[refactor] Sync diagnosis-detection narrowing to sister skills idd-all/idd-list/idd-update (verify follow-up from #53)

[kiki830621]

Problem

From verification of #53 (PR #58):
「Sister substring scans 仍存在 — idd-all/SKILL.md:441,524 (Python substring), idd-list/SKILL.md:115, idd-update/SKILL.md:120 用 ## Diagnosis 做 phase 推斷,同樣 false-positive 風險」
— Source: team:regression + codex

#53 narrow fix 只動 idd-all-chain/SKILL.md。Sister skill 仍有同類 false-positive risk:

• plugins/issue-driven-dev/skills/idd-all/SKILL.md:441 — Python substring if "## Diagnosis" in comment_body
• plugins/issue-driven-dev/skills/idd-all/SKILL.md:524 — same pattern
• plugins/issue-driven-dev/skills/idd-list/SKILL.md:115 — 掃 comments 的標題推斷(prose-level,但底層可能 substring)
• plugins/issue-driven-dev/skills/idd-update/SKILL.md:120 — 同 pattern

#53 Implementation Plan 已 explicit 列為 out-of-scope。本 issue 追蹤是否同步 narrow fix。

Type

refactor (sync narrowing across sister skills)

Priority

P3 (none observed in real-world chain runs to date)

Strategy

For each location:

1. idd-all/SKILL.md:441,524: Python substring → equivalent line-start check (regex ^## Diagnosis on per-comment body)
2. idd-list/SKILL.md:115 + idd-update/SKILL.md:120: 釐清是 prose-only heuristic 還是有底層 jq/grep。若有具體 filter,sync narrow

Acceptance

• Inspect each cited line — confirm whether substring or anchored
• Apply same line-anchor fix as fix: idd-all-chain Step 0.4 jq filter use anchored regex (avoid false-positive on quoted text) #53 if substring
• Smoke test: 對每個 sister skill 跑 quoted-comment fixture 確認 no false-positive

Related

• fix: idd-all-chain Step 0.4 jq filter use anchored regex (avoid false-positive on quoted text) #53 (parent — chain-shell narrowing)
• cluster: chain follow-ups (#49 docs cross-link + #53 jq regex anchor) #58 (PR that fixed chain-shell only)

Source: surfaced during /idd-verify --pr 58 (cluster #49 #53), Step 5b triage. Master report: #58 (comment)

---

Current Status

• Phase: closed
• Diagnosed at: 2026-05-20

[kiki830621]

Diagnosis

Type

refactor (cross-skill consistency — sync narrowing fix from #53 to sister skills)

Root Cause / Analysis

#53 (PR #58) tightened idd-all-chain/SKILL.md Phase 0.4 diagnosis detection from substring "## Diagnosis" in comment_body → anchored regex test("(?m)^## Diagnosis") to eliminate false-positives from quoted / inline mentions (e.g., > see "## Diagnosis" pattern). But #53 verify Regression + Codex reviewers caught that 3 sister skills retain the old substring pattern and have the same false-positive risk:

• plugins/issue-driven-dev/skills/idd-all/SKILL.md:441 — Python substring if "## Diagnosis" in comment_body
• plugins/issue-driven-dev/skills/idd-all/SKILL.md:524 — same pattern
• plugins/issue-driven-dev/skills/idd-list/SKILL.md:115 — same pattern
• plugins/issue-driven-dev/skills/idd-update/SKILL.md:120 — same pattern

Impact

• 4 narrow sites across 3 skill files (idd-all, idd-list, idd-update)
• 0 behavior change for canonical IDD comments (which start ## Diagnosis at col-0)
• Reduces false-positive risk for adversarial comment bodies that quote the heading marker

Strategy

• idd-all/SKILL.md:441 Python substring → equivalent regex anchored multiline (re.search(r"(?m)^## Diagnosis", comment_body))
• idd-all/SKILL.md:524 same change
• idd-list/SKILL.md:115 substring → anchored regex
• idd-update/SKILL.md:120 substring → anchored regex
• Verify the new pattern matches the idd-all-chain Phase 0.4 helper convention (consistency across the family)
• CHANGELOG note about consistency sweep, not a new behavior

Complexity

Plan (cross-file refactor with no external contract change)

Layer P triggers:

• Cross-file refactor but no external contract change — same change applied to 4 sites across 3 sibling skill files
• Decision-heavy is mild — Python regex flavor vs jq test() flavor depends on each skill's existing tooling (bash + jq for idd-all-chain; Python likely for idd-list/idd-update if they're using Python helpers). Verify how each site is invoked

Layer 2 check: no published API change — internal detection logic. Not Spectra.

Risks

• Each site uses its own runtime (Python regex / bash jq / shell grep) — verify the equivalence per site
• If a sister skill silently relied on the substring's looseness (e.g., matching ### Diagnosis substring would still hit), the new anchored regex is a behavioral tightening — but that's the explicit point of the fix

Residue

(none) — 4 atomic sites, each with the same pattern swap, all consistent with the established #53 anchored-regex contract.

Vagueness Pre-check

• V1: 2 — Strategy bullets enumerate concrete file paths + line numbers from fix: idd-all-chain Step 0.4 jq filter use anchored regex (avoid false-positive on quoted text) #53 verify
• V4: 2 — Issue body's "## Expected" / "## Type" sections give explicit completion criteria
• Triggered: no — both axes ≤ 3

[kiki830621]

Implementation Complete

Checklist

• Strategy bullets shipped
• Commit: 9d58660 (cluster branch idd/chain-multi-b9be3134-refactor-sync-diagnosis-detection-narrow)
• PR opened: chain (multi-root): 3 issues — Phase 0.4 diagnosis-detection precision sweep from #53 verify (#114) #114 (cluster, all 3 sister issues)
• CHANGELOG v2.68.0 entry + plugin.json version bump

Scope shipped

2 Python substring sites in skills/idd-all/SKILL.md (line 450 + line 533) swapped from '## Diagnosis' in c['body'] to re.search(r'(?m)^## Diagnosis', c['body']). Cited sister sites in #59 issue body (idd-list:115 / idd-update:120) are narrative prose not executable code; idd-close:416 uses startswith() (already line-1-anchored). Scope narrowed to idd-all only.

Notes

• All changes additive — no behavior change for canonical IDD comments
• Per cluster discipline: per-issue verify findings will be partitioned in the cluster verify report at PR chain (multi-root): 3 issues — Phase 0.4 diagnosis-detection precision sweep from #53 verify (#114) #114

[kiki830621]

Verify (PR mode, cluster) — verify-gated PASS for this issue's portion of PR #114. 0 blocking findings; 3-AI ensemble PASS (Codex process gap recorded). Cluster ready to merge → /idd-close per issue after merge.

[kiki830621]

Closing Summary

Problem

#53 (PR #58) tightened idd-all-chain Phase 0.4 diagnosis detection from substring "## Diagnosis" in comment_body → anchored regex test("(?m)^## Diagnosis") to eliminate false-positives from quoted / inline mentions. #53 verify Regression + Codex reviewers flagged that sister skills retained the old substring pattern with the same false-positive risk.

Root Cause

/idd-all Phase 3 (single-issue orchestrator) and Phase 3b (Spectra context capture) used Python substring '## Diagnosis' in c['body'] for the same detection task. The fix was localized to idd-all-chain in #53; sister-skill consistency sweep was deferred to this issue.

Solution

skills/idd-all/SKILL.md (commit 9d58660):

• Line 450 (Phase 3 complexity readback) — '## Diagnosis' in c['body'] → re.search(r'(?m)^## Diagnosis', c['body'])
• Line 533 (Phase 3b Spectra inline Python) — same swap with explicit import re

Scope narrowed during implementation: cited sister-skill sites in the #59 issue body (idd-list/SKILL.md:115, idd-update/SKILL.md:120) are narrative prose describing the detection STRATEGY, not executable substring code. idd-close/SKILL.md:416 uses startswith("## Diagnosis") (jq) which is already line-1-anchored (stronger than substring, no false-positive risk). Only idd-all had actual substring patterns in real executable code.

Verification

3-AI verify ensemble (PR #114 cluster mode) — verify-gated PASS. Master: #114 (comment)

Requirements + Logic + Devil's Advocate all PASS. DA empirically confirmed idd-close:416 startswith() is line-1-anchored only (not substring), validating the scope narrowing.

Changes

• Squash merge: a15a6a5 (PR chain (multi-root): 3 issues — Phase 0.4 diagnosis-detection precision sweep from #53 verify (#114) #114)
• skills/idd-all/SKILL.md lines 450 + 533 substring → regex with (?m)^## Diagnosis anchor
• Shared CHANGELOG.md v2.68.0 + plugin.json 2.67.0 → 2.68.0

Residue

(none) — Scope narrowed during implementation but rationale is documented in the closing summary above. Cited sister sites that turned out to be prose are explicitly accounted for.