Claude Haiku 5.5 (effort default (no --effort flag) · thinking on) on Fix an async job pool (retries, deadlock, error capture) with hidden tests

[
 {
  "duration_api_ms": 36626,
  "stop_reason": "end_turn",
  "session_id": "[REDACTED]",
  "total_cost_usd": 0.009038659999999999,
  "usage": {
   "input_tokens": 20,
   "cache_creation_input_tokens": 15897,
   "cache_read_input_tokens": 216726,
   "output_tokens": 7380,
   "output_tokens_details": {
    "thinking_tokens": 2528
   },
   "server_tool_use": {
    "web_search_requests": 0,
    "web_fetch_requests": 0
   },
   "service_tier": "standard",
   "cache_creation": {
    "ephemeral_1h_input_tokens": 15897,
    "ephemeral_5m_input_tokens": 0
   },
   "inference_geo": "global",
   "iterations": [
    {
     "input_tokens": 2,
     "output_tokens": 1014,
     "cache_read_input_tokens": 27563,
     "cache_creation_input_tokens": 873,
     "cache_creation": {
      "ephemeral_5m_input_tokens": 0,
      "ephemeral_1h_input_tokens": 873
     },
     "type": "message"
    }
   ],
   "speed": "standard",
   "fallback_credit": null
  },
  "modelUsage": {
   "claude-haiku-5-5": {
    "inputTokens": 20,
    "outputTokens": 7380,
    "cacheReadInputTokens": 216726,
    "cacheCreationInputTokens": 15897,
    "webSearchRequests": 0,
    "costUSD": 0.009038659999999999,
    "contextWindow": 1000000,
    "maxOutputTokens": 128000,
    "thinkingTokens": 2528,
    "canonicalModel": "claude-haiku-5-5",
    "provider": "firstParty",
    "costBasis": "list"
   }
  },
  "permission_denials": [],
  "terminal_reason": "completed",
  "fast_mode_state": "off",
  "fast_mode_disabled_reason": "sdk_opt_in_required",
  "subagent_stats": {
   "spawned": 0,
   "requested": {
    "background": 0,
    "foreground": 0,
    "unset": 0
   },
   "started_in_background": 0,
   "max_depth": 0,
   "spawned_by_subagents": 0,
   "completed": 0,
   "failed": 0,
   "killed": {
    "parent": 0,
    "user": 0,
    "system": 0
   },
   "refused": {
    "depth_limit": 0,
    "concurrency_limit": 0,
    "budget": 0
   },
   "by_type": {}
  },
  "safety_stops": 0,
  "is_error": false,
  "num_turns": 10,
  "subtype": "success",
  "api_error_status": null,
  "result": "All 19 tests pass, including the original two.\n\n## What was wrong and what changed\n\nAll changes are in `jobpool/__init__.py`. Tests are in `tests/test_regressions.py`, with the existing `tests/test_basic.py` left as is.\n\n1. **A failed job crashed the whole run.** `_run_one` re-raised the last exception, so `gather` propagated it out of `run()`, which the README says never raises. Failures now come back as exception instances in the result list.\n2. **The pool hung after a few failures.** The semaphore was acquired but only released on success, so each failure permanently consumed a slot. The semaphore is now held with `async with`, which releases it on every path.\n3. **Retry counts were off.** `range(self.retries)` gave `retries` total attempts, but the README defines `retries` as extra attempts. It now makes `retries + 1` attempts.\n4. **`retries=0` crashed with `UnboundLocalError`.** That was the `last` variable never being assigned when the loop ran zero times. The new loop always runs at least once, so `last` is always set.\n5. **The pool broke when reused.** The `asyncio.Semaphore` was created in `__init__` and bound to whichever event loop first contended on it, so a second `asyncio.run()` could fail. It is now created inside each `run()`.\n6. **`concurrency=0` hung forever.** The constructor now rejects `concurrency < 1` and `retries < 0` with `ValueError`. This is a small addition beyond the report, since it's the same hang with a different trigger.\n\nThe public API is unchanged: `JobPool(concurrency, retries)`, `run(jobs)`, and `stats`.\n\n## Choices to check\n\n- **`stats` accumulates across `run()` calls** rather than resetting per run. The README says it \"counts ok/failed jobs and total attempts,\" so I read that as lifetime totals. If you want per-run stats, that's a small change.\n- **No backoff between retries**, and a job keeps its concurrency slot while it retries. Neither was in the README, so I didn't add either.\n\n## Verification\n\n- `python` isn't on PATH in this environment, and `python3` has no pytest. I ran the tests with `./venv-sb/bin/python -m pytest -q`, which is the interpreter that has pytest installed. Result: 19 passed.\n- Before the fix, 17 of the 19 tests failed, each for the reported reason. The two that passed were the non-regression cases: the original ordering test and the concurrency-limit test.\n- The hang-related tests use a 5-second `asyncio.wait_for` so a regression fails instead of blocking the suite.\n\nI did not commit anything; the directory isn't a git repository.",
  "ttft_ms": 1234,
  "type": "result",
  "duration_ms": 39787,
  "uuid": "[REDACTED]",
  "ttft_stream_ms": 1075,
  "time_to_request_ms": 233,
  "first_content_frame_ms": 1076,
  "queued_turn_count": 0,
  "result_index": 0
 }
]