[ViperJuice]

Summary

This follow-up fixes the four NITs from the delta re-check of #132. That re-check judged #132 MERGE-READY, with 0 blocking and 0 should-fix findings. #132 is merged as e193baf.

These fixes are meant to land before the 1.6.0 release PR. They change no version and release nothing.

Changed: mcp_server/storage/mcp_task_registry.py, mcp_server/cli/task_runtime.py (five lines for R4), the STDIO pre-check in mcp_server/cli/stdio_runner.py, CHANGELOG.md [Unreleased], tests/test_mcp2_stdio_wire.py, and a one-line test-isolation fix in tests/test_v13_pmcp_pilot.py (see below).

R1–R4 → fixes and tests

Nit	Fix	Tests (tests/test_mcp2_stdio_wire.py)
R1: a finished result could be lost before a woken tasks/result read it, either via ttl: 0 or via eviction right after the wake with 255 tasks running	Task TTLs have a one-minute floor (MIN_TASK_TTL_MS; effective_ttl_ms clamps to [1 min, 1 h]). Expiry and eviction never remove a finished task that still has a tasks/result waiter, or one that finished within the floor. Waiters are counted in wait_for_update, so a waiter woken by the completion but not yet scheduled is covered. Eviction takes the oldest evictable finished task, by finished_at. If every slot holds a running, awaited or just-finished task, the new task is refused rather than a result dropped.	test_ttl_zero_task_result_still_reaches_the_waiting_client (ttl: 0 end to end: echoed as 60000, and the parked tasks/result receives the payload) · test_task_ttl_is_clamped_and_defaulted[ttl 0 / ttl 1 → 60000] · test_finished_task_within_the_floor_survives_its_expiry · test_full_registry_keeps_a_just_finished_task_instead_of_evicting_it[with_waiter / without] (255 running plus one just finished; the reviewer's eviction-after-wake ordering: the new task is refused, and the result survives and is read) · test_eviction_takes_older_finished_tasks_before_recent_ones
R2: N3 answered malformed responses, echoing an id from the server's own request-id space	_malformed_line_reply returns a drop verdict for a line that fails the SDK's validator and is an object with result or error and no method. _stdio_input logs it at debug and sends no reply. Malformed requests are answered as before (-32700 / -32600).	test_malformed_stdio_lines_get_json_rpc_errors_and_valid_lines_pass now feeds {"id":11,"error":{"code":"x"}}, {"id":12,"result":1}, {"id":13,"error":{"message":"no code"}} and {"jsonrpc":"1.0","id":14,"result":{}}. None is answered or passed through; the -32700 and -32600 replies and the valid ping are unchanged.
R3: the CHANGELOG stated the N3 replies unconditionally	The [Unreleased] bullet now says "On the POSIX STDIO transport", and adds that Windows uses the SDK's own reader, which still drops malformed lines silently. It also describes the R2 drop.	tests/test_release_notes.py and tests/test_release_metadata.py pass, along with the 61 doc-contract files
R4: a capacity refusal reached the client as "details": "ValueError"	The registry raises TaskCapacityError (a ValueError subclass). TaskRuntime.run_task re-raises it as MCPError, which call_tool passes through, so the tools/call adapter returns its text as the isError result: "Too many tasks (N running or awaiting retrieval); retry after one finishes".	test_full_task_slots_return_the_actionable_message_through_call_tool (through the real call_tool and the handler path with a one-slot registry: the second task call gets the exact message, with no ValueError)

Two existing tests now pass explicit bounds, because the floor changes what they observe. test_task_retention_evicts_finished_tasks_and_refuses_when_all_run uses min_ttl_ms=0, since it tests eviction order rather than the floor, and expects TaskCapacityError. The max_ttl_ms=0 expiry test is unchanged, because the floor is clamped to the maximum.

Verification

dev0. Each affected file ran alone, serially, under nice, with no xdist, after the load and I/O-pressure check (1-minute load 1.0–3.3, I/O pressure ≈0):

File	Result
test_mcp2_stdio_wire.py	34 passed
test_mcptasks_reindex.py	8 passed
test_mcptasks_write_summaries.py	2 passed
docs/test_mcptasks_protocol_contract.py	4 passed
test_mcp_server_cli.py	18 passed
test_mcpauth_stdio_contract.py	4 passed
docs/test_mcpstruct_call_tool_contract.py	2 passed
smoke/test_mcpeval_sdk_surface.py	1 passed
smoke/test_mcpbase_stdio_smoke.py	2 passed
test_release_notes.py	4 passed
test_release_metadata.py	27 passed
doc-contract files	61 passed

black, isort and flake8 are clean. The mypy ratchet output is identical to origin/main.

Full suite on ai. It was run rather than skipped because the change touches task_runtime.py (five lines for R4) in addition to the registry, the STDIO pre-check and the CHANGELOG. Each run used the same isolation as #132: systemd-run --user with MemoryMax=16G and CPUQuota=400%, the shared Code-Index-MCP project lock, a fresh clone, an isolated HOME, a private prefetched grammar cache, MCP_REQUIRE_GRAMMARS=1, MCP_SKIP_PLUGIN_PREINDEX=true, uv sync --locked --extra dev --extra semantic (mcp 2.3.0), and pytest tests/ -n 2 --timeout=300.

• 93b7e9f: 1 failed, 4384 passed, 127 skipped. The failure was test_v13_pmcp_pilot.py::test_fixture_server_and_workers_get_the_prefetched_cache_with_downloads_denied, with ModuleNotFoundError: No module named 'v13_pilot_estimate'. This is a test-isolation bug in a PMCP pilot: pass a prefetched grammar cache to the fixture (offline acceptance blocked) #127 test added by feat: port to the mcp 2.x SDK; grammar prefetch retries; pilot grammar cache; serverInfo version (#124, #128, #127, #131) #132, not in this PR's change. create_fixture imports the pilot's sibling helper module, which resolves only once scripts/ is on sys.path, so the test passed only when an earlier test on the same xdist worker had added it. Running the test alone on dev0 reproduces it. Fixed in e33e639: the test prepends scripts/ with monkeypatch.syspath_prepend, as the file's other create_fixture test already does. It now passes alone (1 passed) and with its file (168 passed).
• e33e639 (final head): 4385 passed, 127 skipped, exit 0 in 15m16s (coverage 56.65%), unit ci-code-index-132nits-e33e639.

🤖 Generated with Claude Code

[chatgpt-codex-connector]

Codex Review Summary: ✅ Completed

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-11-09T23:12:04.515305Z	93b7e9f	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.