[ZachFara]

What changed

tools/eval/github_tasks.py gets 1 more verified GitHub task. After merging #169 and #170 the branch holds 9. Zach wants 20, so this is still short. I hit the turn budget and kept the bar where it was.

Task	Repo (license)	Upstream release and date	upstream_change
gh-table-extraction-bool-form	huss-huss/table-extraction-pro (MIT, 156 KB)	fastapi 0.147.0, 2026-10-19	fastapi-0.147-omitted-bool-form-false

• Pairs: 0. The new key has one task. Every sibling I found failed the bar (see the rejected candidates).
• Upstream change: 🐛 Treat omitted boolean form fields as False fastapi/fastapi#16523, "Treat omitted boolean form fields as False", merged 2026-10-19 as fastapi/fastapi@19c0b7a (authored and committed 2026-10-19). Its first commit was authored the same day. All of it is after the June 2026 cutoff. FastAPI 0.162.0 (2026-10-23) restored the declared default, so the task pins 0.147.0.
• Repo timing: issue include_metadata silently defaults to false with FastAPI 0.147 huss-huss/table-extraction-pro#6 opened 2026-10-19, after the release. Fix docs: accept cq field mapping and rename fields in DESIGN.md #7 merged the same day.
• Merge commit: base_commit is the merge's first parent, 16908b9.
• Fix: it changes code. The three routes take include_metadata and use_fallback as optional strings and parse them, applying the default themselves.
• Setup: uv pip install --exclude-newer=... -c requirements/ci-constraints.txt -e .[dev], then the fastapi==0.147.0 pin.
  • These are the project's own CI constraints. Without them PyMuPDF 2.1 pulls in onnxruntime and the project grows to 353 MB. With them it is about 226 MB, .venv included.
  • The constraints pin fastapi 0.146.0. The pin step installs 0.147.0 over it.
• Laid over: only tests/test_jobs.py. The fix's tests/test_api.py has a comment stating the cause ("FastAPI 0.147 reads a missing bool = Form(True) as False"), so it is left out.
• Hard-set: not claimed. The error doesn't name its fix, but I didn't test whether the obvious fix fails.

Base failing output (containerized)

Run	Result
Base, fastapi 0.147.0	2 failed, 26 passed
Fix, fastapi 0.147.0	28 passed
Base, fastapi 0.146.0	28 passed (previous version, checked)

• The two failures at the base are the fix's new test_job_boolean_form_fields_default_when_omitted and the existing test_result_formats. The fix changed test_result_formats to stop sending include_metadata="true".
• Key line:

  >       assert manager.get(job_id).include_metadata is True
  E       AssertionError: assert False is True
  

• Runs went through a .scratch/ helper that calls repos.materialize, repos.prepare and repos.run_tests with container.run, the same steps agent_eval.py check runs. Containers ran one at a time, none were left behind, and --on-host was never used.

Cause-word checks

• Cause words:
  • omitted bool form fields as false, omitted bool form field as false
  • omitted boolean form fields as false, omitted boolean form field as false
  • html checkbox, unchecked checkbox, checkbox semantics
  • 0.147
• Every phrase pairs FastAPI's object, its bool form field, with its new rule: an omitted field is read as False, the "unchecked HTML checkboxes" of the release notes.
• Dropped on review:
  • bool form field names the object but not the change. A symptom-only fix says it ("make the bool form field Optional"), and word-start matching also catches "bool form fields".
  • An earlier draft's bare checkbox and unchecked were dropped for the same reason.
• Left out:
  • boolean form field: it hits the laid-over test names.
  • default: a fix that only treats the symptom says it.
• I checked each word with grading v2's mentions against:
  • the whole captured base output;
  • the whole laid-over tests/test_jobs.py at the fix sha;
  • the symptom words (include_metadata, metadata), in both directions;
  • the names of the functions the fix edits or adds (extract, extract_batch, create_job, parse_form_bool), the failing tests and the traceback's last call (get);
  • the base source near the fix site: table_extraction/api/main.py and jobs.py, plus README.md and tests/test_api.py (git show 16908b9:<path> from the repo cache).
• No hits. As a control, default hits every source, as expected.
• 0.147 is the only version word. The other seven are non-version words.
• Leak review: I read the laid-over file at the fix sha.
  • The new test's comment states only the symptom: "Regression for docs(decisions): propose cq knowledge unit field mapping #6: an omitted include_metadata must mean true, use_fallback false."
  • It does point at the issue, which states the cause. The record's comment says so, and notes that eval sessions have no web tools.
  • No other comment, docstring or test name mentions FastAPI versions, checkboxes or form parsing.
• New version word: 0.147. It is added to the expected set in test_the_cause_words_grader_v2_skips_are_exactly_the_versions.

Merge

I merged origin/feature/eval-setup twice, after #169 and after #170. The TASKS conflict kept both sides. The version set merged cleanly: 2.0 was already in it.

Rejected candidates

This batch:

• CI rot: Upload ohne Formfeld schreibt mehrsprachig=False (fastapi 0.148 / starlette 1.9) napoleonmm83/Transkribor#687 (fastapi 0.147 bool form): the only test file's new test docstring states the fix ("the field arrives as text and is evaluated itself"), and the new test likely passes at the base.
• fix(doc): keep /parse's default-on flags on under FastAPI 0.147+ ReadyForAI/MantisFetch#335 (fastapi 0.147): the new test file's docstring states the cause. The base's requirements.txt comment states it too.
• API server: with FastAPI 0.147+, a /generate request without streaming no longer streams mstar-project/mstar#713 (fastapi 0.147): the repo is 690 MB.
• FastAPI 0.147 reads an omitted bool form field as False, so some re-tags skip artwork randomphrase/harmonist#808 (fastapi 0.147): the fix PR #810 changes no test.
• Hold a message on a bare POST to its hold page again (FastAPI 0.147) vinitrami-Soc/phishhawk#78 (fastapi 0.147): the fix changes no test.
• fix(stt): omitted bool form fields keep their default; FastAPI 0.162.0 SteffenHebestreit/tts-stt-playground#119 (fastapi 0.147): the repo has no license.
• fix(ui): keep approval and enabled flags unchanged on a one-field op edit evoila/meho#5013 (fastapi 0.147): a large platform backend with database-backed tests; not pursued.
• fix(ingest): keep bot import a dry run when dry_run is omitted bl0rb/Weave#377 (fastapi 0.147): license NOASSERTION, an OCR monorepo.
• Fix main CI: contract-note upload applies by default again under FastAPI 0.148 abhishekwasu-hue/AdvanceTradingPlatform#265 (fastapi 0.148): 519 MB, license NOASSERTION.
• validate default format f so tiles and legend work on fastAPI >= 0.148 earth-mover/xpublish-tiles#345 (fastapi 0.148): the fix changes no test.
• cancervariants therapy-normalization#499, gene-normalization#453, disease-normalization#286 (fastapi 0.145 removed **extra): the fixes change no test.
• Unit suite fails with FastAPI >= 0.155 (Prefect 3.7.8: Body() got an unexpected keyword argument 'min_items') dathere/datapusher-plus#353 (fastapi 0.155, via Prefect): the fix is only a pin.
• UI disconnects with ASGI 'http.response.start' error when Starlette resolves to 1.8.0 PovertyAction/datasure#330 (Starlette 1.8 GZip): the fix is only a pin.
• fix(api): 재시도 미들웨어의 가짜 연결 끊김 — 새로 설치하면 웹 화면이 0바이트 jeongeundev/OpenArchive#283 (Starlette 1.8 FileResponse watches for disconnects): the new test's docstring states the cause and the fix.
• fix: stop file downloads 500ing under gzip-accepting clients on Starlette 1.8 goalsgame/omnigent#68 (Starlette 1.8 GZip): the repo is 296 MB.
• Fix main's unit tests: declare markdown, edge path and webhook test on Starlette 1.9 hubzoid/hubzoid#79 (Starlette 1.9 path escaping and TestClient): the fix mixes a missing dependency with test-harness changes.
• fix: keep ApiError mutable so Starlette can reset its traceback CARI-DAAL/ascend-op-evaluation#75 (Starlette 1.10 traceback reset): no license, no test change.
• fix(wedding): הזמנות מחו״ל page 500 under Starlette 1.7 josefcohen96/expense_tracker#44 (Starlette 1.7 TemplateResponse): no license, no test change, and the change came before the cutoff.
• [Bug] Starlette 1.7.0: public run-capacity slot is released after the streamed response is delivered (test_encoded_sse_cannot_bypass_public_error_inspection fails intermittently) agno-agi/agno#10515 and #11160 (Starlette 1.7, fastapi 0.147): a very large repo.
• [BUG] ASGI lifespan: _Lifespan never closes its anyio streams; with filterwarnings = error an unrelated test fails with ResourceWarning: Unclosed <MemoryObjectReceiveStream> schemathesis/schemathesis#5085 (anyio streams in its own lifespan): a schemathesis change, not an upstream one; 139 MB.
• Test suite aborts at collection (exit 2) under anyio >= 4.15: BlockingPortal DeprecationWarning promoted to error Guard-Core/guard-agent#55 (anyio 4.15 BlockingPortal deprecation): the fix is only a pytest config change.
• AttributeError with urllib3 2.8.0: 'Urllib3UtilUrlPercentReOverride' has no attribute 'sub' streamlink/streamlink#7116 (urllib3 2.8): the fix changes no test.
• public_transport="curl" breaks with urllib3 2.8.0 (curl-adapter incompatibility) — every login fails at the encryption-key fetch subzeroid/aiograpi#473 (urllib3 2.8 via curl-adapter): the fix is a dependency floor, plus tests.
• Configured relax-policies lists are coerced to strings by Click datnguye/dbterd#161 (click lists coerced to strings): not caused by a click release. click.STRING always coerces a default_map list.
• Fix --no-verify not disabling SSL verification korewayume/httpx#1 (click 8.5 flag_value for --no-verify): the fix changes no test.
• Fix CLI error output with Click 8.5 kedro-org/kedro#5847 (click 8.5 error output): license NOASSERTION, 251 MB, no test change.
• Fix daily pipeline BOOLEAN insert (Click 8.5 × Typer bool flags) MKSaber/ai-trader#8 (click 8.5 with Typer bool flags): no license.
• Support setuptools 84.2 read-only abi_tag karellen/wheel-axle#42 (setuptools 84.2 read-only abi_tag): the new tests' docstrings state the cause, and the AttributeError names it.
• Markdown 3.11 c2corg/v6_api#2359 (Markdown 3.11): the fix is only a version bump.
• fix(webhost): rewrite doc image URLs on mistune 3.4 MultiworldGG/MultiworldGG-Beta#196 (mistune 3.4): the repo is 815 MB.
• Callout plugin crashes on mistune 3.3+ (too many values to unpack) ediblesites/wp-poster#42 and fix(hosted): serve /health because Cloud Run reserves /healthz #43 (mistune 3.3): released before the cutoff.
• CI red on main: browser coverage test times out on "Approve & move" with fresh dependencies PhilippMundhenk/photosort#31 (anyio 4.16 and others): the fix is only constraints.

From the brief and PRs #164, #165, #167 and #168: not re-checked.

Unfinished: none. The uncommitted change in the earlier run was a comment on this task (its measured size and the constraints note), not a second task. It is committed with the task.

Why

Ticket: MIT-618 (batch 8), the GitHub-task eval. These are real post-cutoff tasks for the arms run.

How it was tested

• Containerized runs of the new task: the base and the fix on fastapi 0.147.0, and the base on 0.146.0, as described above.
• server/tests/test_eval_grader.py and server/tests/test_eval_repos.py pass.
• make check passes after both merges of origin/feature/eval-setup (1532 passed).

Learned

• Merged-PR search (is:pr is:merged, a quoted release string) finds far more code fixes than the closed-issue search with linked:pr. For fastapi 0.147 it found about 20 fix PRs where issue search found 4. Many of them never had an issue.
• FastAPI shipped a run of behaviour changes from 0.147 to 0.162 (2026-10-19 to 2026-10-23): omitted bool form fields as False (reverted in 0.162), defaults no longer validated (0.148), empty form strings as missing (0.149), unselected files as missing (0.150), JSON null versus omitted (0.151), header names in errors (0.152) and the removal of **extra (0.145, 0.155). Downstream fixes for 0.148 and later are still rare.
• When an install is over the 200 MB cap, check whether the project ships a CI constraints file. A newer major of a transitive dependency (PyMuPDF 2.1 pulls in onnxruntime) can double the install, and the project's own constraints bring it back.
• Fixes written by coding agents often carry test docstrings that state the cause. Read the laid-over test files before anything else, and lay over a sibling file when one file leaks.
• git --git-dir <cache> is refused in worktree shells. Read cached repos through a .scratch/ helper run with python3 -I.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Qc1CQTb8r67J32HpkhS43y