{ "sha": "8c002ae7ee67846b9cd8fd7385195db10916915a", "message": "fix: reserve independent repair and time for final Luna fallback (#300)", "author": { "login": "nnneeeooo-nnneeeooo", "name": "", "email": "" }, "committer": { "login": "web-flow", "name": "", "email": "" }, "created_at": "2026-10-13T01:37:13Z", "files": [ { "filename": "pipeline/write.py", "previous_filename": "", "status": "modified", "additions": 61, "deletions": 8, "changes": 69 }, { "filename": "tests/test_write_offline.py", "previous_filename": "", "status": "modified", "additions": 101, "deletions": 0, "changes": 101 } ], "diff": "@@ -82,6 +82,7 @@\\n MAX_GROUPS_PER_RUN = 10\\n MAX_DRAFT_SECONDS_PER_RUN = 15 * 60\\n MAX_MODEL_ATTEMPT_SECONDS = 60\\n+PREFERRED_ATTEMPT_SECONDS_WITH_LUNA = 30\\n MAX_DRAFT_SECONDS_PER_GROUP = 120\\n PROVIDER_ROUTING_OVERHEAD_SECONDS = 5\\n TRANSIENT_PLATFORM_FAILURE_LIMIT = 2\\n@@ -1484,14 +1485,17 @@ def _normalize_reject_reason(candidate: dict) -> None:\\n f\"to establish a new event. Model detail: {reason}\")\\n \\n \\n-def _provider_attempt_allowance(provider, deadline: float | None = None) -> float:\\n+def _provider_attempt_allowance(provider, deadline: float | None = None,\\n+ max_attempt_seconds: float | None = None) -> float:\\n \"\"\"Return this provider's bounded call budget, clipped to its story time.\"\"\"\\n allowance = getattr(\\n provider, \"max_attempt_seconds\", MAX_MODEL_ATTEMPT_SECONDS)\\n if (isinstance(allowance, bool)\\n or not isinstance(allowance, (int, float))\\n or allowance <= 0):\\n allowance = MAX_MODEL_ATTEMPT_SECONDS\\n+ if max_attempt_seconds is not None:\\n+ allowance = min(allowance, max_attempt_seconds)\\n if deadline is not None:\\n allowance = min(allowance, deadline - time.monotonic())\\n return allowance\\n@@ -1507,9 +1511,34 @@ def _provider_group_budget(provider) -> float:\\n return budget\\n \\n \\n+def _is_luna_provider(provider) -> bool:\\n+ return (getattr(provider, \"name\", None) == \"openai\"\\n+ and getattr(provider, \"model\", None) == \"gpt-6-luna\")\\n+\\n+\\n+def _luna_reserve_seconds(providers) -> float:\\n+ \"\"\"Reserve one initial draft and one independent repair for final Luna.\"\"\"\\n+ return sum(2 * _provider_attempt_allowance(provider)\\n+ + 2 * PROVIDER_ROUTING_OVERHEAD_SECONDS\\n+ for provider in providers if _is_luna_provider(provider))\\n+\\n+\\n def _provider_chain_budget(providers) -> float:\\n \"\"\"Leave room for each preferred route to time out before GPT-6 Luna.\"\"\"\\n providers = tuple(providers)\\n+ luna_reserve = _luna_reserve_seconds(providers)\\n+ if luna_reserve:\\n+ preferred = [provider for provider in providers\\n+ if not _is_luna_provider(provider)]\\n+ # Existing primary/shared fallback repairs remain bounded, while\\n+ # neither can spend Luna's own repair time.\\n+ preferred_budget = sum(_provider_attempt_allowance(\\n+ provider, max_attempt_seconds=PREFERRED_ATTEMPT_SECONDS_WITH_LUNA)\\n+ for provider in preferred)\\n+ if preferred:\\n+ preferred_budget += 2 * PREFERRED_ATTEMPT_SECONDS_WITH_LUNA\\n+ return luna_reserve + preferred_budget + (\\n+ len(preferred) + 2) * PROVIDER_ROUTING_OVERHEAD_SECONDS\\n budget = max(\\n (_provider_group_budget(provider) for provider in providers),\\n default=MAX_DRAFT_SECONDS_PER_GROUP,\\n@@ -1527,7 +1556,8 @@ def _provider_chain_budget(providers) -> float:\\n \\n def _validated_draft(provider, group: dict, tries: int, ai_calls=None,\\n run_trace: _RunTrace | None = None,\\n- deadline: float | None = None):\\n+ deadline: float | None = None,\\n+ max_attempt_seconds: float | None = None):\\n \"\"\"Call `provider` up to `tries` times until a draft survives validation\\n and quote verification.\\n \\n@@ -1556,7 +1586,8 @@ def _validated_draft(provider, group: dict, tries: int, ai_calls=None,\\n previous_candidate)\\n if run_trace is not None:\\n run_trace.add_duration(\"promptAssembly\", prompt_started)\\n- allowance = _provider_attempt_allowance(provider, deadline)\\n+ allowance = _provider_attempt_allowance(\\n+ provider, deadline, max_attempt_seconds)\\n with draft_timeout(allowance):\\n candidate = provider.draft(\\n SYSTEM_PROMPT, user_prompt, DRAFT_SCHEMA)\\n@@ -2863,6 +2894,10 @@ def main() -> None:\\n else:\\n print(\"write: provider order: \"\\n + \" -> \".join(p.label for p in providers))\\n+ if _luna_reserve_seconds(providers):\\n+ print(\"write: Luna final fallback policy: preferred calls capped \"\\n+ f\"at {PREFERRED_ATTEMPT_SECONDS_WITH_LUNA}s; independent \"\\n+ \"draft/repair time reserved\")\\n try:\\n # Thin-story companion retrieval: excerpt-only items (e.g.\\n # Simple Flying) act as a discovery index and the pipeline\\n@@ -2962,6 +2997,11 @@ def main() -> None:\\n continue\\n if drafted >= MAX_GROUPS_PER_RUN:\\n break # LLM budget for this run is spent; the rest waits\\n+ if draft_deadline - time.monotonic() < _luna_reserve_seconds(alive):\\n+ print(\"write: insufficient run time for Luna draft and repair; \"\\n+ \"retaining remaining groups for retry\")\\n+ run_trace.finish(\"retry_pending\", \"retry_pending\")\\n+ break\\n drafted += 1\\n draft, writer, verified = None, None, None\\n final_model = None\\n@@ -2980,17 +3020,30 @@ def main() -> None:\\n # Removed mid-group by a platform-level auth/quota\\n # death earlier in this same iteration.\\n continue\\n- # One primary retry and one responsive-fallback repair per\\n- # group; transport failures do not spend the repair budget.\\n- tries = 2 if provider is providers[0] or not fallback_repair_used else 1\\n+ is_luna = _is_luna_provider(provider)\\n+ luna_reserve = _luna_reserve_seconds(alive)\\n+ provider_deadline = group_deadline\\n+ attempt_cap = None\\n+ if luna_reserve and not is_luna:\\n+ provider_deadline -= luna_reserve\\n+ attempt_cap = PREFERRED_ATTEMPT_SECONDS_WITH_LUNA\\n+ if time.monotonic() >= provider_deadline:\\n+ print(f\"write: {provider.label} skipped on group \"\\n+ f\"{group.get('id')}: preserving Luna draft/repair time\")\\n+ continue\\n+ # Luna has its own repair, independent of the shared repair\\n+ # for preferred fallbacks. Transport failures still fail over.\\n+ tries = 2 if (is_luna or provider is providers[0]\\n+ or not fallback_repair_used) else 1\\n try:\\n candidate, facts = _validated_draft(provider, group,\\n tries, ai_calls,\\n run_trace,\\n- deadline=group_deadline)\\n+ deadline=provider_deadline,\\n+ max_attempt_seconds=attempt_cap)\\n except DraftInvalid:\\n content_failed = True\\n- if provider is not providers[0]:\\n+ if provider is not providers[0] and not is_luna:\\n fallback_repair_used = True\\n _record_provider_response(\\n provider, transient_failure_counts)\\n@@ -16,6 +16,7 @@\\n import sys\\n import tempfile\\n import traceback\\n+from contextlib import contextmanager\\n from datetime import timedelta\\n from pathlib import Path\\n \\n@@ -975,6 +976,104 @@ def test_luna_is_used_after_preferred_service_failures():\\n print(\"test_luna_is_used_after_preferred_service_failures: done\")\\n \\n \\n+def test_luna_repairs_after_shared_fallback_repair_is_spent():\\n+ \"\"\"A responsive preferred fallback cannot consume Luna's repair.\"\"\"\\n+ for valid_repair in (True, False):\\n+ reset_data_dir()\\n+ pending = load(DATA / \"pending.json\")\\n+ group = pending[\"groups\"][1]\\n+ pending[\"groups\"] = [group]\\n+ common.save_json(DATA / \"pending.json\", pending)\\n+ bad = json.loads(json.dumps(DRAFT_BIZ))\\n+ bad[\"cat\"] = \"not-a-cat\"\\n+ short = json.loads(json.dumps(DRAFT_BIZ))\\n+ short[\"zh\"][\"body\"] = [\"短訊\", \"短訊\"]\\n+ fabricated = json.loads(json.dumps(DRAFT_BIZ))\\n+ fabricated[\"facts\"][0][\"sourceQuote\"] = \"This quote is fabricated.\"\\n+ primary = FakeProvider(\"gemini\", [providers.ProviderError(\"HTTP 503\")])\\n+ fallback = FakeProvider(\"wechat\", [bad])\\n+\\n+ class Luna(FakeProvider):\\n+ model = \"gpt-6-luna\"\\n+ max_attempt_seconds = 90\\n+ group_budget_seconds = 180\\n+\\n+ def draft(self, system_prompt, user_prompt, schema):\\n+ self.prompts.append(user_prompt)\\n+ return super().draft(system_prompt, user_prompt, schema)\\n+\\n+ luna = Luna(\"openai\", [short, DRAFT_BIZ if valid_repair else fabricated])\\n+ luna.prompts = []\\n+ original = write.build_providers\\n+ write.build_providers = lambda: [primary, fallback, luna]\\n+ try:\\n+ write.main()\\n+ finally:\\n+ write.build_providers = original\\n+ check(primary.calls == 1 and fallback.calls == 2 and luna.calls == 2,\\n+ \"Luna independently repairs after preferred fallback uses both calls\")\\n+ check(\"body does not meet\" in luna.prompts[1],\\n+ \"Luna repair receives the actual failed body validation\")\\n+ if valid_repair:\\n+ check(len(all_articles()) == 1\\n+ and all_articles()[0][\"writer\"] == \"openai:fake\",\\n+ \"a valid Luna repair publishes the formerly short draft\")\\n+ else:\\n+ check(not all_articles() and load(DATA / \"pending.json\")[\"groups\"],\\n+ \"fabricated Luna repair remains unpublished and retryable\")\\n+ check(common.norm_url(group[\"items\"][0][\"url\"])\\n+ not in load(DATA / \"seen.json\")[\"urls\"],\\n+ \"failed Luna repair does not consume the source\")\\n+\\n+\\n+def test_preferred_timeout_preserves_luna_and_run_budget():\\n+ \"\"\"A clipped run deadline still leaves two full Luna calls, then stops.\"\"\"\\n+ reset_data_dir()\\n+ pending = load(DATA / \"pending.json\")\\n+ pending[\"groups\"] = [pending[\"groups\"][1], pending[\"groups\"][0]]\\n+ common.save_json(DATA / \"pending.json\", pending)\\n+ clock, allowances, order = [0.0], [], []\\n+ bad = json.loads(json.dumps(DRAFT_BIZ))\\n+ bad[\"cat\"] = \"not-a-cat\"\\n+\\n+ class TimedProvider(FakeProvider):\\n+ def draft(self, *args):\\n+ order.append(self.name)\\n+ clock[0] += allowances[-1]\\n+ return super().draft(*args)\\n+\\n+ first = TimedProvider(\"gemini\", [providers.ProviderError(\"ReadTimeout\")])\\n+ skipped = TimedProvider(\"wechat\", [DRAFT_BIZ])\\n+ luna = TimedProvider(\"openai\", [bad, DRAFT_BIZ])\\n+ luna.model = \"gpt-6-luna\"\\n+ luna.max_attempt_seconds = 90\\n+ luna.group_budget_seconds = 180\\n+\\n+ @contextmanager\\n+ def fake_timeout(seconds):\\n+ allowances.append(seconds)\\n+ yield\\n+\\n+ original = (write.build_providers, write.time.monotonic,\\n+ write.draft_timeout, write.MAX_DRAFT_SECONDS_PER_RUN)\\n+ write.build_providers = lambda: [first, skipped, luna]\\n+ write.time.monotonic = lambda: clock[0]\\n+ write.draft_timeout = fake_timeout\\n+ write.MAX_DRAFT_SECONDS_PER_RUN = 220\\n+ try:\\n+ write.main()\\n+ finally:\\n+ (write.build_providers, write.time.monotonic,\\n+ write.draft_timeout, write.MAX_DRAFT_SECONDS_PER_RUN) = original\\n+ check(order == [\"gemini\", \"openai\", \"openai\"],\\n+ \"preferred routes yield to final Luna when its reserved time is reached\")\\n+ check(allowances == [30, 90, 90] and clock[0] == 210,\\n+ \"preferred request is capped and Luna gets two full calls within run cap\")\\n+ check(len(all_articles()) == 1\\n+ and len(load(DATA / \"pending.json\")[\"groups\"]) == 1,\\n+ \"completed article is saved; next story waits for a full fallback budget\")\\n+\\n+\\n def test_editorial_gate():\\n \"\"\"Model reject skips the group; fabricated quotes are dropped/blocking.\"\"\"\\n reset_data_dir()\\n@@ -1328,6 +1427,8 @@ def main():\\n test_model_chain_and_routing_policy,\\n test_provider_specific_time_budget,\\n test_luna_is_used_after_preferred_service_failures,\\n+ test_luna_repairs_after_shared_fallback_repair_is_spent,\\n+ test_preferred_timeout_preserves_luna_and_run_budget,\\n test_editorial_gate,\\n test_completeness_precheck,\\n test_major_title_only_is_retained_for_material_retry," }