fix: restore bugfixes silently dropped by the v0.5.0 release commit

[atulya-singh]

I went looking for the cause of one reverted fix (#234, restored in #304) and found it wasn't a one-off. The v0.5.0 release commit reverted most of a month of merged PRs.

What happened

12d3fd8 ("release: AutoResearchClaw v0.5.0") is a single squashed commit — 318 files, +42,268/−8,632 — sitting directly on top of 84dad0a. Judging by what survived and what didn't, the v0.5.0 work was branched around 2026-04-09/10 and landed on 05-19, overwriting everything merged in between. Nothing in the commit mentions reverting anything.

The date cutoff is unusually clean:

merged	commit	fate
04-06	ad806b5 MiniMax preset URL / from-stage	intact
04-06	28478cc MCP path traversal, dispatcher failover	intact
04-07	5dc7fcc Responses API / MiniMax 404	intact
04-08	be89838 HITL collab fixes	intact
04-10	afccebe scripted HITL adapter	intact
04-10	e6a6dfa perf: lift regex compilation (#231)	reverted
04-10	7b57457 --from-stage without --output (#217)	fix kept, 212-line test file deleted
04-10	684ddb2 Gemini native adapter (#227)	file deleted
04-11	8ec230d retry error tracking (#234)	reverted (→ #304)
04-20	9f8d11d doctor gateway probe	reverted
04-20	61a92d8 Volcengine/BytePlus presets (#240)	reverted
04-20	28508f2 SOTA hallucination (#239 / fixes #238)	reverted
04-21	9f15247 search plan schema variants (#243)	reverted

Everything ≤04-08 survives; almost everything ≥04-10 is gone. 61a92d8 is the sharpest case — it came in via 84dad0a, the release commit's own parent, and was wiped by its own child.

I verified each of these against HEAD rather than trusting a diff heuristic — the refactor moved prompts.py into a prompts/ package, so "file gone" needed checking by hand. 5dc7fcc looked lost by line-matching but its content genuinely moved into prompts/ml.py; it's fine.

What this PR restores

Three correctness fixes whose removal is unambiguous, each mapping to an issue that's closed but reproducible again at HEAD:

Doctor gateway probe — health.py:177 was back to method="HEAD" on /models, byte-for-byte pre-fix. Gateways that reject HEAD but serve GET normally get reported as unreachable, which is what #292 and #241 were about.

SOTA hallucination (#238) — this one is the clearest evidence of an accident rather than a decision. The release didn't just drop the fix, it restored the exact prompt line the fix removed:

# researchclaw/prompts/ml.py:186 at HEAD — the line #239 deleted
"- State whether SOTA results exist on this benchmark and what they are.\n"
"- Add a 'Benchmark' subsection listing: name, source, metrics, "
"current SOTA (if known)."

The goal.md disclaimer is gone too. Notably prompts/ml.py:182 still has the anti-fabrication line from 5dc7fcc (04-07) while line 186 has the pre-#239 text (04-20) — the file is a snapshot from between those two dates, which is exactly the squash artifact you'd expect.

Search plan schema variants (#243) — _literature.py was back to plan.get("search_strategies", []), so a plan returned as search_phases / phases, or with dict-wrapped queries, silently produces zero queries.

What this PR deliberately does not restore

These may have been removed on purpose, and that's your call, not mine:

• Gemini native adapter (Gemini Native API Adapter #227) — llm/gemini_adapter.py is deleted and no gemini preset remains. Cleanly removed, no dangling imports, so this could well be intentional.
• Volcengine/BytePlus presets (LLM: add Volcengine and BytePlus presets #240) — gone from PROVIDER_PRESETS entirely (HEAD has: anthropic, deepseek, kimi-anthropic, minimax, novita, openai, openai-compatible, openrouter), and tests/test_volcano_provider.py was deleted.
• e6a6dfa (perf: lift repeated regex compilations to module level #231) — the regex hoisting is undone (_helpers.py:527 compiles per-call again). Perf only, and a separate concern.
• 7b57457 (fix: --from-stage without --output now finds existing run directory #217) — the fix survives, but its 212-line tests/test_from_stage_run_dir.py was deleted, so it's now unprotected.

Happy to send any of those as follow-ups if you confirm they weren't intentional.

Testing

pytest tests/ → 2799 passed, 56 skipped (2797 on main + the 2 tests added here).

#239's tests were dropped by the release too, so they're restored: the test_rc_prompts.py guard, and the test_rc_executor.py assertion that had to change because goal.md now leads with the disclaimer. I added a probe-method test for the doctor fix. I checked the new tests actually fail against un-restored source rather than passing vacuously:

FAILED tests/test_rc_health.py::test_check_llm_connectivity_probes_with_get_not_head
FAILED tests/test_rc_prompts.py::TestPromptManagerDefaults::test_topic_init_no_sota_hallucination

#243 has no test and I didn't add one. Its parsing sits inline in _execute_search_strategy, which needs run-dir artifacts, adapters and an LLM path to drive, so any test I wrote would be mostly mock scaffolding. Worth noting the original also shipped without tests — which is a large part of why it disappeared unnoticed for three months. If you want it covered, extracting the plan-parsing into a helper would make it a one-liner to test; happy to do that separately.

Worth considering

The single most useful change here isn't in the diff: whatever produced 12d3fd8 should rebase rather than squash onto main, or at minimum diff against main before landing. Five merged community PRs were reverted with no notice, and the only reason any of it surfaced is that I chased an unrelated 429 error message. I'd suggest a CI check that fails when a PR deletes a test file it didn't add — that alone would have caught #217, #240 and #227.

I'd also hold off closing #238, #241, #292 and #243 as fixed until this (or an equivalent) lands, since all four currently reproduce on main.

[Jiaaqiliu]

Merged. This is the most valuable report in the queue, and the finding is confirmed independently rather than taken on trust.

Verification

12d3fd80 is a single squashed commit, 318 files, +42,268/−8,632. The sharpest case you named checks out exactly as described — 61a92d8 (Volcengine/BytePlus presets, #240) came in via 84dad0a4, the release commit's own parent, and was wiped by its own child:

parent 84dad0a4  researchclaw/llm/__init__.py : 4 volcengine references
HEAD  (pre-merge) researchclaw/llm/__init__.py : 0

tests/test_volcano_provider.py shows in the release commit's own stat block as 165 --. A release commit deleting a test file it never wrote is not a judgement call about scope; it's a squash artifact.

The three fixes this PR restores also reproduced at HEAD before merge: health.py back to method="HEAD", prompts/ml.py:186 carrying the exact SOTA line #239 removed, and _literature.py back to a bare plan.get("search_strategies", []).

One merge conflict, resolved by keeping both sides

_literature.py conflicted: main had grown handling for a query_strategies.<sub_question>.{boolean_seeds, queries} schema, while this PR adds dict-wrapped queries and a top-level queries list. Those are complementary — different schema variants, not competing implementations — so I unified them behind one helper rather than picking a winner:

def _collect_queries(raw: Any, out: list[str]) -> None:
    """Append query strings from raw into out."""

It accepts plain strings and dict-wrapped entries, ignores non-lists and junk, and all three call sites (search_strategies, query_strategies, top-level queries) now go through it. Behaviour check:

plain strings      -> ['a', 'b']
dict-wrapped       -> ['x', 'y']
mixed + junk       -> ['ok', 'z']
non-list input     -> []
None               -> []

That also answers your note about #243 being untestable where it sat — the parsing is now a free function, so it's a one-liner to cover. Worth a follow-up.

On the process point

Your closing observation is the part worth acting on, and I'd rather say so plainly than let it sit under a merged PR: five merged community PRs were reverted with no notice, and the only reason any of it surfaced is that you chased an unrelated 429 message. That's a real cost to people who contributed in good faith. The CI check you propose — fail a PR that deletes a test file it didn't add — would have caught #217, #240 and #227, and is cheap to add.

Not restored, as you left them

Gemini native adapter (#227), Volcengine/BytePlus presets (#240), and the e6a6dfa regex hoisting (#231) are still out, along with the deleted tests/test_from_stage_run_dir.py. You were right to leave those as a maintainer call. Given the above, none of the three look intentional to me either — follow-up PRs restoring them would be welcome, test file first.

Post-merge: 2956 passed (baseline before this batch: 2869), no regressions.