Start authorized file-only P08-03 implementation without shell permission changes

docs/plans/original-completion-2026-10-02/execution/p08-03/ACTUAL_OPUS_FILE_ONLY_RUN_DIRECTION.json

@@ -0,0 +1,7 @@
+{
+  "action": "run",
+  "task_id": "P08-03",
+  "reason": "P08-03 is ready and its registered contract has an empty depends_on list, so a safe, independent file-only path is available. The run must use the isolated, already-verified AGY 1.3.3 authentication and strict stream (/init/user_input confirmed) and keep the existing outer sandbox. AGY may use only file view/edit tools. Shell and run_command are forbidden entirely. That avoids the printf marker command waiting for owner approval: do not re-run, rephrase or route it through another tool, and do not treat this run as approving it. This run does not certify full acceptance of RP24-02 (0a7b0148) or production adoption of the 1.3.3 adapter. File-only instructions, limited to allowed_paths: (1) Before editing, read docs/spec/acceptance/P08-03.json, ACTUAL_OPUS_START_DIRECTION.json, CURRENT_ACCEPTANCE_AUDIT.json, autoload/simulation_runtime.gd (submit_command, demolish_facility, refusal codes, STAIR_LANDING_MISSING) and tests/transport/test_runtime_stairs.gd. Using DIAGNOSTIC_SCENARIO_SOURCE_AUDIT as reference, write CONDITION_MATRIX.md stating that preview validate_placement has no connection check and that STAIR_LANDING_MISSING is a pre-construction Runtime refusal. Keep these two oracles separate, so a preview rejection is never presented as proof of a Runtime connection rejection. (2) In scenes/main/main.gd and world_preview.gd, add a normal-palette demolition entry, mode, selection, confirmation, cancel and result/status display that call only SimulationRuntime.submit_command demolish_facility. Do not duplicate domain logic, mutate Runtime directly or touch Runtime/service/construction data. Keep build, excavation, pan, zoom, selection and modal behavior coherent. (3) Write diagnostic-tagged GDScript fixtures with stable UIDs under tests/fixtures/p08-03, choosing existing normally constructible definitions for support/connection and rebuild cases. Add one new strict marker entry to tests/ui/test_p08_normal_placement_ops.gd. (4) Write the Root driver tools/run_p08_normal_placement.py using existing execute_phase, the profile guard and owned_outer. (5) Write checker tools/check_p08_normal_placement_evidence.py with --verify-only reading BUILD_OUTPUT_DIR/p08-normal-placement-evidence/index.json; it must recompute hashes and state. (6) Write tools/test_p08_normal_placement_evidence.py with one whole positive case and semantic corruption negatives: reason mismatch, cell drift, stale source/input, fake OS provenance, partial demolition or orphan occupancy, bad resume, omitted members. (7) Only if the existing cleanup helper is used, add one source-SHA-pinned policy to suite-policy.json. (8) Update implementation.json and the completion marker contract. Because no Shell is allowed, the worker must record Python syntax, checker tests, Godot, native, OS and full regression all as unexecuted/pending. Running them is not the worker's job: Root runs local syntax and meaningful tests, guarded native, real-OS pan/zoom and the separate-PID resume. If an ordinary edit permission is refused, stop and report the exact denial without widening permissions or working around it. The original P08-03 conditions in full, 30 packages/88 units, the Windows open item and the need for independent review and integration are all preserved. Do not start a new Root or Codex duplicate implementation or a duplicate full regression.",
+  "codex_model": "gpt-6.1-sol",
+  "codex_model_reason": "The assigned worker is AGY, so the Codex model is only a fallback or for later repairs. The work is broad but well specified: UI wiring onto an existing Runtime API, fixtures, and a checker with defined corruption cases. No hard reasoning failure or repeated failed repair has been observed for this task. Execution is file-only, and Root owns native/OS verification. The default model gives enough quality at lower time and usage cost. Escalate to gpt-6-astra only if a submitted candidate shows a concrete semantic defect in the checker invariants or the two-oracle split that the default model fails to repair."
+}

docs/plans/original-completion-2026-10-02/execution/p08-03/CURRENT_EXECUTION.json

@@ -5,7 +5,7 @@
     "units": 88,
     "completed": 11
   },
-  "UI_impl_worker_started": false,
+  "UI_impl_worker_started": true,
   "necessary_prerequisite": {
     "task": "RP24-02",
     "state": "review_pending",
@@ -22,7 +22,18 @@
     "counting_scope": "infrastructure task, no original game-unit completion credit"
   },
   "Claude_auth": "existing native subscription client automatically renewed; actual authenticated request success, no user login wait or API fallback",
-  "P08_03_next": "registered ready, actual Opus scope confirmed wait; UI worker not started. Continue independent evidence preparation, no duplicate code or repeated game regression.",
+  "P08_03_next": "actual Opus run direction recorded, file-only AGY source implementation started in existing registered task; Root runs all tests",
   "current_game_sources_changed": false,
-  "permission_request": "single readonly printf command temporary, not applied"
+  "permission_request": "still pending and not applied; independent source work does not require marker command",
+  "UI_worker_method": "file view/edit only; no shell/run_command or denied marker action",
+  "RP24_acceptance_not_inferred": true,
+  "actual_UI_run": {
+    "id": "a33343b08feb43f1ae18e3adc8f8dfcf",
+    "state": "implementing",
+    "pid": 17973,
+    "session": 32391,
+    "service": "agy",
+    "method": "file-only; no shell/run_command",
+    "worker_count": 1
+  }
 }