KAP-990: suppress comment author self-triggers

[milymarkovic]

Summary

• suppress implicit squad-assignee wake when comment author is assigned squad leader
• make explicit @agent / @squad self-mentions loop-safe no-ops with honest outcomes
• preserve wake-up for comments from other agents and explicit delegation to different owners

Verification

• DATABASE_URL=.../kap990_test go test ./internal/handler -count=1
• go vet ./...
• independent pre-commit review: no blocking logic or security findings

Migration impact

None.

Closes KAP-990

[vercel]

@milymarkovic is attempting to deploy a commit to the IndexLabs Team on Vercel.

A member of the Team first needs to authorize it.

[multica-eve]

Thanks for working on the loop-safety problem. We can't merge this as proposed because it removes supported workflows where an assistant working on a child item explicitly wakes itself on the parent item, and where a squad leader who also completes a worker slice needs its result to resume coordination. Treating every same-identity mention or comment as a no-op changes those workflows without a replacement handoff mechanism. We need a narrower design that prevents accidental loops while retaining intentional cross-item handoffs, so we're not taking this PR forward in its current form.

[milymarkovic]

Addressed review with narrow, source-task-bound suppression in 2361968ec.

• Preserved explicit cross-item self-handoff when trusted X-Task-ID belongs to another issue.
• Preserved same-squad worker-slice → leader coordination resume when the exact authoring task is a worker task for that squad.
• Suppressed same-issue leader-run self-loops and same-issue explicit agent self-mentions.
• Removed dependence on “latest task role”; unrelated historical tasks can no longer change routing.
• Added/updated regressions for both intentional workflows and accidental-loop cases.

Verification: full go test ./internal/handler -count=1 PASS, go vet ./... PASS, GitHub CI all PASS. No migrations. Ready for re-review.

[milymarkovic]

@multica-eve re-review requested: the corrective DoD from your 2026-07-19 09:06Z review has been addressed on new head 2361968ec (pushed 09:44Z).

Summary of the redesign: blanket same-identity suppression replaced with narrow same-issue/same-context criteria; intentional cross-item self-handoff and squad-leader worker→coordination resume are preserved and covered by regression tests, accidental same-issue self-loops remain suppressed. CI green, MERGEABLE.

[milymarkovic]

Independent QA completed on exact head 2361968ec2042105ebbbd63f3f89a8adf213ee0c: PASS_WITH_NOTES.

Verified on isolated PostgreSQL test DB:

• same-issue self-trigger: 0 queued follow-up;
• same-issue explicit self-mention: suppressed/deferred, no duplicate;
• trusted cross-item self-handoff: exactly 1 expected run;
• same-squad worker→leader resume: exactly 1 coordination run;
• other-author routing and handoff to another owner: no regression;
• full go test ./internal/handler -count=1: PASS.

CI remains 5/5 PASS and PR merge state is clean. Requesting fresh upstream review of this exact SHA; no merge/release/deploy performed.

[apollion69]

Related data point from a self-hosted v0.4.10 deployment — the wake surface seems wider
than comment self-triggers:

• A plain top-level comment by a workspace member (no @-mention of any agent or squad)
  on a squad-assigned issue enqueued a new squad-leader task the same second the comment
  was created.
• Cancelling that task (issue cancel-task) was itself followed ~4 minutes later by
  another leader task being enqueued for the same issue, with no new comment or human
  action in between — i.e. lifecycle events also appear to re-wake the assignee.

With --max-concurrent-tasks 1 on the runtime this becomes a starvation loop: each
informational comment or cancel re-queues the leader, and scheduled/autopilot tasks wait
behind it. We had to unassign the squad from the issue to stop the cycle.

Suggestion: treat member comments without a mention, and task lifecycle events
(cancel/complete), as non-waking by default — or make the wake policy configurable per
issue/squad.