技能 2026-05-06 13:59 UTC
autodecision
由提交人 harshilmathur 提交,主要更新类型为:问题缺陷修复与稳定性提升。提交说明:fix: CLARIFY follow-ups — Sources table format, schema description, parse validation (#14)。解决了已知运行时异常与边界错误,提升系统稳定性。
Commit提交信息原文
fix: CLARIFY follow-ups — Sources table format, schema description, parse validation (#14)
Three small fixes flagged during PR #11 review and folded into a single
post-merge commit.
1. output-format.md Sources table example + "Required columns" footer were
describing the ELICIT-only format ("user-inputs.md, provided Apr 2026"),
while decide.md Step 4b instructs the brief writer to use
"Source: ELICIT Q1 / CLARIFY Q1". The brief writer had conflicting
instructions across the two files. Updated the example with both an
ELICIT-sourced row and a CLARIFY-sourced row, and rewrote the footer
so the convention is documented in one place.
2. brief-schema.json content_check description for unsourced_dollar_figure
still said "[U#] = user-inputs.md claim" — describing the tag as
ELICIT-only. The validator regex itself already accepts [U#] regardless
of origin, so no behavior change, but the description string appears
in HARD_FAIL error messages, so a contributor debugging a CLARIFY-cited
number would have seen misleading guidance. Updated to mention both
user-inputs.md and iteration-1/clarify-answers.json with a note that
the Sources column disambiguates origin.
3. clarify.md Step 1.5 fallback (orchestrator writes the questions file
when the subagent returned JSON inline instead of calling the Write
tool) was writing the subagent's payload verbatim with no parse step.
Malformed inline JSON (trailing comma, bad escape, partial output)
would have produced a broken clarify-questions.json that crashes
Step 2's read. Tightened to parse first, fall through to
clarifier_failed on parse error so Step 2 always reads a parseable
file or sees a typed skip marker.
The fourth item flagged in review (decide.md Step 4b "n = ELICIT count"
not deterministically defined) is not addressed here — it's a theoretical
reproducibility concern about the LLM-driven brief writer's interpretation
of user-inputs.md, the [U#] tag numbers are run-local so cross-brief
inconsistency affects nothing observable. Worth revisiting only if a real
miscount surfaces.
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>