[johnteee]

Summary

Why: after #97, four read-only decision packets re-checked the remaining "owner decision" items against code and the project's own records. Two of them were already decided:

• Deny presets: the owner chose "block in all modes" (docs/reviews/dogfood-g1-g22-2026-09-15-risk.md:22). docs/cli.md:923 and the permission playbook say the same. In prompt mode the code still offered a deny-matched call for interactive approval, so this is a defect, not an open question.
• Cockpit and background lifecycle: DR-006 already carves this work out under owner-override: co-maintainer dogfood (docs/strategy/dr-006-owner-decision-2026-06-22.md:35). That covers DU-011 and DU-014.

The packets also found that the WS2-002 residual was cheaper than I had reported. The runner already has a cancel_token (teaagent/runner/_core.py:167,907). The subagent batch just never set it.

What:

Action	Change	Gate	Test evidence
G-P1-4	POLICY_DENIED now goes to record_blocked and re-raises before any interactive approval (teaagent/runner/_governed_execution.py). approval why-denied describes policy_denied instead of "Unknown denial reason"	governance-gap	tests/test_deny_preset_runtime_enforcement.py. The handler is never called and the tool never runs. Falsifier: approval revoke <grant_id> lets the same approving handler run the tool, so revoke is the escape hatch
G-P1-5	Each subagent batch creates its own cancel_token, passes it to the children, and sets it at the deadline and on every exit. The parent's cancel still propagates. The runner re-checks the token before each tool dispatch (+3 lines in _core.py)	governance-gap	tests/test_subagent_batch_cancel_token.py. After the batch returned, 4 more tool calls ran before the fix and 0 after; the bound is at most 1 call already in flight. Cancellation is audited as the existing run_failed (system)
U-P1-7 (DU-011)	ControlCockpitState and the TUI cockpit show branch, undo availability, last run id and active rules. Each source falls back to an explicit unknown, and the failure is logged	owner-override (M4 carve-out)	tests/test_control_cockpit.py: 6 new tests, including no-git, no-runs and failure cases
U-P1-8 (DU-014)	Heartbeat ticks carry tool, tool_in_flight, elapsed_seconds, cost_cents and checkpoint through an allow-listed progress provider. heartbeat.py does not import runner internals	owner-override (M4 carve-out)	tests/test_heartbeat_progress_fields.py: 5 unit tests plus one real-run test
Docs	New action-register rows. Backlog: the four items above are done, and PHASE-01…06 plus DU-008/010/015/017/018/020 get PROPOSED verdicts (1 close-as-superseded, 11 holds; not ratified). Roadmap-status has a dated update. Acceptance count is 674	—	Docs consistency, OKF and aging checks pass

How: four worktree-isolated agents, each limited to its own files and to writing the test first. The coordinator reviewed every production diff, verified the agents' key claims against the code (the owner decision, the carve-out text, cancel_token wiring), cherry-picked the commits, and ran every CI lint and governance step plus the full suite locally.

Done: see below.

Not changed (noted for the owner)

• Protected-path codes: WORKSPACE_WRITE_MODE, SYMLINK_BLOCKED and SKILL_WRITE_BLOCKED still reach the interactive approval prompt for destructive tools in prompt mode. That looks intended, since prompt mode means "ask", but no owner decision on record pins it.
• DU-014 limits: checkpoint records the last completed call, because the runner has no "next checkpoint" concept. The liveness file still holds only the tick and the interval.
• DU-011 private import: the cockpit imports the private _grant_expired helper.
• Pre-existing failure: tests/test_hybrid_approval_queue_performance.py::test_large_scale_queue_performance fails on main too. It is a wall-clock assertion (22s against a 15s limit) and unrelated to this PR.

Action ID

Action: G-P1-4, G-P1-5, U-P1-7, U-P1-8 (registered in docs/retrospective/06-action-register.md in this PR)

Risk Class

Risk class: medium

This PR changes approval behavior in one direction only. A deny that already matched now always blocks, which is what the docs and the owner decision already promise. It also cancels subagent work at the batch deadline. The cockpit and heartbeat changes are additive observability.

Self-Review Checklist

General (all PRs)

• ruff check + ruff format --check pass (locked ruff 0.15.20)
• mypy teaagent/ tests/ --explicit-package-bases reports 0 issues
• python3 scripts/run_test_tier.py --tier smoke passes (inside the full suite)
• Coverage at least 75% (CI reported 79% on fix: governance-gap follow-ups (tick cost cap, batch timeout, deny usage error) + telemetry CI fix #97)
• check_root_module_count.py ≤ 184
• check_complexity.py ≤ 99 (62)
• No circular imports
• Event-spine wiring passes
• Docs consistency passes
• PR has Why / What / How / Done sections

High-Risk (if risk class = high)

• n/a (medium)

Documentation

• No contradiction with current-truth docs: this PR removes one, since prompt-mode deny override contradicted docs/cli.md:923
• Commands in docs are executable
• Stale dated docs updated or marked as superseded

Verification notes

• Full suite at 4d826103 (locked .venv): 6830 passed, 34 skipped, 0 failed.
• CI steps run locally: all 17 lint and governance-gate steps exit 0. That covers docstrings, config access, complexity, Any count, ruff, ruff format, mypy, bandit, test-quality severe and untyped, assertion regression, god modules, circular imports, event spine, root module count, selftest and tool lint.
• Commit gates: every commit passes .githooks/commit-msg, check_dr006_gate_trailer and check_action_register_link.
• Size: runner/_core.py grew from 1,062 to 1,065 lines; god-module exemptions are unchanged at 18.

---

Generated by Claude Code