mirror of
https://github.com/rennf93/roboco.git
synced 2026-08-03 07:23:24 +02:00
eb0dcb6ecbf141a2160c5bd8a50f7707544f2be4
6
Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
|
e9ca7d4036 |
Delegation detail-fidelity + PM-loop hardening (#541)
* feat(gateway): delegation detail-fidelity — details survive hand-off, both directions
Details thinned out at every delegation hop: a PM child task mapped to no
parent criterion was legal (coverage only surfaced at submit_up, after the
whole wave ran — a 12-subtask docs tree grew through 8 review rounds that
way, one child titled 'docs page and route wrapper' shipping only the
page), and QA could pass work on a gestalt read (a 4-scene video brief
shipped 3 scenes past every gate because the features existed only in
prose). Three chokepoint gates:
- delegate (down): every child must declare covers_parent_criteria
resolving against the parent's real acceptance criteria — no mapping or
an unresolvable ref rejects naming every offending child and the valid
criteria; the success envelope carries parent_ac_coverage
{covered, uncovered} so a wave-planning PM sees remaining gaps in the
same turn. Full coverage stays enforced at submit_up (waves stay legal).
- pass_review (up): mandatory criteria_verified — one {criterion,
evidence} entry per task AC, matched by the findings ledger's
id-or-exact-text matcher, evidence soup-checked and capped; rejects
naming the unverified criteria; entries render deterministically into
qa_notes as '[AC] <criterion> — verified: <evidence>' lines. The old
count-only ac_verdicts gate is superseded (arg kept for back-compat).
- video briefs (structured detail at origination): an enumerable feature
list (release highlights, or input_props.highlights carried onto a
reject re-author) becomes its own scene acceptance criterion, bounded to
the AC caps; a re-author without highlights carries the
feedback-addressed criterion instead.
Extracted findings.py's criterion matcher into shared unmatched_criteria /
uncovered_acceptance_criteria instead of duplicating it; criteria_verified
joins the WAF free-text exclusion set like findings/issues.
* fix(gateway): break the block/unblock wedge — four hardening fixes from the live PM loop
A cell task looped fe-pm/main-pm block/unblock for hours (10 cycles, 43
spawns): a transient GitHub API error resolving CI became an unwaivable
blocker finding whose own fix text said no code change was required, the
submit freshness guard then demanded a commit no finding called for,
escalate_up auto-blocked, and main-pm's correct recovery plan 422'd on
the approach length cap, degrading it to a bare unblock. Four fixes:
- pr_pass CI-unresolvable refusal is now explicitly transient-worded:
retry pr_pass shortly, do NOT pr_fail over a CI-status lookup error —
a platform blip is not a code finding
- submit freshness guard grants ONE unchanged-head resubmission per
head sha when the findings ledger has zero open rows (all addressed
without code changes) — stamped via the resubmit_unchanged_head
marker so the same head can never loop a second time
- unblock carries a flip breaker: block_flip_count marker, and at the
third flip a one-shot CEO notification flags the task as structurally
wedged (unblock itself still succeeds — the breaker signals, it does
not wedge recovery)
- i_will_plan's approach cap truncates at 800 chars instead of
rejecting — an over-detailed plan must never cost the PM its turn
---------
Co-authored-by: Renn F <rennf93@users.noreply.github.com>
|
||
|
|
cea3e56628 |
feat(lifecycle): revision findings ledger — structured failure feedback, persisted and delivered down the chain (#486)
* feat(lifecycle): revision findings ledger — structured QA/PR/PM/CEO failure feedback, persisted and delivered down the chain Every bounce used to survive only as flattened prose: rounds overwrote each other in notes_structured, request_changes persisted nothing, two raw dev_notes appends were silently destroyed by the next handoff note, and the dev prompt pointed at fields (qa_notes via evidence(), pm_notes) the API never delivered. Agents re-interpreted and re-discovered every failure before they could start fixing it. - task_review_findings (migration 071, append-only): file/line/severity/ criterion(AC-id-validated)/expected/actual/fix/evidence per finding, with origin (qa|pr_gate|pm|ceo), round, and an open->addressed->verified lifecycle (waived reserved); new tasks.pm_notes + PmReviewContent give request_changes a structured home - producers: fail_review/pr_fail/request_changes take findings=[...] (prose issues shimmed+merged for one release, deprecation-logged); ceo_reject validates its reason (no 500), lands an origin=ceo finding, and bumps round+audit on branchless coordination roots; guardrails at the verb chokepoint (nudge >5, hard reject >10, field caps, traversal-safe file); the dev_notes data-loss appends are removed; new task.request_changes + task.ceo_reject audit events close rework attribution - delivery: qa_notes/pr_reviewer_notes/pm_notes carry the deterministic [F-id8] rendering; claim briefings, evidence(), the REVISION_REQUIRED spawn prompt, PM triage bounced-blocks, and A2A bodies deliver open findings; round-N+1 QA and gate reviewers get the full prior ledger; panel Findings tab + bounced-xN chip; metrics pm_rejects/ceo_rejects + findings counts; vault task notes render a Findings section (fail-open) - resolution closes for every origin: i_am_done and submit_up/submit_root take resolved_findings gated by FINDINGS_ADDRESSED (owner-gated so a stale non-owner PM can never mutate the ledger); pass_review/pr_pass/ complete verify-stamp same-transaction; ceo_approve stamps best-effort - 24 real-DB integration tests drive the full loop through the real choreographer; full suite 12856 green * docs: revision findings ledger sweep — CLAUDE.md, map, RAG corpus - CLAUDE.md: new ledger section + corrected request_changes row - docs/map/review-findings.md (new subsystem map) + surgical updates to task-service/pr-gate-review/metrics-observability/vault/panel maps - docs/rag: producers' findings contract across qa/pr-reviewer/developer/ cell-pm/main-pm/ceo role docs (the PM docs were missing request_changes entirely), verb references, and a new architecture/review-findings.md disambiguating ledger findings from convention findings * test(e2e): resubmit resolves the pr_fail finding per the ledger contract The scripted pr_fail revision loop resubmitted submit_up without resolved_findings — correctly rejected now that FINDINGS_ADDRESSED gates the PM resubmit verbs (green locally, red only in CI since the e2e suite skips without ROBOCO_E2E_SMOKE=1). The scripted PM now reads the open ledger row pr_fail persisted (new open_finding_ids arc helper) and resolves it on resubmit, asserting the open set drains — exercising the coordinator half of the new contract end to end. --------- Co-authored-by: Renn F <rennf93@users.noreply.github.com> |
||
|
|
8f6dde9a50 |
feat(sandbox): pluggable per-engine registry (postgres/redis/mongo) (#324)
* feat(sandbox): pluggable per-engine registry (postgres/redis/mongo) Replaces the hardcoded postgres+redis branches in the provisioner and the env emitter with a registry of SandboxEngine specs (image, run args, readiness probe, connection, ROBOCO_TEST_* env) in a pure low module (roboco/models/sandbox.py). VALID_SANDBOX_SERVICES is derived from the registry — single source of truth — and the provisioner + orchestrator iterate it, so adding an engine is one class + one registry line, not another branch. Adds a mongo:8-alpine engine (ROBOCO_TEST_MONGO_*) as the third service alongside postgres/redis. Also fixes the cold-pull loop that stranded v0.19.0 board agents with empty error strings: docker run pulled inline under a 20s deadline, so a NAS cold pull was killed, cancelled, and re-pulled from scratch forever. _ensure_image now inspects + pulls (300s) before run; provisioning errors log type+message so a bare TimeoutError no longer shows as "". Panel edit-project dialog: postgres/redis toggles -> a Set<string> multi-select driven by a SANDBOX_SERVICES catalog, so new engines appear in the UI by adding to the catalog. Tests: engine parity (allowlist==registry, unique slugs/images, no None leak in env, SandboxInfo aggregates every engine), mongo provision + env injection, plus the existing postgres/redis provision/env/spawn/janitor suite updated to the registry shape. 821 unit / 5 skip green; ruff + mypy (360 files) clean. * docs(sandbox): reflect pluggable engine registry + mongo across docs CHANGELOG (0.19.0): Added entry for the pluggable sandbox engine registry (postgres/redis/mongo) + Fixed entry for the cold-pull loop/empty-error strand that boarded v0.19.0 board agents. docs/map (9 files): sandbox subsystem blurbs, SandboxProvisioner rows, _maybe_provision_sandbox/_append_sandbox_env rows, feature-flag rows, the migration-057 row + v0.17.0 delta, and the models.md VALID_SANDBOX_SERVICES note — all retitled to DB/Redis/Mongo via the engine registry (roboco/models/sandbox.py), with the one-class-one-line extension story and the _ensure_image cold-pull fix. Production-network (roboco_data) lines left as postgres+redis — mongo is sandbox-only, not a prod service. docs/rag (3 files): sandbox-db.md rewritten around the registry (engine list, generic _provision_engine, image pre-pull, ROBOCO_TEST_DB_*/REDIS_*/MONGO_* incl. MONGO_AUTH_DB=admin, single emit_env); config-reference sandbox flag row + subsection retitled; db-network-isolation framing broadened to postgres/redis/mongo. preconditions-and-rejections left untouched (its hit was an unrelated gateway see-also link). * test(e2e): harden umbrella close terminal reads with bounded wait-for-state The MegaTask umbrella close test flaked once on CI (ceo-approve returned 200 but the re-fetch saw awaiting_pm_review) then passed on re-run. The production path is deterministic: complete -> main_pm_complete -> submit_pm_review -> escalate_to_ceo -> ceo_approve -> commit, all on one session, all awaited; the fire-and-forget completion hooks are isolated (own session, best-effort, never touch task.status or the request session). 20 local runs could not reproduce it. The one real surface is the read pattern: the e2e stack commits on the uvicorn thread's loop and reads via a separate loop (run_db -> asyncio.run with a fresh engine), so a terminal single point-read can race a still-draining completion hook on a contended runner. Replace the two terminal point-reads with a bounded wait_for_status poll. Strictly better than a one-shot read: absorbs the transient, and a genuine state bug still surfaces via the timeout branch asserting against the last-read state. --------- Co-authored-by: Renn F <rennf93@users.noreply.github.com> |
||
|
Renn F
|
da0fa2e33a |
test(e2e): feature-spotlight end-to-end scenario (catches the unregistered-verb gap)
Drives the Head of Marketing calling propose_feature_spotlight through the real
do_server -> /api/v1/do route -> ContentActions -> XEngine, then asserts a held
x_feature draft is created (confirmed_by_human=False) and the exploration task
completes. Red-then-green verified: reverting the do_server registration
(
|
||
|
|
48f2944086 |
MegaTask umbrella e2e scenario + batch root-subtask completion fix; comms dead-code deletion (#296)
* chore(panel): delete the five dead comms components The comms audit found communications-view, channel-sidebar, channel-item, message-list, and message-item exported but rendered by no page — the live /communications page and the session detail render their own inline content and import only MessageComposer and MessageTypeBadge, which stay. Verified zero consumers outside the dead cluster before deletion; panel gates green (tsc, lint, 187 tests). * feat(tests): e2e scenario 4 — MegaTask umbrella; fix batch root-subtask completion wall Scenario 4 seeds an umbrella + two dependency-linked root-subtasks: sequencing hold proven (unmet_dependency on RS2's i_will_plan while RS1 lives), RS1 completed through the entire real chain to a master merge, hold lifts, RS2 completes, umbrella closes branchless via ceo-approve and never carries a PR. Product fix it surfaced on first run: _main_pm_complete_guard and escalate_to_ceo refused ANY parented task as 'not a root', but a batch root-subtask is parented (the umbrella) BY DESIGN — both sites now consult is_batch_root_subtask, plain subtasks stay refused. Live root-subtasks previously needed CEO god-mode to close. Regression tests added; built subagent-driven (Sonnet 5) and reviewed. --------- Co-authored-by: Renn F <rennf93@users.noreply.github.com> |
||
|
|
d1cf6ecbf3 |
Wave 1: PR-gate turn cut, task search, trace timestamps, Secretary edits + e2e scenarios 2–3 (#295)
* feat(tests): e2e scenario 2 — the PM merge chain through the PR gate
Shared arcs extracted (arcs.py: canonical-company seeding + dev/qa/doc
segments); scenario 2 seeds a root->cell->dev hierarchy mid-flight, rides
the child through the scenario-1 arc into the cell branch (real squash
via the fake GitHub), then submit_up -> claim_gate_review/pr_pass ->
dispatcher re-claim (mirrored) -> PM complete merging cell->root. This is
the exact PM->reviewer->PM turn sequence the wave-1 turn cut shortens —
the BEFORE-net. Learned seams scripted: commit-subject validator (>=20
chars), reviewer learning-note gate, pr_pass clears ownership by design.
* feat(runtime): PR-gate turn cut — assembled parents auto-submit to the reviewer
When every child of an assembled parent is terminal, the closure
dispatcher now runs the real submit_up/submit_root through the internal
API as the owning PM (_try_auto_submit) instead of spawning the PM for
that turn — the submit's substance is deterministic gate code. Any gate
refusal falls back to the classic PM closure spawn; pr_fail routing and
the PM's final merge turn are unchanged; umbrellas never auto-submit.
ROBOCO_PR_GATE_AUTO_SUBMIT_ENABLED default-on; task.auto_submitted audit
row per cut. Proven by e2e scenario 2b (real API, real gates, real git)
against scenario 2 as the before-net.
* feat(notes): structured note sections carry a written_at trace stamp
Sections are overwrite-in-place, so without a stamp there was no way to
reconstruct WHEN a dev/qa/doc/reviewer note landed (CEO reMarkable item:
trace TIMESTAMPS). apply_structured_note stamps ISO written_at beside
the model fields; the panel notes tab renders it next to each card
title (pre-stamp rows render nothing). Progress updates, commits, and
journal entries already carried timestamps — this was the one gap.
* feat(tasks): server-side task search — title, details, and id prefix
The task list's search box only matched titles client-side, and the
trimmed summary payload deliberately carries no description — so
keyword/details/id search was impossible in the browser by design.
GET /tasks/summary gains q (ILIKE over title+description, id-prefix
match, composed with team/status and the view-permission scoping);
the panel debounces the box into the summary fetch and drops the
title-only client filter that would have hidden description matches.
* feat(wave-1): trace timestamps, real task search, Secretary task edits
- apply_structured_note stamps written_at per section; the panel notes
tab shows it (the one trace surface without a timestamp).
- GET /tasks/summary?q= searches title+description+id-prefix server-side
(summaries carry no description by design); panel debounces into the
fetch and drops the title-only client filter.
- Secretary control_task gains a CEO-gated edit action over the content
allowlist, and GET /secretary/tasks?q= resolves task names to ids for
the chat. PM-side expansion deferred per the CEO's 'not that much'.
* fix(workspace): dep-update probe scrubs the inherited venv pin
Under uv run the orchestrator's process tree carries VIRTUAL_ENV, and a
uv-based dep_update_command in the throwaway probe clone would target
that venv instead of the clone's — the same hazard _uv_subprocess_env
already guards on the install path.
* build: private per-repo uv cache — isolate from machine-wide uvx servers
Root cause of the recurring rich/pip/bandit rot, with evidence: uv cache
clean timed out on the ~/.cache/uv lock ('is another uv process
running?') — three uvx mcp-server-fetch processes (Claude Code fetch MCP,
one alive since Wednesday) share that cache and race repo syncs on it;
poisoned entries then survive venv rebuilds because rm -rf .venv never
touches the cache, and every re-link reproduces the breakage. UV_CACHE_DIR
now pins <repo>/.uv-cache (gitignored). The earlier UV_NO_SYNC
serialization stays as defense-in-depth but was not the whole story.
* feat(tests): e2e scenario 3 — pr_fail revision loop + root→CEO chain
3a: reviewer pr_fail with a concrete issue -> needs_revision ->
i_will_plan re-entry (full plan gates) -> real fix lands on the cell
branch (the unchanged-PR hard gate refuses resubmit until it does) ->
clean second pass -> merge. 3b: submit_root -> gate -> Main PM complete
escalates the root to the CEO -> the REAL approve-and-merge endpoint
squash-merges to the origin's master. Harness gains the tasks router, a
seeded CEO identity, origin_commit, and a fake GitHub whose head.sha is
recomputed live (real-GitHub semantics the unchanged gate reads). Seeds
now encode the real shape: delivery roots are team=main_pm and
planning-typed.
---------
Co-authored-by: Renn F <rennf93@users.noreply.github.com>
|