* fix(mcp): delegate tool carries the collision surface the B1a gate demands
TASK_AT_DELEGATE (5fc85419) requires intends_to_touch on code delegations,
but the MCP delegate tool never gained the parameter — PMs were rejected
with incomplete_input and could never comply (live fleet-wide delegation
wall, 2026-07-02). Adds intends_to_touch / adds_migration / touches_shared /
depends_on to the tool and forwards them; parity test locks the invariant.
* fix(git): assembly-integrity guard accepts squash-merged children
git cherry patch-matches each child commit individually, so a squash merge
(N patches -> one commit, new patch-id) read as 'work missing' and the #11
guard refused every legitimate submit_up (live 2026-07-02: S6 cell, three
squash-merged children at the branch tip). A parent commit carrying the
child's [taskid8] prefix now proves the child landed; children with no
marker stay flagged — the original incident the guard exists for.
* fix(git): diff head prefers origin when the local ref is behind it
Assembled branches advance on ORIGIN as child PRs squash-merge on GitHub,
but _resolve_head_ref preferred the inspecting clone's parked local ref —
the PR-gate reviewer's evidence diff was built from a pre-merge snapshot
and re-flagged work that had already landed (two false pr_fail verdicts
on the S6 cell PR, live 2026-07-02). When both refs exist and the local
ref is strictly behind origin, resolve to origin/<branch>; local-ahead
(unpushed) and diverged refs keep priority, single-ref cases unchanged.
* test(mcp): plan-gate fields must be tool parameters (parity class lock)
Extends the delegate parity test to every choreographer plan-depth gate:
a gate that can reject with missing=[field] must name only fields the
corresponding MCP tool can send, else the agent can never comply.
* perf(api): wire TaskSummaryResponse into a bounded /tasks/summary route
The panel fetched /api/tasks unbounded and full-fat — 2MB per refresh
measured live (2026-07-02), ~21KB/task, and the trimmed
TaskSummaryResponse was dead code. /tasks/summary returns exactly the
fields list views render (~50x lighter); the status-only branch of
/tasks now honors its limit, and the eleven unbounded task list routes
are capped.
* perf(panel): kill the per-page request flood and fat payloads
Every page load funneled ~85 default-prefetch RSC requests + 665KB of
images + the 2MB task list through the browser's six HTTP/1.1
connections — real data calls queued ~2s before being sent (measured
via Playwright resource timing, 2026-07-02).
- prefetch={false} on all 59 Links (sidebar, task rows, kanban cards,
list rows) — ~85 requests/refresh down to a handful
- icon/apple-icon/logo resized to render size: 665KB -> 54KB; unused
219KB PNG removed
- task list fetches the trimmed /tasks/summary (2MB -> ~100KB),
normalized into the Task shape so list consumers keep their types
- ReactQueryDevtools rendered only in development
* fix(api): Annotated limit defaults so direct-call tests get real ints
Query(...) positional defaults arrive as Query objects when a route
function is invoked outside the HTTP layer (integration tests call
handlers directly) and broke the new [:limit] slices.
* fix(api,panel): summary carries completed_at + board_review_complete
The metrics page computes velocity client-side from completed_at and the
CEO approval queue gates on board_review_complete — both were nulled by
the summary normalizer, so Completed Today/Week read 0 against 63 real
completions and approved-board tasks could vanish from the queue. The
queue also renders quick_context, so it fetches the full list (small,
status-scoped) via tasksApi.listFull instead of the summary.
* fix(runtime): spawn manifest workspace_path follows the task's project
_build_manifest_for_agent hardcoded the roboco project workspace for
every agent; a guard-core task's manifest claimed /data/workspaces/roboco
while the container cwd sat in the task worktree. The manifest now takes
the same _resolve_workspace_cwd the container -w uses — one resolver,
both surfaces agree by construction.
* fix(runtime): respawn breaker catches status ping-pong loops
Any status CHANGE fully reset the strike counter, so a blocked <->
in_progress oscillation — which changes status on every spawn while
advancing nothing — never tripped the gate (live 2026-07-02: 8 spawns
over two hours). A status never seen on the (agent, task) still fully
resets; a REVISITED status gets a bounded reset budget mirroring
tracing_resets, after which strikes accrue and the gate fires.
* fix(runtime): unassigned-QA dispatch spawns without pre-claiming
The transitioning pre-claim moved awaiting_qa -> claimed before the QA
agent existed; the spawned agent's claim_review/pass_review both demand
awaiting_qa, so it bounced twice and unclaimed (live 2026-07-02,
ba7b751c). Matches _spawn_assigned_qa and the external-PR reviewer
dispatch: no pre-claim, the agent claims itself via claim_review.
* fix(tests): narrow await_args before kwargs access (mypy union-attr)
* Minor upgrades
* fix(policy): team-match gate gains org-wide exemption; resume/unblock/activate now team-matched
needs_team_match sat in its permissive fallback since shipping (no
caller supplied Context.agent_team) and three PM verbs opted out
entirely — a misrouted frontend cell PM blocked, escalated, and held a
backend task through exactly that gap (live 2026-07-02). Org-wide roles
(main_pm, board, CEO, PR reviewer) are exempt so escalation handling
and root-PR gating keep working; cell-scoped roles are now enforced
wherever the caller supplies the team.
---------
Co-authored-by: Renn F <rennf93@users.noreply.github.com>
