed by compare-and-set (UPDATE ... WHERE reescalation_count = :n) BEFORE delivery — the previous draft leaned on the 60s dedup window, which never engages for BLOCKER_ESCALATION (_LOOP_PRONE_TYPES excludes it), so concurrent sweeps would have double-delivered. A lost claim skips delivery outright. Legacy rows read as count=0 and keep today's first-fire semantics. 61 tests incl. a two-session CAS race and a real alembic upgrade/downgrade round trip. Co-authored-by: Renn F <rennf93@users.noreply.github.com> |
||
|
|
fa459998b4 |
feat(git): env-ladder rung protection at the shared remote-delete chokepoint (#651)
Rung protection lived only in delete_task_branch; the post-merge PR- source cleanup (and the stale-branch sweep's shared primitive) could still delete a branch that IS a ladder rung. _protected_branches_for_ deletion(slug) — field ∪ rung names, null-ladder shim included — now feeds _delete_remote_branch_best_effort, so every remote deletion path is covered; delete_task_branch's local rung check is removed as exactly subsumed (verified byte-identical comparison semantics). Bonus closed gap: a renamed trunk (default_branch 'trunk', null ladder) is now delete-protected, which the hardcoded main/master floor never covered. Per adversarial review, the deletion lookup fails CLOSED: a raised project lookup skips the delete with a warning (a skipped best-effort delete just retries next sweep — free safety), while a genuinely-gone project proceeds with the hardcoded floor (its ladder is meaningless). The rebase/sync resolver stays fail-open — a refused rebase on a DB blip would wrongly block work, a different tradeoff, now documented. Panel tooltip updated to the new truth. 29 tests. Co-authored-by: Renn F <rennf93@users.noreply.github.com> |
||
|
|
3806317aa7 |
fix(guard): operator-scoped XFF hop peers + tailnet allowlist (live-incident fix) (#650)
Two coupled hardenings from the chain-peers adversarial rounds plus the
root-cause fix for the live post-deploy incident where the CEO was
blocked from the panel ('IP not allowed: 100.x.x.x').
Hop peel-set: the whole docker bridge pool leaves the XFF hop set — hops
are now loopback plus operator-named single addresses only
(ROBOCO_GUARD_TRUSTED_CHAIN_PEERS, plain IPs; CIDR entries rejected with
a warning because a range readmits sibling containers). Default-empty
closes the CGNAT-forge residual outright; a gateway-fronted Tailscale
Serve deploy sets its real gateway IP, and a rate-limited detection log
names exactly that IP when an unconfigured host-proxied tailnet chain is
seen, so the silent-regression shape is observable. The connecting-peer
gate (may nginx present XFF at all) deliberately keeps the broad bridge
pool — different check, unchanged.
Incident root cause: guard-core's whitelist is an EXCLUSIVE allowlist
(any non-member is refused), so honestly resolving the tailnet client IP
made ip_security reject the CEO. The tailnet CGNAT range joins
_guard_whitelist() deliberately: Tailscale authenticates device
membership before a packet arrives, real-IP stamping still buys correct
attribution, and any future non-tailnet exposure keeps full scrutiny.
Both compose files now pass ROBOCO_GUARD_EMERGENCY_WHITELIST through to
the orchestrator (the operator escape hatch previously did nothing in a
compose deploy).
NAS is running ROBOCO_GUARD_PASSIVE_MODE=true as interim mitigation —
flip back to false when this deploys. 66 tests.
Co-authored-by: Renn F <rennf93@users.noreply.github.com>
|
||
|
|
da4d9b333d |
feat(git): protected-branches enforcement + panel editor (#649)
projects.protected_branches existed end-to-end but nothing consulted it — the panel had no editor and the git safety checks used hardcoded sets. Now: GitService._protected_branches_for(slug) (frozenset, stripped, fail-open to the hardcoded floor with a warning log) is unioned — never replacing, only tightening — into rebase()'s refusal set, the shared _delete_remote_branch_best_effort skip set (threaded through every caller: task cleanup, PR merge/close cleanup), and sync_task_branch, which now refuses to force-push a protected-named head (the dev-facing sync_branch verb path the HTTP-only fix would have missed). Matching is exact and case-sensitive; an empty list degrades to exactly the old hardcoded behavior, pinned by union-floor regression tests (master/main stay refused regardless of the project list). Panel: chips editor for the field in the edit-project dialog (add via Enter/comma, paste-splitting on comma-separated lists, dedup, clear-to- empty persists []) with an honest tooltip scoped to what is actually enforced. Tests cover both the incumbent GitHub-App dialog suite and the new Protected Branches suite in one harness. Co-authored-by: Renn F <rennf93@users.noreply.github.com> |
||
|
|
4585a248ce |
feat(x): redraft loop on CEO reject — feedback re-enters the draft flow (#648)
* feat(x): redraft loop on CEO reject — feedback re-enters the draft flow A rejected X draft's reason used to die with the cancel. reject() with a non-blank reason now schedules a redraft after its commit (defer_after_commit; fresh session; never blocks or fails the HTTP response): XEngine.redraft_from_rejection re-drafts the same source kind via the local model with the reason and rejected body folded in as revision guidance, originating ONE fresh held draft — mirroring the video pipeline's reauthor_from_rejection. Local-model failure or empty output originates nothing (no degraded copies); markers carry forward whole so a redrafted reply/spotlight stays fully functional downstream; bodies ride the same 280 clamp; the open-posts cap holds. Hardened per adversarial review: reject() is now idempotent on an already-CANCELLED target at both check sites (mirroring approve's already_rejected guard — a replayed reject schedules nothing), and the dedup check+originate runs under a non-blocking identity-keyed Redis lock (SET NX + compare-and-del, matching the approve/reject mutex style) so racing rejects can't stack duplicate drafts — lock held or Redis down skips the redraft, which is always safe. Tests pin the fresh-session contract by session identity, the replay no-op, the lock-skip, and clean up their own committed rows. * fix(tests): runtime UUID import + typed task-id coercion in x cleanup helper CI's quality gate runs mypy over tests/ (the local pass covered only roboco/): the _delete_tasks calls handed ORM-typed ids where uuid.UUID was expected. Coercing at the call sites then exposed that UUID was imported under TYPE_CHECKING only — a runtime NameError. Import moved to runtime; both call sites coerce explicitly. --------- Co-authored-by: Renn F <rennf93@users.noreply.github.com> |
||
|
|
583d43f213 |
fix(guard): resolve the tailnet client behind host-proxy hops for the WAF (#646)
fastapi-guard peels a fixed trusted_proxy_depth=1 from X-Forwarded-For (the rightmost entry, which nginx itself recorded). That is correct for every chain except host-proxied tailnet traffic (Tailscale Serve → nginx), which arrives as [tailnet-client, loopback-or-bridge-gateway] — depth-1 resolves it to a whitelisted hop IP, leaving WAF/ban/rate-limit inert for the whole /tg surface (the documented ceiling). ClientIpResolutionMiddleware (pure ASGI, wraps SecurityMiddleware so it runs first) stamps guard_core's request.state.client_ip cache — its supported pre-resolution seam — for EXACTLY that shape: peel known local hops (loopback + docker bridge pool) from the right, stamp only when at least one hop was peeled AND the candidate is in the tailnet CGNAT range (100.64.0.0/10). Every other shape abstains, so direct LAN clients, agent containers relaying through nginx (even with forged public-IP prefixes), and all-hops operator traffic resolve byte-for-byte as before. Documented residual: a same-bridge container forging a CGNAT prefix only DE-privileges itself (loses its whitelist exemption). XFF is read first-occurrence to match Starlette's own header semantics, and a wiring test pins the middleware mount ORDER, not just presence. Co-authored-by: Renn F <rennf93@users.noreply.github.com> |
||
|
|
2e889c7009 |
feat(lifecycle): inherit advanced upstream base on work re-claims (#644)
* feat(lifecycle): inherit advanced upstream base on work re-claims A re-claim reused a branch cut at an earlier claim, so upstream work merged since (UX/UI landing on the root after the cell branch was cut) never reached BE/FE branches — divergence and avoidable conflicts. _finalize_claim now merges the advanced base into the pre-existing branch via the dependency-lineage merge: already-ancestor is a no-op, a conflict aborts at the cut point and leaves a transition note steering the agent to sync_branch, a clean merge logs an audit trail; never fails the claim. Double-gated: by role (developer/cell_pm/main_pm — QA/documenter/gate claims review the branch as pushed and never move it) AND by pre-claim status (pending/needs_revision only — a PM's i_will_plan re-claim of its own awaiting_pm_review task must not move a branch that already passed QA + the PR gate). Fresh cuts already branch from the live remote base. Cell-PM prompt now orders reading the upstream design docs before planning. * fix(lifecycle): extract base-inheritance gate predicate for xenon budget The four-condition inline gate pushed _finalize_claim to cyclomatic rank C; the quality gate caps blocks at B. The decision moves to a pure module-level predicate, byte-for-byte the same logic. --------- Co-authored-by: Renn F <rennf93@users.noreply.github.com> |
||
|
|
b91229f487 |
fix(orchestrator): break the notification-driven respawn loop (#643)
The escalation/approval dispatchers spawn a notification's recipient every cooldown window for as long as it stays pending. These spawns carry no task_id, so the PM respawn breaker never sees them — a single wedged alert/escalation whose recipient never resolves it respawns that recipient forever. Observed live: fe-pm's unacked alerts kept main-pm/fe-pm spawning every ~2-3 min for 6+ hours. Two guards, both gating the spawn after the existing cooldown: - A hard per-(agent, notification) attempt cap (notification_spawn_max_attempts, default 5): once a notification has respawned its target that many times without being acknowledged, stop and log once. The count is id-scoped and survives map pruning (re-stamp), so a fresh escalation is unaffected. - A live-work check before spawning: skip when the notification has expired, is stale past notification_spawn_max_age_seconds (default 6h — wedged or reloaded from before a restart), or its related task is already terminal. Fail-open — a failed task fetch or unparseable field never suppresses a real escalation. Co-authored-by: Renn F <rennf93@users.noreply.github.com> |
||
|
Renn F
|
f74131a122 |
fix(release): commit the approve() COMPLETED write under the release lock
approve() flushed the published proposal's COMPLETED status but left the durable commit to the background caller (_run_approve_background), which runs after approve()'s finally has already released the Redis lock. In that window a concurrent reject() could acquire the freed lock, re-read a row whose COMPLETED write was only flushed (invisible to its own session under READ COMMITTED), pass its guard, and flip the just-published proposal to CANCELLED — last writer winning the row. The bug-sweep (#638) fixed reject()'s side of this but left approve()'s, so reject()'s 'fails closed' guarantee didn't hold end to end. Commit COMPLETED while still holding the lock (mirroring XPostService._post), so it's durable before release and a racing reject sees it and refuses. A cross-session regression test proves a fresh connection sees COMPLETED the moment approve returns (it read 'pending' before the fix). |
||
|
|
17de29545a |
[6788ce7f] Silent bug sweep: concurrency, state integrity, engine edge-cases, panel data freshness (#638)
* [943d8c4d] Frontend data freshness and approval-queue reliability audit (#631)
* [233a8b0f] WebSocket reconnect message-loss audit and fix (#625)
* [233a8b0f] fix(panel): add REST catch-up to useNotificationStream on WS reconnect
connection.ts has no message buffering/replay, so a notification published
while the CEO bell's socket was down (disconnected/reconnecting) was lost
forever instead of merely delayed. Add a reconnect-triggered GET
/notifications?unread_only=true catch-up folded into the existing
notification_id dedup so a notification delivered both via catch-up and
live WS is never double-counted, and make clearMessages drop the held
catch-up batch too. use-a2a-live.ts and use-rate-limit-websocket.ts were
audited and already have working reconnect-triggered REST fallbacks
(verified via a2a/page.tsx, rate-limit-banner.tsx, usage-overview-panel.tsx
and their existing F083 tests) so no fix was needed there.
* [233a8b0f] docs(panel): add comprehensive WebSocket hooks reference and reconnect architecture guide
Add panel/docs/frontend/hooks.md with full API reference for useWebSocket, useNotificationStream (with new REST catch-up behavior), useAgentStream, useA2ALiveStream, and useConnectionStatus. Include examples, best practices, and testing guidance.
Add panel/docs/architecture/websocket-reconnect.md documenting the message-loss mitigation pattern: Strategy 1 (REST catch-up for events, used by useNotificationStream) and Strategy 2 (REST invalidation for state, used by A2A/rate-limit consumers), plus the dedup logic ensuring no notification is double-counted on reconnect.
---------
Co-authored-by: Frontend Developer 1 <fe-dev-1@roboco.tech>
Co-authored-by: Frontend Documenter <fe-doc@roboco.tech>
* [d5315683] fix(frontend): add distinct toast feedback for silently-swallowed x-post and release-proposal statuses, plus regression tests for all 4 approval queues (#626)
Co-authored-by: Frontend Developer 2 <fe-dev-2@roboco.tech>
* [cd953838] Data-hook null-guard audit and API client 429 retry-by-method fix (#630)
* [cd953838] fix(panel): gate 429 retry by HTTP method, add hook null-guard regression tests
* [cd953838] chore(conventions): waive test-fixture wrapper in hooks null-guard test
* [cd953838] docs(frontend): document API rate-limit retry behavior and null-guard audit results
Added `docs/frontend/api-rate-limiting.md` to document the 429 retry strategy: GET/PUT auto-retry, POST/PATCH/DELETE require X-Idempotency-Key header. Updated `docs/frontend/hooks.md` to confirm the data-hook null-guard audit found all hooks already have correct `enabled` guards and include a regression test suite for the board-review poll on/off behavior and enabled-guard assertions.
---------
Co-authored-by: Frontend Developer 1 <fe-dev-1@roboco.tech>
Co-authored-by: Frontend Documenter <fe-doc@roboco.tech>
---------
Co-authored-by: Frontend Developer 1 <fe-dev-1@roboco.tech>
Co-authored-by: Frontend Documenter <fe-doc@roboco.tech>
Co-authored-by: Frontend Developer 2 <fe-dev-2@roboco.tech>
* [4534c71a] Backend concurrency, state-machine, and engine audit (#634)
* [41de844a] fix(lifecycle): sync CLAIM_RULES with runtime + clear stale claimant on PM hand-off (#627)
Two confirmed state-machine gaps found while auditing lifecycle.py,
task_lifecycle.py, the _ESCALATABLE_TO_BLOCKED bypass, and every
_REVIEW_QUEUE_STATES entry point:
- lifecycle.py's CLAIM_RULES/claim-ActionSpec/StatusTransition table
did not grant CELL_PM/MAIN_PM re-claim of AWAITING_PM_REVIEW even
though task.py's runtime _ROLE_CLAIM_STATUSES already granted it
and claimed the spec agreed -- the two tables had silently drifted,
breaking i_will_plan re-claim on an awaiting_pm_review task.
- docs_complete's _maybe_advance_to_pm_review pre-assigns a specific
owning PM via assigned_to but left claimed_by/active_claimant_id
pointing at the outgoing documenter, unlike every sibling transition
into a review-queue state. A stale active_claimant_id makes
content_actions.py's _active_claim_violation wrongly reject the
newly-assigned PM's own content writes before it formally claims.
Reassign claimed_by + active_claimant_id to the owning PM alongside
assigned_to.
Adds a regression test asserting the documenter's stale claim does not
survive the docs_complete -> awaiting_pm_review hand-off.
Co-authored-by: Backend Developer 2 <be-dev-2@roboco.tech>
* [0c46f666] Engine dedup race + sequencing.py edge-case audit (#628)
* [0c46f666] fix(sequencing): dedup race audit + collision-edge fallback bug
Audited the list-open-then-originate dedup pattern across six engines:
RoadmapEngine, XEngine.run_cycle, DepUpdateEngine, and CIWatchEngine each
run inside exactly one sequential orchestrator-loop asyncio task (no other
call site invokes run_cycle), so they cannot race with themselves; their
in-cycle dedup sets/keys are correctly built before any commit. SelfHealEngine
is the same shape. VideoEngine.open_video_task is genuinely different: it is
reachable from the release-publish hook, the feature-spotlight hook, and the
on-demand POST /video/request route, so two overlapping calls for the same
occasion can both pass the "no open task yet" check before either commits.
Fixed by wrapping the check+insert in a short-lived Redis mutex (reusing
HeartbeatMutex) keyed by occasion, mirroring XPostService's existing
lock pattern, with a regression test proving only one of two concurrent
calls creates a task.
Verified ReleaseExecutor's half-landed retry path (release_commit_sha):
apply_version_bumps and write_changelog_entry both run as uncommitted
working-tree edits before commit_and_push's single `git add -A` + commit,
so a bumped-version-without-changelog state can never reach origin (and
therefore can never be observed by a fresh retry clone) - confirmed correct
with a real-git-repo regression test, no fix needed.
Fixed sequencing.py's dev_task_collision_edges: the `if edges: return edges`
short-circuit dropped the same-assignee-lane fallback entirely whenever ANY
surfaced sibling pair produced a collision edge, even for a completely
unrelated same-assignee pair with no declared surface. Now the fallback
always runs, skipping only pairs the analyzer already ordered (so the two
mechanisms can never disagree on direction for the same pair).
Verified sequencing.py rule 3 (all-shared batch generates no edges): correct
by inspection (_shared_last_edges skips every pair when both are shared) and
confirmed with a regression test - no fix needed.
* [0c46f666] docs(reference): concurrency audit summary - engine races, fixes, verified patterns
---------
Co-authored-by: Backend Developer 2 <be-dev-2@roboco.tech>
Co-authored-by: Backend Documenter <be-doc@roboco.tech>
* [8f7f167a] Redis mutex pre-lock write audit (#629)
* [8f7f167a] Redis mutex pre-lock write audit: add cross-session regression test for XPostService.approve
Audited x_post_service.py, video_post_service.py, release_proposal.py, and
heartbeat_mutex.py for the pre-lock DB-write anti-pattern (a session write
that happens before the SET NX / HeartbeatMutex acquire returns a token,
letting a losing racer's stale write clobber a winner's committed state).
XPostService.approve, VideoPostService.approve, and
ReleaseProposalService.approve/reject already implement the correct
validate-pure-pre-lock, apply-under-lock pattern (the XPostService fix
already shipped per CHANGELOG.md: "X edited_body write deferred into the
single-flight lock (M5)"). HeartbeatMutex holds no AsyncSession at all, so
the anti-pattern is structurally inapplicable there.
Adds a genuine cross-session concurrency regression test to
test_x_post_service.py (a real second DB connection, not an in-process
mock) mirroring VideoPostService's existing cross-session test, proving a
concurrently-committed post survives and the CEO's edited body never lands
on the just-posted row.
* [8f7f167a] Remove redundant inline comments flagged by QA in cross-session regression test
Both comments restated what the surrounding docstrings already say
explicitly, per QA findings F-dbadd8f0 (line 294) and F-27ac051e (line
631) — no behavior change, tests re-verified green against a sandbox
Postgres.
* [8f7f167a] Remove inline trailing comments flagged by QA (correct file this time)
QA findings F-e6f3e6a6 and F-24189858 cited tests/unit/services/
test_x_post_service.py:294 and :631 across 5 revision rounds, but that
file never contained the flagged comment text — a repo-wide grep for
the exact quoted strings shows both comments actually live in the
mirrored tests/unit/services/test_video_post_service.py file, in its
own cross-session concurrency regression tests (the caption-edit and
tiktok-skip tests). Removed both there:
- "# externally visible to the "concurrent" session below" on the
db_session.commit() call
- "# never attempted without credentials" on the tiktok_poster.calls
assertion
Both restated what the surrounding docstrings/test names already say;
no behavior change. Verified with the full make quality gate against a
sandbox Postgres/Redis: 13,717 passed, 94.41% coverage, clean except
one pre-existing unrelated failure in tests/unit/api/test_cloud_auth.py
::test_login_route_parses_oauth2_form_not_query_params, which connects
to the app's default localhost:5432 Postgres (not the db_session
sandbox fixture) and is unreachable in this sandboxed environment —
structurally unrelated to the auth subsystem this task never touches.
* [8f7f167a] Redis mutex pre-lock write audit (round 7): add cross-session regression tests for reject() lock protection
Round-7 QA findings F-7eb9fbcb, F-06f39a2e, and F-4d56e49b claim
XPostService.reject(), ReleaseProposalService.reject(), and
release_executor._await_proc() lack lock protection / a CancelledError
handler — but their cited line ranges (255-267, 429-454, 241-257)
describe a pre-fix, shorter version of these functions that predates
commit
|
||
|
Renn F
|
a1233b2aeb |
fix(panel): V6 review gaps — honest errors, safe secretary start, real tests
CEO A2A mutations now invalidate the Mine-list query key so the list refreshes without the socket; the tg Metrics tab renders explicit error notes instead of confident zero stats when a section's fetch fails; the tg Secretary chat checks a new registry-backed /secretary/live/active route before auto-starting, showing a Take Over button instead of silently killing a session live on another device; the AI-routing card surfaces roster fetch errors instead of an empty grid; the shared acceptance-criteria editor caps at the backend's 7-item limit; the board-tab comment no longer calls the task sheet read-only. Tests: task-sheet approve/reject interactions (not just visibility), a non-demo metrics error-state test, secretary take-over branches, and dashboard-router auth-gate coverage (the e2e harness now mounts the dashboard router so the gate is actually exercised). |