Preserve P18 native evidence and deliver recorded Opus repair direction

2026-10-13 14:34:06 +09:00
parent c77639a9d7
commit a0f556e135
15 changed files with 186 additions and 7 deletions
@@ -17,3 +17,5 @@ OWNER_RUNNER_CONTROL=CONTROL python3 -B owner_model_cli.py --project PROJECT dis
 New-project defaults are also set in the Runner policy. Existing frozen project configurations are not rewritten; the explicit adapter applies owner-authorized model selection at invocation. Bare legacy CLI/MCP dispatch does not provide this routing; use this entry point for future Codex invocations. Antigravity model is unchanged.

 Configuration and selected argv are distinct from actual inference evidence. Confirm the worker execution artifact after its next authorized run; no implementation is launched solely to test this setting.
+
+The same validated recorded run decision is also delivered in full to the Codex implementer prompt. Its concrete execution direction is subordinate to the immutable original spec, allowed paths and acceptance. Model choice alone is insufficient for task handoff. Check the actual run packet for the exact recorded direction before relying on delivery. This corrects the original model-only adapter; it does not change Runner permissions, registration, counters or acceptance.
@@ -31,6 +31,7 @@ def owner_model_llm(self, task_id, role, service, prompt, cwd, schema):
        chosen = decision.get('codex_model')
        if decision.get('task_id') != task_id or decision.get('action') != 'run' or chosen not in ALLOWED_CODEX_MODELS or not str(decision.get('codex_model_reason', '')).strip():
            raise ValueError('A recorded Claude run decision with explicit model and reason is required')
+        prompt = ("Recorded Claude commander run direction for this exact fixed task. Apply within the immutable spec, scope and acceptance; do not widen them.\n" + json.dumps(decision, ensure_ascii=False) + "\n" + prompt)
        self._owner_codex_model = chosen
        try:
            return original_llm(self, task_id, role, service, prompt, cwd, schema)
@@ -0,0 +1,10 @@
+{
+  "id": "13d58f6cc19f4257a93a6c44e250d81a",
+  "role": "implementer",
+  "service": "codex",
+  "status": "running",
+  "pid": 6833,
+  "artifact": "/Users/korizatou114/Library/Application Support/agent-control/the-tree-p18-01-model/runs/13d58f6cc19f4257a93a6c44e250d81a",
+  "argv_model": "gpt-6.1-sol",
+  "command_sha256": "2d35aba4a8cd6ea1fd22291214dfc86ad4516c8e2428b6ba2a6e4ae574617dab"
+}
@@ -0,0 +1,12 @@
+{
+  "id": "aa74e0e182094e83afe2dba1009b31b6",
+  "role": "implementer",
+  "service": "codex",
+  "status": "running",
+  "pid": 9620,
+  "artifact": "/Users/korizatou114/Library/Application Support/agent-control/the-tree-p18-01-model/runs/aa74e0e182094e83afe2dba1009b31b6",
+  "argv_model": "gpt-6-astra",
+  "direction_delivered": true,
+  "packet_sha256": "89983ae8e93a5244e8b89d004ae274d87a9f194278a382f38921feea51b96e48",
+  "command_sha256": "0d69dd08ea5c1219fbc7d12eef6814d95ff787f8dfa5a9acef73610fa8dbdb34"
+}
@@ -0,0 +1,39 @@
+{
+  "task": "P18-01-CODE",
+  "base_commit": "d2a1f80a613e7c77b02b941026bc49c2d85db863",
+  "existing_affected_suite_pins": {
+    "tests/s4/test_relationship_event_service.gd": {
+      "bytes": 32069,
+      "sha256": "ce1aa05716feed435cb68a1435a28db219cc3f056e23a442d2eaf08840771ffe",
+      "test_body_change_allowed": false
+    },
+    "tests/s4/test_runtime_relationship_events.gd": {
+      "bytes": 33566,
+      "sha256": "466f9a40b3471a9bb39c670cdd1106e990d297bc8bfff0db557cd63bb74c8ef3",
+      "test_body_change_allowed": false
+    },
+    "tests/s6/test_runtime_snapshot_serialization.gd": {
+      "bytes": 6753,
+      "sha256": "d9106f9defb6cc195b969ff040d24f1f8ceebe055ba66c39d4634a68620cde84",
+      "test_body_change_allowed": false
+    },
+    "tests/s6/test_runtime_save_lifecycle.gd": {
+      "bytes": 30502,
+      "sha256": "de94d588d6ce204aa8d0221c2989b8dd900cf4e36ca51f83b464518a382a887f",
+      "test_body_change_allowed": false
+    },
+    "tests/ui/test_staff_management.gd": {
+      "bytes": 35774,
+      "sha256": "7e37dc47583325a564ef48de31596adeeccfb7151ac5fce952efc6873c8ab704",
+      "test_body_change_allowed": false
+    }
+  },
+  "required_new_tests": [
+    "tests/s4/test_original_relationship_model.gd",
+    "tests/s4/test_original_relationship_runtime.gd"
+  ],
+  "reuse": "Test inputs and unchanged collector/ownership/diagnostic protocols reusable; changed relationship service and Runtime invalidate transfer of previous affected-suite execution success to new candidate. Run affected set on exact submitted source. Full required roster after code/UI fixes batched remains a P18-01 unit requirement.",
+  "previous_P24_candidate": "666c87bd7d7c961ae6d7d2ad8d52d6456c31015b",
+  "ordinary_UI_still_missing": true,
+  "formal_acceptance": false
+}
@@ -0,0 +1,9 @@
+{
+  "before": "Model was forwarded; concrete commander direction absent from Runner implementer packet.",
+  "after": "Exact recorded commander run decision and directions prepended to worker prompt. Immutable spec/scope/acceptance still authoritative.",
+  "controller_sha256": "e342904f012b14ad2856bd10bee03a227a151da88f6a9c9c891c564cfa253f1e",
+  "source_edits_by_root": false,
+  "policy_or_permission_change": false,
+  "repair_counter_reset": false,
+  "verification": "AST parse passed; actual aa74e0e182094e83afe2dba1009b31b6 packet.txt contains the exact recorded Opus decision and concrete repair direction (SHA in actual-collector-repair-dispatch-r02.json)."
+}
@@ -0,0 +1,64 @@
+{
+  "task": "P18-01-CODE",
+  "implementation_process": "b1931a8dcbc24d88bb3cbdce45140e02",
+  "implementation_exit": 124,
+  "retained_uncommitted_source_baseline": "d2a1f80a613e7c77b02b941026bc49c2d85db863",
+  "native_evidence": "/Users/korizatou114/Projects/とりにく工房ゲーム制作部/artifacts/original-completion-2026-10-02/p18-01-code-local-native-r01",
+  "native_ended_sha256": "00d85455f1c6745a1d0eb0c4e83d10db42bc11b9c3bd1491ef40b51069d7c712",
+  "native_input_sha256": "9c8f0720aa8293bb867516f86416b08840c9f7dff1ae17b93d94cb4cd4ea21cd",
+  "source_files": 1311,
+  "local_and_affected_suites": 7,
+  "checks": 843,
+  "passed": true,
+  "guard_errors": [],
+  "cleanup_errors": [],
+  "native_phases": [
+    [
+      "import",
+      5540,
+      true
+    ],
+    [
+      "parse",
+      5632,
+      true
+    ],
+    [
+      "000",
+      5662,
+      true
+    ],
+    [
+      "001",
+      5681,
+      true
+    ],
+    [
+      "002",
+      5708,
+      true
+    ],
+    [
+      "003",
+      5715,
+      true
+    ],
+    [
+      "004",
+      5742,
+      true
+    ],
+    [
+      "005",
+      5754,
+      true
+    ],
+    [
+      "006",
+      5810,
+      true
+    ]
+  ],
+  "required_repair": "Collector must strictly verify precollected Root-owned native evidence in registered validator sandbox, using same fixed argv/output, without child engines or player reads. Receipt mode triggered by existing output directory, fail closed for stale/missing/forged evidence. Native collection stays Root-owned; sandbox/caps/criteria unchanged. Collector change requires exact final-source revalidation; do not claim current tests whole-unit acceptance.",
+  "formal_acceptance": false
+}
@@ -0,0 +1,7 @@
+{
+  "action": "run",
+  "task_id": "P18-01-CODE",
+  "reason": "Task is repair_ready, with repairs 1/2 used and scope, spec, baseline and caps unchanged. Root's guarded native collection passed and found no game behavior defect: 7 tests, 843 checks, import PID5540, parse PID5632 (308 scripts), all 9 owned engines ended, and no guard or cleanup errors. The only blocker is the fixed validation argv. It runs collection inside Seatbelt, where player files and fresh Godot profiles are intentionally unreadable. This needs one narrow collector-only repair in tools/run_p18_relationship_tests.py, plus protocol notes under docs/plans/original-completion-2026-10-02/execution/p18-01-code/** only. The worker must continue the exact retained code, with no reimplementation. Requirements: (1) The default mode with the same fixed argv `python -B tools/run_p18_relationship_tests.py --output {build_dir}/p18-relationship-evidence` only verifies a preexisting Root-owned output. It starts no engines and does not read or write player or profile files. (2) A new explicit optional `--collect` flag is for Root-only fresh native collection under the unchanged guarded protocol. During collection it copies the required native raw logs into the output. (3) Verification checks exact current source-map content hashes and settings, never absolute worktree paths. It checks the engine, collector, helper and marker pins, and the exact 7-test roster. It also checks all 9 import/parse/native rows and all raw phase artifact hashes. It checks the strict markers, diagnostics, ownership, profile and cleanup proof, and that the default player data is fully equal before and after. (4) Fail closed on missing, stale, forged, incomplete or unknown receipts, on path escapes or symlinks, and on malformed flags. Never fabricate success. The worker may add verify-negative static checks and run static or prepare-only checks. It must not run engines or full regression, edit existing product code or tests, weaken guards, alter Runner, rights, state, caps or registration, or commit or push. It must submit an ordinary code-phase result promptly and never claim that the whole P18-01 unit is accepted. After this, Root runs the exact final source once, then formal validate, review and integration. The existing preflight in artifacts/p18-01-code-local-native-r01 is evidence of game behavior only, not acceptance of the changed collector.",
+  "codex_model": "gpt-6.1-sol",
+  "codex_model_reason": "This is the default model, and the remaining work is narrow and fully specified: one collector file plus docs, with no game logic, engine runs or product or test changes. The Astra worker already timed out (exit 124 at 900s), and this is the last repair attempt (2/2). Sol is faster and cheaper, so it is less likely to hit the timeout again. Astra's extra depth isn't needed here, because the verification rules are fixed in the instructions rather than left to design."
+}
@@ -0,0 +1,7 @@
+{
+  "action": "run",
+  "task_id": "P18-01-CODE",
+  "reason": "Resume existing candidate f76ad96990a4d7371b6ab667df0a068cc06be798 only. This is the recorded repair_ready repair. Repairs stay at 2/2, with no counter reset and no changes to caps, model policy, scope, spec or rights. Accepted game behavior source is the exact retained Astra preflight: Root actual guarded native import 5540 / parse 5632 all 308, 7/843 pass, no guards or cleanup errors, all 9 PIDs ended. Do not reconstruct or change any game or test source (scripts/relationship/*, autoload/simulation_runtime.gd, tests/s4/*, .uid files, suite-policy.json, completion_marker_contracts.json). Formal validator bc8bc48c on f76 failed with PermissionError (Operation not permitted: ps) in live_engines before any engine started. That shows collection must stay Root-owned and must not run inside Seatbelt. The earlier Sol run 13d58f6c only updated the handoff because the worker never received these instructions.\n\nWHAT TO CHANGE: only tools/run_p18_relationship_tests.py and protocol notes under docs/plans/original-completion-2026-10-02/execution/p18-01-code/.\n\n(1) DEFAULT MODE: fixed argv `python -B tools/run_p18_relationship_tests.py --output {build_dir}/p18-relationship-evidence`. It strictly verifies output that Root already finished collecting. It must not use subprocess or ps, start an engine, or read or write any player or profile outside the output directory. It must work in the ordinary registered sandbox. Missing, outdated, forged or incomplete input must make it reject with a nonzero exit. Never fabricate success.\n\n(2) `--collect`: Root-only fresh guarded native collection. Reuse tools/run_all_tests.py source_refs/execute_phase and the existing owned_outer/Guard/diagnostic/marker/logger contracts, a fresh private CoW source and owned profiles. Keep the existing exact native isolation, Guard, logger, process ownership and cleanup conditions exactly as they are. No raw Godot against the player profile. Run import, then all-script parse (expected parse count 308), then the 7 native tests: 2 new P18 tests, s4 relationship_event_service and runtime_relationship_events, s6 runtime_snapshot_serialization and runtime_save_lifecycle, and ui staff_management. Do not run a duplicate full regression. Write self-contained raw phase files (existing phase/engine-native.log etc.) with SHA-256 hashes recorded in the receipt.\n\n(3) `--prepare-only`: preview bytes, capacity and roster without starting any engine.\n\nVERIFIER MUST CHECK:\n- the current exact source-map content SHA and project settings SHA, independent of checkout path;\n- the collector, shared/helper, engine and marker pins;\n- exactly 7 unique tests and 9 import+parse+test rows, with all 308 parse expected;\n- full raw marker, diagnostic, phase, profile, ownership and cleanup files, each matching its recorded SHA;\n- complete default player before/after equality in the receipt;\n- every PID ended and no guard or cleanup error.\n\nConfine all paths to the supplied --output. Reject `..` escapes, absolute escapes, symlinks, duplicate or unknown or malformed flags, and mutually exclusive mode combinations. No source exemptions and no weakening for convenience.\n\nWORKER MAY: run AST/py_compile/static checks, --prepare-only, and negative verify checks against a temporary synthetic output directory (missing, tampered-hash, symlink, escape and incomplete cases must reject). The worker must NOT: run native or full tests or --collect, edit the Runner, change caps or model policy, commit, push or touch main, or read credentials. Report any environment denial with exact evidence. Write a short protocol note describing the modes, the receipt schema, and Root's next steps: after this collector change, Root reruns the exact final native --collect once into the task-allocation build_dir, then ordinary Runner validation (default verify mode), a separate Opus review, then integration. Finish quickly and return an ordinary submitted result. Original P18-01 UI rendering, approved initial values, full final roster and Windows final acceptance remain unmet; do not claim the unit complete.",
+  "codex_model": "gpt-6-astra",
+  "codex_model_reason": "This is the last recorded repair (2/2) and there is no counter reset, so a weak result can't be cheaply retried. The work is in a single file, but it is security- and provenance-sensitive. The worker has to split one collector into three modes (verify-only that works in the sandbox, Root-only collect, prepare-only) and keep the exact existing run_all_tests/Guard/owned_outer/marker contracts unchanged. On top of that it needs strict hash, pin and roster verification (7 tests, 9 rows, parse count 308, player equality) and path and symlink confinement. A subtle mistake here would either break the formal validator again or weaken isolation. The accepted game behavior source also came from Astra, so continuing with Astra keeps the context consistent. The extra usage and time are justified by the higher chance of passing on the first try; game source is out of scope, which keeps the run short."
+}
@@ -0,0 +1,6 @@
+{
+  "task": "P18-01-CODE",
+  "repairs": 1,
+  "max_repair_attempts": 2,
+  "reason": "Commander model rationale mistakenly says2/2; authoritative state after actual ordinary repair is1/2. Model decision retained unedited. No counter change."
+}
@@ -0,0 +1,8 @@
+{
+  "task": "P18-01-CODE",
+  "submitted": "f76ad96990a4d7371b6ab667df0a068cc06be798",
+  "review_actual_run": false,
+  "packet_ready": false,
+  "reason": "packetが大き過ぎます。タスクを小さく分割してください",
+  "needed_action": "Prepare scoped complete changed-function/transitive context with original full source/diff retained and hashes; keep500000byte guard and independent review criteria. No review passed or task state change."
+}
@@ -0,0 +1,9 @@
+{
+  "submitted_commit": "f76ad96990a4d7371b6ab667df0a068cc06be798",
+  "original_actual_native_evidence": "/Users/korizatou114/Projects/とりにく工房ゲーム制作部/artifacts/original-completion-2026-10-02/p18-01-code-local-native-r01",
+  "tested_files": 1311,
+  "submitted_files": 1311,
+  "mismatching_paths": [],
+  "settings_and_source_match": true,
+  "scope": "Exact candidatef76 original model and affected7/843 macOS native preflight. Not adopted, not whole-unit acceptance. Future collector repair requires new exact final-source evidence."
+}
@@ -30,19 +30,18 @@
    },
    "latest_registered_task_selection": {
      "task": "P18-01-CODE",
+      "phase": "collector-only repair2",
      "model": "gpt-6-astra",
      "decision_owner": "Claude Opus5.5",
-      "evidence": "docs/plans/original-completion-2026-10-02/execution/p18-01-contract-r01/opus-run-decision-r01.json",
-      "dispatch_evidence": "docs/plans/original-completion-2026-10-02/execution/p18-01-contract-r01/actual-model-dispatch-r01.json",
+      "evidence": "docs/plans/original-completion-2026-10-02/execution/p18-01-contract-r01/opus-collector-repair-decision-r02.json",
      "worker_state": "implementing",
-      "actual_argv_confirmed": true,
      "unit_acceptance": false
    }
  },
  "agy": "unchanged",
  "existing_project_override_controller": {
    "path": "/private/tmp/the-tree-owner-model-cli.py",
-    "sha256": "bca36b7129b80404ca2b719749e3795097f5c3b725fd9d1deba94bfd61bf3709",
+    "sha256": "e342904f012b14ad2856bd10bee03a227a151da88f6a9c9c891c564cfa253f1e",
    "persistent_path": "/Users/korizatou114/Projects/とりにく工房ゲーム制作部/docs/plans/original-completion-2026-10-02/execution/owner_model_cli.py"
  },
  "codex_source": "Direct human instruction: codexは6.1 solを使用して。必要であればastra。判断はclaude opus 5.5が行なう"
@@ -254,6 +254,6 @@
    "validate_facility_value_contracts.py": "91640e2a76cd6cc9186f7b85c0bf379b24b6cac0d48962815235ee628e7ae43e",
    "validate_plan.py": "23678227d9b3c556f1bd903f4b5b3147c4a82ce5085e10055481026803451b80",
    "validator_update_checks.json": "f997d0f3de2cea29550840b1d5cb23647d64f81329de46758551e2b6c8fecc27",
-    "work_packages.json": "dd68057aa4a00291dd9107c989983277568dcd8d5dc6e3560e89119e8583bf98"
+    "work_packages.json": "62b89096e0c8c8183a41e6c0ed3536f235a3a17b91b7955b702063501b94bfa8"
  }
 }
@@ -8503,7 +8503,7 @@
    },
    "legacy_state": "planned",
    "state_scope": "package-level; completed means whole package accepted; partial implementation is not completion",
-    "current_summary": "P18-01 code phase registered at base d2a1f80 and actually implementing with gpt-6-astra selected by Opus5.5. Original model/legacy raw preservation/Runtime save and read-model scoped; UI and initialization values remain unmet. No new delivery unit or acceptance.",
+    "current_summary": "P18-01 candidatef76ad969 preserves actual native7/843 and all308parse/source1311. Formal sandbox validation failed before engines on ps. First repair only documented because commander direction was missing in worker packet; coordinator handoff corrected. Collector-only repair2/2 now dispatched with Astra selected by Opus. Unit UI/initial values/final whole acceptance remain unmet.",
    "remaining_deliverables": [
      {
        "id": "P18-01",
@@ -8524,10 +8524,16 @@
            "sha256_at_integration": "e0b7094808385097fba664d9284d967a044d53fc8bd08e9d4a484f1f31d0032f",
            "scope": "Actual Opus contract/model proposal only, action wait; no implementation acceptance",
            "current_game_acceptance_inferred": false
+          },
+          {
+            "path": "docs/plans/original-completion-2026-10-02/execution/p18-01-contract-r01/native-preflight-r01.json",
+            "sha256_at_integration": "43b66545d88a4ac8406911e27a40b448efd89863a2eb5ec2ca554ff2dead5fe9",
+            "scope": "Retained uncommitted actual native source1311/model+Runtime+legacy affected7/843; preliminary code evidence only, collector changed later requires exact final revalidation",
+            "current_game_acceptance_inferred": false
          }
        ],
        "updated_at_jst": "2026-10-13T14:01:57.377446+09:00",
-        "progress_summary": "P18-01 code phase registered at base d2a1f80 and actually implementing with gpt-6-astra selected by Opus5.5. Original model/legacy raw preservation/Runtime save and read-model scoped; UI and initialization values remain unmet. No new delivery unit or acceptance.",
+        "progress_summary": "P18-01 candidatef76ad969 preserves actual native7/843 and all308parse/source1311. Formal sandbox validation failed before engines on ps. First repair only documented because commander direction was missing in worker packet; coordinator handoff corrected. Collector-only repair2/2 now dispatched with Astra selected by Opus. Unit UI/initial values/final whole acceptance remain unmet.",
        "remaining_gaps": [
          "Implement original three-axis meaning and protagonist Affection-bound shield maximum across model/UI/save.",
          "Preserve old generic five-axis facts without guessed migration; pin new/old/missing/unknown/corrupt save contract.",