IDD human-in-the-loop should conform to the NSQL confirmation protocol (v4.1.0)

[kiki830621]

Problem

Original directive (Che, 2026-05-19): 「根據這個新的專案來更新 idd」 —— 具體為「IDD 裡面的 human in the loop 應該要符合 NSQL 的原則」。

NSQL (kiki830621/NSQL, now v4.1.0 — registered as a reference project in this repo's CLAUDE.md) is the Human-AI Confirmation Protocol that IDD's human-in-the-loop should conform to.

NSQL v4.1.0's model: a request triggers a dryrun — the AI renders "what I would do" without committing — which branches on two properties of that dryrun:

• clarity — exactly one interpretation?
• reversible — is a mistake recoverable?

clear and reversible  -> run, then report   ("traceable")
clear and not-reversible -> confirm
not-clear             -> disambiguate

NSQL principles: P1 Read-Only for Humans (humans confirm/correct, never write), P2 Explicit over Implicit, P3 Dynamic Confirmation Loop.

An audit of IDD's human-in-the-loop touchpoints against NSQL (2026-05-19) found IDD is ~80% already conformant — idd-plan EnterPlanMode and the AskUserQuestion checkpoints (Layer V 3-option, sister-concern, IC_R011) are textbook "AI renders -> human confirms / picks".

Type

feature (IDD methodology)

Expected

IDD's human-in-the-loop touchpoints are explicitly reconciled to NSQL's confirmation protocol — the conformant ones documented as such, the non-conformant ones fixed. A surgical reconciliation, not a redesign: IDD contains an NSQL-shaped confirmation loop (its front half — issue + idd-diagnose + complexity routing), it is not itself an NSQL extension.

Actual

IDD has never been audited against NSQL's protocol. Known non-conformances for idd-diagnose to verify and scope:

1. idd-issue has no dryrun / confirm step — it renders the structured issue and creates it without showing the rendering for the human to confirm or correct (violates P3 confirm-before-commit; should at least dryrun -> disambiguate-if-unclear -> report).
2. idd-diagnose Layer V "clarify now" makes the human write free-text answers — NSQL P1 says the AI should render candidate interpretations and the human picks / corrects.
3. IDD diagnosis has no residue concept — the part of an issue's intent that cannot be operationalized (its horizon / purpose) is never explicitly marked (NSQL section 4.6).
4. IDD complexity routing (Simple / Plan / Spectra) gates on clarity-of-complexity, not on reversibility — NSQL's traceability gate says an irreversible change should confirm regardless of how "Simple" it looks; Layer P "risk-sensitive boundary" only partially covers this.

Impact

• Affected surfaces: skills/idd-issue/, skills/idd-diagnose/ (Layer V), the complexity-routing rule (rules/sdd-integration.md), possibly MANIFESTO.md.
• Related: [enhancement] acceptance-review 缺在 IDD mode model — 「人不 review」需要 first-class 做法 #102 (acceptance-review / verify-gated — already established IDD's human-in-the-loop maps to the NSQL confirmation loop).
• Reference protocol: kiki830621/NSQL — see CLAUDE.md "Reference Projects".---

Current Status

Phase: verified
Last updated: 2026-05-20 by idd-verify

Key Decisions

• 6-AI verify on PR IDD human-in-the-loop should conform to the NSQL confirmation protocol (v4.1.0) #104 → PASS (0 blocking; 2 LOW polish + 4 INFO follow-ups); pre-merge gates Step 0.7/0.8 clean — see verify master
• In-PR polish committed bcdb3a0 — Layer V option description (idd-diagnose:281 + sdd-integration.md:64) realigned to candidate-render handler; Residue (none) requirement promoted to > paragraph
• 3 follow-up issues filed: idd-close: Residue acknowledgement prompt — close the F3 write-only loop #105 (idd-close Residue acknowledgement) Verify /plugin-tools:plugin-update rewrites plugin.json description chain (not only marketplace.json) #106 (plugin.json description chain via plugin-update) idd-issue Step 5: add CI/loop hard-parse warning for echo expansion (parallel to --no-multi-finding warning) #107 (idd-issue Step 5 echo CI/loop warning)
• Implemented F1–F4 on branch idd/103-nsql-conformance → PR IDD human-in-the-loop should conform to the NSQL confirmation protocol (v4.1.0) #104; plugin bumped to v2.64.0
• Implementation Plan approved via EnterPlanMode; Complexity re-routed Spectra → Plan
• F1 = report-echo (not a gate); F2 = Layer V renders candidates; F3 = residue field; F4 = Layer P wording
• [enhancement] acceptance-review 缺在 IDD mode model — 「人不 review」需要 first-class 做法 #102 owns the MANIFESTO doctrine; IDD human-in-the-loop should conform to the NSQL confirmation protocol (v4.1.0) #103 references it

Scope Changes

• (none)

Blocking

• (none — PR IDD human-in-the-loop should conform to the NSQL confirmation protocol (v4.1.0) #104 ready for merge → /idd-close #103)

Commits

• bcdb3a0 fix: align Layer V option description + Residue (none) emphasis (Refs IDD human-in-the-loop should conform to the NSQL confirmation protocol (v4.1.0) #103)
• 71d1044 chore: bump issue-driven-dev to v2.64.0 (IDD human-in-the-loop should conform to the NSQL confirmation protocol (v4.1.0) #103)
• a747c34 feat: reconcile IDD human-in-the-loop touchpoints to NSQL (IDD human-in-the-loop should conform to the NSQL confirmation protocol (v4.1.0) #103)---

Current Status

Phase: closed
Last updated: 2026-05-20 by idd-close

Key Decisions

• PR IDD human-in-the-loop should conform to the NSQL confirmation protocol (v4.1.0) #104 squash-merged to main as 675d7ba (3 commits collapsed; 5 plugin files +27/−10); plugin v2.64.0 lands in main
• Closing summary posted #issuecomment-4490409449 — 5-section format (Problem / Root Cause / Solution / Verification / Changes) per IDD discipline
• 6-AI verify on PR IDD human-in-the-loop should conform to the NSQL confirmation protocol (v4.1.0) #104 → PASS (0 blocking; 2 LOW polish + 4 INFO follow-ups); pre-merge gates Step 0.7/0.8 clean — verify master
• In-PR polish bcdb3a0 — Layer V option description (idd-diagnose:281 + sdd-integration.md:64) realigned to candidate-render handler; Residue (none) requirement promoted to > paragraph
• 3 follow-up issues filed: idd-close: Residue acknowledgement prompt — close the F3 write-only loop #105 (idd-close Residue acknowledgement) Verify /plugin-tools:plugin-update rewrites plugin.json description chain (not only marketplace.json) #106 (plugin.json description chain via plugin-update) idd-issue Step 5: add CI/loop hard-parse warning for echo expansion (parallel to --no-multi-finding warning) #107 (idd-issue Step 5 echo CI/loop warning)
• Implementation Plan approved via EnterPlanMode; Complexity re-routed Spectra → Plan
• F1 = report-echo (not a gate); F2 = Layer V renders candidates; F3 = residue field; F4 = Layer P wording
• [enhancement] acceptance-review 缺在 IDD mode model — 「人不 review」需要 first-class 做法 #102 owns the MANIFESTO doctrine; IDD human-in-the-loop should conform to the NSQL confirmation protocol (v4.1.0) #103 references it

Scope Changes

• (none)

Blocking

• (none — closed)

Commits

• 675d7ba IDD human-in-the-loop should conform to the NSQL confirmation protocol (v4.1.0) (IDD human-in-the-loop should conform to the NSQL confirmation protocol (v4.1.0) #104) [squash of a747c34 + 71d1044 + bcdb3a0]

[kiki830621]

Diagnosis

Type

feature — IDD methodology / human-in-the-loop reconciliation. (#103 authored this session; plain-text issue, no attachments.)

Analysis (requirements breakdown)

The issue asks one thing: reconcile IDD's human-in-the-loop touchpoints to NSQL's confirmation protocol (v4.1.0) — conformant ones documented, non-conformant ones fixed. The 4 findings in the body, evaluated:

F1 — idd-issue has no dryrun / confirm step. Real, but the framing needs sharpening. Under NSQL v4.1.0, creating a GitHub issue is reversible (an issue can be edited / closed) → clear ∧ reversible → run → report. So idd-issue creating an issue without a confirmation gate is NSQL-conformant for the clear case, and idd-issue already has the report half (report_and_stop). The genuine gap is narrower: idd-issue never does the dryrun clarity check — it renders the issue and creates it without ever assessing "is the human's intent clear enough to render as one issue, or ambiguous?". Fix = a dryrun step (render draft → if unclear, disambiguate → if clear, create + report), not a blanket confirm gate.

F2 — Layer V "clarify now" makes the human write. Real, clean. The clarify now handler has the AI ask 1-3 questions and the human write free-text answers. NSQL P1 (Read-Only for Humans): the AI should render candidate interpretations (the disambiguation shape) for the human to pick / correct. Fix = clarify now renders structured candidates where the AI can enumerate them; free-text only as the genuine fallback.

F3 — IDD diagnosis has no residue concept. Real. NSQL §4.6 residue = the part of intent that cannot be operationalized (its horizon / purpose) — marked, never silently dropped. IDD's diagnosis has Layer V (vagueness) and the IC_R011 sister-concern sweep, but nothing that marks "this part of the issue we cannot operationalize." Fix = a residue field in the Diagnosis report.

F4 — routing gates on complexity-clarity, not reversibility. Partially real — the weakest finding. IDD's Layer P already has "risk-sensitive boundary — concurrency, migrations, security-critical, save-durability, atomic operations", which is reversibility-adjacent. So IDD partially weighs reversibility, framed as "risk-sensitive". The gap is framing + completeness, not absence. Needs the most scoping.

Cross-issue overlap with #102. #102's discuss-conclusion already established "IDD's human-in-the-loop is the NSQL confirmation loop", and its deliverable #1 is a MANIFESTO.md doctrine. #103 is the touchpoint-level implementation of that conformance. The two must be reconciled so the doctrine (#102) and the fixes (#103) do not duplicate or contradict.

Impact

Surface	Finding
skills/idd-issue/SKILL.md	F1 — dryrun clarity step
skills/idd-diagnose/SKILL.md	F2 — Layer V clarify now handler; F3 — residue field in the Diagnosis report
rules/sdd-integration.md	F4 — routing reversibility signal
MANIFESTO.md	the "IDD's human-in-the-loop ⊂ NSQL confirmation loop" doctrine (overlaps #102 deliverable #1)
#102	reconcile scope — doctrine + --review vs touchpoint fixes

Strategy

Discuss agenda (routed Spectra → spectra-discuss):

• Reconcile IDD human-in-the-loop should conform to the NSQL confirmation protocol (v4.1.0) #103 with [enhancement] acceptance-review 缺在 IDD mode model — 「人不 review」需要 first-class 做法 #102 — IDD human-in-the-loop should conform to the NSQL confirmation protocol (v4.1.0) #103 scope vs [enhancement] acceptance-review 缺在 IDD mode model — 「人不 review」需要 first-class 做法 #102's MANIFESTO doctrine + --review
• F1 — idd-issue: confirm gate vs dryrun-clarity-check vs report-only (creating an issue is reversible → leans clarity-check, not a gate)
• F2 — Layer V clarify now: render candidates; free-text fallback when un-enumerable
• F3 — residue: where in the Diagnosis report? interaction with Layer V?
• F4 — reversibility: explicit routing signal, or recognise Layer P already covers it (framing tweak)?
• Scope: one change, or split into per-finding sub-issues?

Complexity

Spectra

• Layer 2 — the change reconciles 2+ published IDD skills (idd-issue, idd-diagnose) to an external reference protocol (NSQL), establishing an ongoing conformance contract that future IDD human-touchpoints must follow. That contract belongs in a documented spec.
• Layer 3 — (a) modifies normative behavior of 2+ published skills; (b) cross-cuts idd-issue + idd-diagnose + sdd-integration.md + MANIFESTO.md — needs consistency-checking; (c) an architectural decision with long-term implications (every future IDD human-touchpoint inherits "conform to NSQL").

Spectra default = discuss-first (not opt-out): genuine open questions remain — the per-finding approach, the #102 reconciliation, and the one-change-vs-split scope decision.

Risks

• Scope balloon — 4 findings + doctrine + contract. Discuss must scope tightly; the one-vs-split decision is load-bearing.
• [enhancement] acceptance-review 缺在 IDD mode model — 「人不 review」需要 first-class 做法 #102 overlap — risk of duplicate / contradictory doctrine. Discuss must reconcile [enhancement] acceptance-review 缺在 IDD mode model — 「人不 review」需要 first-class 做法 #102 first.
• Over-fix on F1 — adding a blanket confirm gate to idd-issue would itself violate NSQL (creating an issue is reversible → NSQL says run). The fix must be the dryrun-clarity-check.
• F4 may be a near-non-finding — Layer P may already cover reversibility; discuss decides real-gap vs framing-tweak.

Vagueness Pre-check (Step 3.4)

Axis	Score	Reasoning
V1 (vague WHAT)	3 (mostly clear)	The WHAT — conform IDD's human-in-the-loop to NSQL — is clear; the body names 4 specific findings + the reference protocol. The per-finding HOW is decision-heavy, not vague — Spectra-discuss territory.
V4 (vague ACCEPTANCE)	3 (mostly clear)	Expected spells out "touchpoints reconciled, conformant documented, non-conformant fixed". Completion = the 4 findings resolved + the conformance documented. Not crisp-testable (design work) but not vague.

• Triggered: no — both axes ≤ 3.

Next Step

Complexity = Spectra → /spectra-discuss (topic: "IDD human-in-the-loop conformance to NSQL"). Opt-out to spectra-propose is not available — genuine open questions remain (per-finding approach, #102 reconciliation, scope split).

Sister Concerns Filed (Step 3.6)

none surfaced — the 4 findings are the issue scope; #102 is an existing issue (reconcile, not file new).

[kiki830621]

Discuss Conclusion — IDD human-in-the-loop conformance to NSQL

/spectra-discuss (2026-05-20). Captures the conclusion + the Spectra → Plan re-route.

Re-evaluation under NSQL v4.1.0

The diagnosis's 4 findings were audited against NSQL's P1/P2/P3 principles, before NSQL v4.1.0's traceability gate was finalized. Re-evaluated under v4.1.0 (dryrun → {clear∧reversible: run→report, clear∧¬reversible: confirm, ¬clear: disambiguate}):

• F1 dissolves — creating a GitHub issue is reversible → NSQL v4.1.0 says run, no confirm gate. idd-issue already renders (dryrun) + creates (run) + reports. The clarity branch correctly lives downstream at idd-diagnose Layer V. F1's residual is tiny: idd-issue's report_and_stop should echo the rendered issue draft (not just the URL) so a misparse is catchable at a glance — run → report done properly. A confirm gate would itself violate v4.1.0.
• F2 stands — the genuine fix. Layer V "clarify now" makes the human write free-text answers; NSQL P1 (Read-Only for Humans) says the AI should render candidate interpretations for the human to pick. Free-text only as fallback.
• F3 stands — small additive: a residue field in the Diagnosis report (NSQL §4.6 — mark non-operationalizable horizon).
• F4 downgrades to wording — Layer P already weighs reversibility ("risk-sensitive boundary"); F4 = add "irreversible side effects" to that bullet's wording. No new mechanism.

Decision

#103's scope = 4 small touchpoint reconciliations:

#	Change	File
F1	idd-issue report_and_stop echoes the rendered issue draft	skills/idd-issue/SKILL.md
F2	Layer V "clarify now" renders candidate interpretations to pick; free-text fallback	skills/idd-diagnose/SKILL.md
F3	add a residue field to the Diagnosis report	skills/idd-diagnose/SKILL.md
F4	Layer P risk-sensitive bullet — add "irreversible side effects"	rules/sdd-integration.md

Complexity re-route: Spectra → Plan. None of the four is a frozen-contract change — they are behavior tweaks to 2 skills + 1 rule, no new published surface. (Same pattern as #102: discuss shrank the Spectra diagnosis to Plan.)

Scope: one change, not split per-finding — small and coherent enough to plan + implement together.

#102 boundary: #102 owns the MANIFESTO.md doctrine ("IDD's human-in-the-loop IS the NSQL confirmation loop") + the terminal disposition (verify-gated / --review). #103 owns the front-half touchpoint fixes (the 4 above) and references the doctrine — it does not re-author it. No overlap.

Next Step

/idd-plan #103.

[kiki830621]

Implementation Plan (Plan tier)

PR path — feature branch idd/103-nsql-conformance. Scope = the 4 touchpoint reconciliations from the discuss conclusion (F1 report-echo, F2 Layer V candidates, F3 residue field, F4 Layer P wording).

Files & Changes

• plugins/issue-driven-dev/skills/idd-issue/SKILL.md — F1: Step 5 "回報並停止" currently outputs issue number / URL / labels / type. Add: echo the AI-rendered interpretation (Type / Expected / Actual + the plain-language interpretation — the parts the AI produced, where a misparse would live) inline in the report. The verbatim "Original text" the human already gave is not re-echoed. This is NSQL run → report done properly — the report states what was done, not just where it is.
• plugins/issue-driven-dev/skills/idd-diagnose/SKILL.md — F2: Step 3.4 D.1 clarify now handler. Currently the AI picks 1-3 unclear points and the user writes free-text answers. Rewrite: the AI renders candidate interpretations per unclear point (numbered options — the NSQL disambiguation shape, "1. X (recommended) 2. Y 3. Z"); the user picks. Free-text stays as the named fallback for genuinely un-enumerable questions. (NSQL P1 — Read-Only for Humans.)
• plugins/issue-driven-dev/skills/idd-diagnose/SKILL.md — F3: Step 3 Diagnosis report template — add a ### Residue section + a one-paragraph explanation. Residue = the part of the issue's intent that cannot be operationalized (its purpose / horizon), per NSQL §4.6 — marked, never silently dropped. The explanation MUST distinguish residue from Layer V vagueness: Layer V = the issue is unclear; residue = the issue is clear but part of its intent is non-operationalizable.
• plugins/issue-driven-dev/rules/sdd-integration.md — F4: Layer P "risk-sensitive boundary" bullet — add "irreversible side effects" to the enumerated list. Wording only — no new routing mechanism. Sync the inline Layer P summary in idd-diagnose/SKILL.md Step 3.5 if it enumerates the same list.
• plugins/issue-driven-dev/CHANGELOG.md — entry for this change.
• plugins/issue-driven-dev/.claude-plugin/plugin.json — minor version bump (behavior change to 2 skills + 1 rule).

Sequencing & Dependencies

F1–F4 are independent — different files / regions, no ordering dependency, no file's change affects another's. Edit in any order. CHANGELOG + plugin.json version bump land last (they describe the change). marketplace.json sync is not in this PR — it is a close/distribution-time step (idd-close Step 6.5 / plugin-update).

Decision Points

• F1 is a report-strengthening, NOT a confirm gate. The discuss established (NSQL v4.1.0) that creating an issue is reversible → NSQL says run, no gate; a gate would violate v4.1.0. Implement ONLY the report echo. Do not "upgrade" F1 to a gate.
• F2 keeps a free-text fallback. Not every clarification is enumerable. The handler renders candidates where it can; genuinely-open questions still go free-text. Both paths stay.
• F4 is wording, not mechanism. Layer P already weighs reversibility. No new signal / Layer.

Risks & Mitigations

• F1 over-implementation → a confirm gate. Mitigation: plan + IDD human-in-the-loop should conform to the NSQL confirmation protocol (v4.1.0) #103 discuss conclusion both state "report echo, not gate" explicitly.
• F3 / Layer V overlap → residue conflated with "vague WHAT". Mitigation: the ### Residue explanation explicitly contrasts the two.
• F2 → losing the free-text path. Mitigation: the handler spec names free-text as the fallback.
• Skill-prose change, no automated harness. Mitigation: dogfood + consistency check (see Test Plan); the 6-AI idd-verify is the real gate.

Test Plan

IDD skills are prose (SKILL.md) — no automated harness. Verification = consistency + dogfood:

1. Consistency — each of the 4 edits is internally consistent with its SKILL.md / rule (no dangling refs; examples still parse).
2. Dogfood F2 — a vague issue diagnosed via /idd-diagnose has Layer V "clarify now" render numbered candidates, not ask free-text.
3. Dogfood F1 — /idd-issue echoes the rendered interpretation in its Step 5 report.
4. F3 — a diagnosis report carries the ### Residue section.
5. The PR-path 6-AI idd-verify ensemble is the real gate.

Out-of-scope (will NOT be touched)

• MANIFESTO.md — the "IDD human-in-the-loop = NSQL confirmation loop" doctrine is [enhancement] acceptance-review 缺在 IDD mode model — 「人不 review」需要 first-class 做法 #102's deliverable. IDD human-in-the-loop should conform to the NSQL confirmation protocol (v4.1.0) #103 references it; [enhancement] acceptance-review 缺在 IDD mode model — 「人不 review」需要 first-class 做法 #102 authors it.
• The --review flag / verify-gated terminal disposition — [enhancement] acceptance-review 缺在 IDD mode model — 「人不 review」需要 first-class 做法 #102.
• Any new routing Layer / new skill / new mechanism — IDD human-in-the-loop should conform to the NSQL confirmation protocol (v4.1.0) #103 is 4 small touchpoint tweaks only.
• Documenting the already-conformant touchpoints (idd-plan EnterPlanMode, the AskUserQuestion checkpoints, idd-close confirm) — doc-polish, deferred.

Tangential Observations (filed mid-plan, v2.42.0+ #524)

none surfaced — the plan is self-contained; no off-scope discoveries during drafting.

[kiki830621]

Implementation Complete

PR: #104 · branch idd/103-nsql-conformance

Checklist (synced from TaskList)

• F1 — idd-issue Step 5 report echoes the AI-rendered interpretation → a747c34
• F2 — idd-diagnose Layer V clarify now renders candidate interpretations → a747c34
• F3 — ### Residue section added to the Diagnosis report template → a747c34
• F4 — Layer P "risk-sensitive boundary" adds "irreversible side effects" (sdd-integration.md + idd-diagnose inline) → a747c34
• CHANGELOG entry → 71d1044
• plugin.json version bump 2.63.0 → 2.64.0 → 71d1044

Legend: - [x] done. idd-close will refuse to close if any - [ ] remains.

Changes

• a747c34 — feat: reconcile IDD human-in-the-loop touchpoints to NSQL (F1–F4)
• 71d1044 — chore: bump issue-driven-dev to v2.64.0 + CHANGELOG

Files Changed

• plugins/issue-driven-dev/skills/idd-issue/SKILL.md — F1
• plugins/issue-driven-dev/skills/idd-diagnose/SKILL.md — F2 + F3 + F4 inline
• plugins/issue-driven-dev/rules/sdd-integration.md — F4
• plugins/issue-driven-dev/CHANGELOG.md + .claude-plugin/plugin.json — v2.64.0

Scope Compliance

On-plan — exactly the 4 touchpoint reconciliations + version/CHANGELOG. MANIFESTO.md untouched (the IDD↔NSQL doctrine is #102's deliverable, per the plan's Out-of-scope). No new mechanism / Layer / skill.

Sister Bugs Filed (mid-impl, Step 5.7)

none surfaced — F4's Layer P list appears in two places (sdd-integration.md canonical + idd-diagnose inline summary), but that summary/canonical split is a deliberate documented pattern (the skill states 完整邏輯見 sdd-integration.md), not a sister bug. Both copies were updated in a747c34.

Next: verification pending

/idd-verify --pr 104

[kiki830621]

Verify (via PR #104)

Result: PASS — 0 blocking · 2 LOW (in-scope polish) · 4 INFO/follow-up · 6-AI consensus

Full report: #104 (comment)

Per-#103 highlights:

• F1–F4 all FULLY addressed (5/5 Claude reviewers + Codex agree)
• Scope clean: MANIFESTO.md untouched ([enhancement] acceptance-review 缺在 IDD mode model — 「人不 review」需要 first-class 做法 #102's deliverable), exactly the 5 planned files changed
• Backward compat ✓: ### Residue is additive, no downstream parser breaks
• Devil's Advocate: 5/5 angles dismissed (F1 reversibility, F2 enumerable, F3 vs vagueness, F4 ritual, audit completeness)

Recommended same-PR polish before merge (per repo feedback_verify_fix_same_pr): update Layer V option-description (idd-diagnose/SKILL.md:281 + sdd-integration.md:64) to match the new candidate-render handler.

[kiki830621]

Closing Summary

Problem

IDD's human-in-the-loop touchpoints had never been audited against an explicit confirmation protocol. NSQL (kiki830621/NSQL v4.1.0 — Human-AI Confirmation Protocol) was registered as a reference project in this repo's CLAUDE.md (commit fd2f21c); a 2026-05-19 audit then surfaced 4 specific non-conformances in IDD's front half:

1. F1 — idd-issue rendered the issue body and posted it, but Step 5 reported only metadata (number / URL / labels / type) — a misparse of the user's intent was invisible until they clicked through to the issue.
2. F2 — idd-diagnose Layer V clarify now asked the user to write free-text answers, contradicting NSQL P1 (Read-Only for Humans, AI renders → human picks).
3. F3 — Diagnosis reports had no place to mark the purpose / horizon of an issue that couldn't be operationalized; non-operationalizable intent silently dropped (NSQL §4.6).
4. F4 — Layer P "risk-sensitive boundary" enumerated migrations / save-durability etc. but not the broader category of irreversible side effects (Telegram send, Stripe charge, webhook fire, file delete) — NSQL's reversibility axis was implicit.

Root Cause

IDD predates NSQL v4.1.0's traceability gate. The implicit confirmation loop in IDD (EnterPlanMode + AskUserQuestion checkpoints) was already ~80% NSQL-conformant by accident, but the front-half touchpoints (idd-issue + idd-diagnose Layer V + Layer P vocabulary + Diagnosis template structure) had drifted in 4 specific ways. Without a reference protocol to align to, the drift was invisible.

The audit's first read flagged all 4 as gaps requiring new mechanisms. Re-evaluating under NSQL v4.1.0's actual model (dryrun → {clear∧reversible: run→report, clear∧¬reversible: confirm, ¬clear: disambiguate}) showed F1 dissolved (creating an issue is reversible → no confirm gate, just a better report) and F4 downgraded (vocabulary add, not mechanism add). F2 and F3 remained as substantive changes.

Solution

Surgical reconciliation across 5 files (no new skill / Layer / mechanism — Plan-tier scope, EnterPlanMode-approved):

Finding	Change	File
F1	Step 5 report echoes AI-rendered ## Type / ## Expected / ## Actual + plain-language interpretation (NSQL run → report, not a gate)	plugins/issue-driven-dev/skills/idd-issue/SKILL.md:876-878
F2	Layer V clarify now handler renders candidate interpretations for the user to pick (NSQL P1); free-text named as fallback for un-enumerable questions	plugins/issue-driven-dev/skills/idd-diagnose/SKILL.md:290-304
F3	Diagnosis report template gains ### Residue field + > explanatory paragraph distinguishing residue (non-operationalizable intent of a clear issue) from Layer V vagueness (unclear issue); (none) empty-state required	plugins/issue-driven-dev/skills/idd-diagnose/SKILL.md:204-205, 221
F4	Layer P "risk-sensitive boundary" enumerated list adds "irreversible side effects" (both canonical English rule + Chinese inline copy)	plugins/issue-driven-dev/rules/sdd-integration.md:148 + plugins/issue-driven-dev/skills/idd-diagnose/SKILL.md:412

Plugin bumped 2.63.0 → 2.64.0. CHANGELOG entry under [2.64.0] - 2026-05-20 records F1's dissolution and F4's downgrade explicitly (the audit trail of why the implementation differs from the initial audit framing).

MANIFESTO.md was deliberately not touched — the IDD↔NSQL doctrine ("IDD's human-in-the-loop is an NSQL confirmation loop, not a separate model") is #102's deliverable. #103 references it (via the CLAUDE.md Reference Projects entry) but does not author it.

Verification

6-AI cross-model verify on PR #104 (5 general-purpose Claude reviewers + Codex gpt-5.5 xhigh in run_in_background) → PASS, 0 blocking findings. All 4 findings labeled FULLY addressed by every reviewer.

Source	Verdict
Requirements	PASS — F1–F4 全 FULLY addressed
Logic	PASS — 3 LOW polish
Security	PASS — 0 findings
Regression	PASS — 1 LOW + 1 INFO
Devil's Advocate	PASS — 5/5 angles dismissed (F1 reversibility / F2 enumerable / F3 vs vagueness / F4 ritual / audit completeness)
Codex (gpt-5.5 xhigh)	PASS — 2 INFO

Pre-merge gates clean:

• Step 0.7 PR↔issue correspondence ✓ (Refs #103)
• Step 0.8 auto-close trap ✓ (closingIssuesReferences + per-commit messageHeadline/messageBody regex both empty)

In-PR polish committed as bcdb3a0 (verify-driven, same-PR per repo feedback_verify_fix_same_pr):

• F2 polish — Layer V clarify now AskUserQuestion option description (idd-diagnose/SKILL.md:281 + sdd-integration.md:64) was still describing the pre-PR free-text flow; aligned to the candidate-render handler (3 reviewers convergent: regression R1, DA D1, Codex)
• F3 polish — (none) empty-state requirement promoted from curly-brace placeholder to the > explanatory paragraph for discoverability (logic L3)

3 follow-up issues filed from Step 5b triage:

• idd-close: Residue acknowledgement prompt — close the F3 write-only loop #105 — idd-close: Residue acknowledgement prompt (close the F3 write-only loop; D2)
• Verify /plugin-tools:plugin-update rewrites plugin.json description chain (not only marketplace.json) #106 — Verify /plugin-tools:plugin-update rewrites plugin.json description chain, not only marketplace.json (R2)
• idd-issue Step 5: add CI/loop hard-parse warning for echo expansion (parallel to --no-multi-finding warning) #107 — idd-issue Step 5: add CI/loop hard-parse warning for echo expansion (Codex)

Reports:

• Master verify: IDD human-in-the-loop should conform to the NSQL confirmation protocol (v4.1.0) #104 (comment)
• Pointer on IDD human-in-the-loop should conform to the NSQL confirmation protocol (v4.1.0) #103: IDD human-in-the-loop should conform to the NSQL confirmation protocol (v4.1.0) #103 (comment)

Changes

• 675d7ba IDD human-in-the-loop should conform to the NSQL confirmation protocol (v4.1.0) (IDD human-in-the-loop should conform to the NSQL confirmation protocol (v4.1.0) #104) — squash-merge of a747c34 + 71d1044 + bcdb3a0
• fd2f21c (precondition, earlier session) docs: add NSQL as a reference project in CLAUDE.md

Files changed (5 in plugin, 1 in repo root pre-existed):

• plugins/issue-driven-dev/skills/idd-issue/SKILL.md — F1
• plugins/issue-driven-dev/skills/idd-diagnose/SKILL.md — F2 + F3 + F4 inline
• plugins/issue-driven-dev/rules/sdd-integration.md — F4 canonical
• plugins/issue-driven-dev/CHANGELOG.md — v2.64.0 entry
• plugins/issue-driven-dev/.claude-plugin/plugin.json — version bump

Total diff: +27/−10 across 5 files.

Distribution Sync (Step 6.5)

• Detected: plugin (issue-driven-dev) — self-hosted marketplace at .claude-plugin/marketplace.json
• Choice: chain /plugin-tools:plugin-update issue-driven-dev
• Outcome:
  • 2f63feb chore: sync marketplace.json — version 2.63.0 → 2.64.0 + v2.64.0 description prefix prepended
  • db25514 docs: sync README Version History — extended v2.56.0–v2.62.0 row to v2.56.0–v2.64.0 + 2 new entries (cluster-PR mode 強制 PR 與 direct-commit path 衝突：override 未文件化且 --no-pr 行為未定義 #96-backlog Simple cluster + this IDD human-in-the-loop should conform to the NSQL confirmation protocol (v4.1.0) #103 PR)
  • claude plugin marketplace update issue-driven-development → ✓
  • claude plugin update issue-driven-dev@issue-driven-development → ✓ 2.63.0 → 2.64.0
• Empirical finding for #106: /plugin-tools:plugin-update v1.16.0 syncs marketplace.json description chain only. plugins/issue-driven-dev/.claude-plugin/plugin.json description still leads with "v2.63.0:" after this sync — the skill does NOT rewrite plugin.json description. Verify /plugin-tools:plugin-update rewrites plugin.json description chain (not only marketplace.json) #106 captures this gap; either fix the skill or document it as the contract.
• At: 2026-05-19T23:05:59Z