[trissim]

P448: a failed goal turn has one block writer and retains its turn

Implementation checkpoint; not merged, installed or live-accepted.

• FailedGoalTurn.settle_launched writes the terminal's own failure reason and TurnId once. Deleted Done.observe_goal's failure block, EndedWithoutResult's extra block, and the now-unused permit/account blocking path.
• BlockedGoal owns verified-success supersession through its declared ActiveGoal successor. Only the same captured block, observed before a distinct successful productive turn, qualifies. Explicit goal-action blocks have no failed turn and keep explicit Retry.
• Verified settlement also handles fresh message turns without a goal launch permit; blocked goals otherwise never enter permit settlement. Same goal ID, incarnation and worktree are required. UUID ordering is not invented: the pre-turn captured state establishes chronology.
• Removed blank-success early blocking; unsuccessful/unproductive goal turns go through the same terminal failure writer.
• Completed the related family: goals created during a turn now use that same failure writer and carry the same TurnId. Deleted FailedGoalOrigin, Goals.block_failed_origin, GoalTurnPermit.fail_origin and both implementations, and the obsolete observe_goal callback/production invocation.

Source evidence

Before and after: refactor-audit Package.load parsed 427 production modules with no omissions. AST declarations and name/attribute consumers retained in owned scratch ast-before.json and ast-after.json (126 before / 90 after related sites; spelling set expanded after for new/deleted members, so these are source inventories, not a score). Relevant declarations/consumers: goal_states.py, goals.py, goal_management.py, goal_attempts.py, turn_goal_account.py, goal_turn_permit.py, turn_terminal.py, turn_progress.py, agent_events.py. Dynamic MroDispatch handler execution is confirmed by behavior checks, not claimed as AST-resolved. External scheduler launch refusal calls have no launched turn and remain explicit/non-recoverable.
Patterns: IDEN-5 (block detached from failed turn), IMPL-12 (three failure-block decisions), TIME-1 (removed superseded early block methods).
Whole-source required check: before 2, after 0:
grep -rnE 'Backend turn failed; inspect|Backend turn ended without a result; inspect' src | wc -l

Read-only live history (SQLite mode=ro, before editing)

Attribution is from the exact recorded reason and source writer, not a persisted stack trace. Turn chronology comes from retained turn files; goal history does not record the failure turn today.

All six recent backend blocks have Done.observe_goal's exact reason (Backend turn failed; inspect local diagnostics before resuming.); none has FailedGoalTurn.settle_launched's or EndedWithoutResult's reason:

• unit-checkrule history1352: preceding turn9 160f84d09fd24a5898af7be3b3b453ab; next turn10 d489fa29bc7aa51b65c926a6d3c11c68.
• unit-checkrule history1395: preceding turn14 6128762164df4c218772eb143537fcb1; next turn15 9f32d23ae35605a048430eade9deedd6.
• fork-boundaries history1353: preceding turn115 ef8730ea67e0463d9f633f962575536e; next turn116 a0f3eeeb14b51aedb240cf6eebf5fb69.
• fork-boundaries history1381: preceding turn119 a4dc0efa0aee4232ab9e2ee10cb70e62; next turn120 d94636a0974e44d21bbb55e2e48509c5.
• unit-runtime-root history1349: preceding turn17 0b89f6c08fac4b19b9da6ccb042f07ad; next turn18 64252b3c35be041ecf2ce76679be119c.
• unit-runtime-root history1374: preceding turn21 4c8805bbe469419cb29e2ee2111c8d87; next turn22 8149a7e47bffaadf620d169e0994b838.
  Older boundaries history1003 is an explicit model block about UNKNOWN completion (not any of the three writers); next turn11 8c5f6021f020a9ba4711f0a1e6900a67. History1136 is an administrative unresolved-attempt block, also not one of the three; next turn97 25326452b8f963c016d00d886c8841ea.

Verification

Three requested tests added to existing copied-bus test family. 6 passed in 6.19s, exit 0, using real registry/history/turn/attempt owners on isolated copied buses. Final obsolete-observer deletion rerun is recorded separately. The first test parametrizes original/origin goals and reported failure/missing result:

• test_failed_goal_turn_writes_one_block_naming_turn
• test_later_verified_turn_supersedes_failed_turn_block
• test_goal_action_block_survives_later_verified_turn
  Ran only these under background timeout 300, PYTHONPATH=src, shared interpreter /home/ts/code/projects/agent-comms/.venv/bin/python, -n 0 -o faulthandler_timeout=60 -q, owned --basetemp=/home/ts/.cache/agent-scratch/unit-goalblock/pytest-published. No ratchets/full suite run locally. Tests load real state through Comms after settlement; installed Toad/native message journey remains merger qualification.

Assumed:

• Verified progress requires a successful terminal plus assistant output or a successful tool; mere tool calls during a still-running/failed turn cannot clear a block.
• Existing blocks without a retained failed TurnId are not retrospectively attributed or cleared.
• Unlaunched scheduler refusals and explicit goal-action blocks remain retry-controlled.
• Recovery of displayed goal state does not authorize/replay the original failed/UNKNOWN input. The failed attempt ledger is retained. Merger must qualify this scheduling boundary and the actual installed Toad path; no installed acceptance claimed.
• Fresh installation qualification belongs to the merger; worker preserves live source, sessions and uncertain inputs.

[trissim]

Closed by claude-ai: P448 landed as #2529 (derived goal execution, owner stop wins over the backend's failure report) and #2564 (generation-owned evidence, report-only blocked work, no reissue). This earlier approach to the same problem is superseded.

---

Generated by Claude Code