intent-improve

name	intent-improve
description	After implementation, cross-check the .intent/ deliverables against the implementation reality on the three axes of completeness / correctness / coherence, classify the drift, and propose corrections. Reflects changes only after user approval. Guides missed write-backs to /intent-writeback.
disable-model-invocation	true
allowed-tools	Read, Write, Glob, Grep, AskUserQuestion
argument-hint	<target scope (optional)>

intent-improve Skill

Core Mission

• Success Criteria:
  • The .intent/ deliverables and the implementation reality (codebase, tests, and progress in the selected specification tool) are evaluated on the three axes (completeness / correctness / coherence)
  • The evaluation results are presented in the 5 classifications (aligned / intent reinforcement recommended / corrective packet recommended / Decision Rules update recommended / invariant violation detected), with evidence (file / relevant text)
  • Only the corrections the user approved are reflected into .intent/ (the approval unit is per proposal)
  • When an unrecorded write-back learning is detected, the skill does not write a delta itself but prompts the user to run /intent-writeback
  • On a run whose 5 classifications include Decision Rules update recommended or invariant violation detected, following the provisions of rules/improve-axes.md, it also prompts the user to run /intent-validate (the check for conformance catch-up) (on a run that does not include them, it does not prompt; it only guides and does not itself make a conformance judgment)
  • When drift-watch is on, the coherence detections are recorded into drift-log as stage:improve / outcome:missed and a pattern×outcome improvement report is presented (when off / missing / invalid / missing section / missing mode.md, nothing is done; the 5 classifications are unchanged)
  • No application code has been changed at all

Execution Steps

Step 1: Collect the current state

• Read the .intent/ deliverables (intent-tree.md / intent-compass.md / .intent/packets/index.md + the packet files under active/ (cross-cutting read for the completeness axis; do not read archive/) / the per-packet drafts under the selected exit (cc-sdd/ / openspec/ / speckit/) / deltas.md). If .intent/ is absent, guide the user through setup (installing intent-planner and running /intent-discover) and stop.
• Read the mode state in order: the inherited issue directory's discovery/<slug>-<rand>/mode.md (A34; inherit the issue name that discover output) → else the single .intent/mode.local.md (legacy) → else old .intent/mode.md (the CONTRACT.md read fallback contract). If both are absent, continue with the standard default and announce it.
• Collect the implementation reality: the codebase (read-only via Read/Glob/Grep), the presence and placement of tests, the progress of .kiro/specs/ (only if it exists), and the promoted / pending entries of deltas.md.
• If the selected specification tool's deliverables are absent, continue without that context. If deltas.md is absent, continue treating it as "no delta records" (non-blocking).
• If a target scope is specified by argument, narrow down to it; otherwise target the whole of .intent/.
• You may partially load compass by domain scope (opt-in; federated-governance / INV101): when the target domain is determined by the domain argument or the case context, read and apply rules/domain-scope.md. Do not full-load compass; pull only "the case's domain tag + always" with grep + inline tags (extending the INV47 pull discipline into the maintenance loop). The coherence axis's safety net (the cross-cutting matching of drift not tied to a packet, decayed rules, and broken premises) targets all symbols, not narrowed by domain scope (do not weaken detection; B-fed5). For a legacy scaffold with no determinable domain and no tags, fall back to the legacy full read (backward-compatible; O3). Do not change the three-axis evaluation or the five-way classification logic (only the range of loaded symbols changes).

Step 2: Evaluate on the three axes

• Read rules/improve-axes.md and cross-check .intent/ against the implementation reality on the three axes of completeness / correctness / coherence.
• Always attach evidence (file / relevant text) to the evaluation. Do not present an evaluation whose evidence cannot be shown.

Step 3: Classify and report

• Classify the evaluation results into the 5 classifications (aligned / intent reinforcement recommended / corrective packet recommended / Decision Rules update recommended / invariant violation detected; multiple may apply) and present them organized per classification.
• When unrecorded write-back learnings or declined items with the "on-hold" tag are detected, also include guidance to /intent-writeback following the provisions of rules/improve-axes.md.
• On a run whose 5 classifications include Decision Rules update recommended or invariant violation detected, following the "Validate catch-up guidance" provisions of rules/improve-axes.md, also include guidance to /intent-validate (the check for conformance catch-up) alongside the writeback guidance (on a run that does not include them, do not include it; only guide and do not make the judgment yourself).
• When drift-watch is on (when off / missing / invalid / missing section / missing mode.md, do nothing): check the drift-watch value in the ## Drift-watch (user-managed) section of .intent/mode.md, and only when it is on, following the provisions of rules/improve-axes.md, record the drift detected on the coherence axis (invariant violation / anti-direction conflict) into .intent/drift-log.md as a stage: improve / outcome: missed draft, and present a pattern × outcome cross-tabulated improvement report. Delegate the recording details (the fixed 9-key order, append-only, obtaining commit, creating drift-log when absent) to rules/improve-axes.md (do not duplicate them here). This recording does not create a new correction class (the 5 classifications above are unchanged), and does not write into deltas.md or hook writeback. When off / missing / invalid / missing section / missing mode.md, do not record or aggregate drift and proceed as before (byte-equivalent to current behavior). Note that the report of the 5 classifications above is always done regardless of the drift-watch value.

Step 4: Confirm approval per proposal for the corrections

• For each item that needs correction, present the correction proposal (a deliverable update proposal or a corrective packet proposal) and confirm the user's approval per proposal (do not force bulk approval).
• Proposals that were not approved end as presentation only (do not rewrite).

Step 5: Reflect only the approved corrections

• Reflect only the approved corrections into the canonical deliverables (intent-tree.md / intent-compass.md / under .intent/packets/ (the target packet file / plan.md)).
• When the canonical under .intent/packets/ has been changed (including when a delta promotion is reflected into the target packet file), regenerate .intent/packets/index.md from the frontmatter under active/.
• Corrections that change the Decision Rules follow the change convention of rules/improve-axes.md (add a new entry in ADR form + annotate the old entry as superseded with a reference to its successor and move it to .intent/compass-archive/<rule-slug>.md (a per-rule file)).
• Do not write into deltas.md (recording deltas and finalizing declined-item tags are the responsibility of /intent-writeback).

Output Description

The output target is the terminal. Use no raw HTML (<details> / <summary>, etc., collapsible UI) in the output; separate details with plain Markdown headings instead (in a terminal the raw tags are shown literally and become unreadable). Internal notations such as [[...]] (wikilinks for memory / delta) are legitimate in records written to delta / memory files, but in human-facing terminal output do not emit them raw — open them into ordinary words (spell the linked name out in plain prose).

Reader: a human developer who, after implementation, approves and corrects the drift between intent and implementation. What this output makes them grasp first: "here is the drift between implementation and intent (invariant violations come first if any). There are N items pending approval. If there is a missed write-back, go to /intent-writeback." The breakdown of the three-axis evaluation is detail for the decision.

Lead the output with the conclusion (the drift and the pending approvals).

• Domain-scope target count (only on runs under domain scope; just before the summary): on runs that partially loaded via rules/domain-scope.md, state the number of loaded symbols and the total (e.g. "並行 + always = 41 symbols / of 829 total; the coherence safety net scanned all symbols"). Omit it on full-read runs (as before).
• Divergence summary (top): show the key points of the detected corrections by classification. If there is an invariant violation detected, lead with it as the top priority. If everything is aligned (no drift), state explicitly "aligned, no correction needed".
• Approval-pending list (next, per proposal): attach evidence (file / relevant text) to each correction proposal. Phrased so it is clear what gets reflected by approving what.
• Writeback guidance (when applicable): when an unrecorded write-back learning is detected, guidance to run /intent-writeback.
• Validate catch-up guidance (when applicable): on a run whose 5 classifications include Decision Rules update recommended / invariant violation detected, attach the guidance to run /intent-validate (the check for conformance catch-up) alongside the writeback guidance.
• Details: the three-axis evaluation summary (completeness / correctness / coherence) and the breakdown by classification (aligned / intent reinforcement recommended / corrective packet recommended / Decision Rules update recommended / invariant violation detected).
• Improvement report (when drift-watch=on): a report cross-tabulating drift-log by pattern × outcome. Always attach the honesty notes (missed=0 is a suspicion of missing records / frequent false-positive suggests the anti-direction is too broad), align the aggregation keys to the type (pattern), and establish the group comparison (without group / with group) by the type id and the commit column of drift-log only (do not create an additional comparison mechanism).

Plainness check for reports (user-facing reports; right before output; shared)

Right before emitting a user-facing report (progress, completion, items needing confirmation — including the end-of-turn summary), run this check (INV105, DR208). It applies only to user-facing report text, not to how internal records (canonical files and logs under .intent/) are written.

• Do not transcribe internal documents verbatim: text you just read or wrote in internal artifacts (tree, compass, packets, Open Questions) is written in internal vocabulary. In the report, restate that content in words a first-time reader understands (without changing facts or meaning).
• Identifiers must not be the subject of the sentence: when presenting an item to confirm or a unit of work, first write one sentence that stands on its own ("what and why"), then append identifiers (Open Question numbers, packet names, symbols, stage names) after it as references (e.g. "... please verify this before starting (ref: OQ-xxx-1)"). Do not delete identifiers or references to records for the sake of plainness (the trail back to the record is lost).
• Signal for overload: three or more unexplained internal terms in one sentence signal overload (read by meaning, not by mechanical count). If a sentence does not stand on its own, rewrite it in plain words before sending (without changing facts or meaning).
• Do not convey meaning only through a metaphor or a vague qualifier: the foundation of a report is precision — write so the meaning reads unambiguously (plain language is a means of staying easy to read while preserving it). Do not report results only with ungrounded qualifiers (e.g. "significantly", "nicely"); state observable facts. If you use a metaphor, pair it immediately with a precise restatement (do not force established technical terms, or ordinary words in their everyday sense, into strained paraphrases).
• This check works as a pair with the after-the-fact record (prevention alone is never enough): when a report failed to get through, log the case to the drift log while drift-watch is on, and feed the next prevention.

Safety & Fallback

• Do not rewrite the .intent/ deliverables without user approval. Confirm approval per proposal.
• Do not change application code (INV6. Code is read-only via Read/Glob/Grep).
• Do not write into the selected specification tool's deliverables (progress is read-only). If those deliverables are absent, continue without that context.
• Do not write directly into deltas.md. Handling of missed write-backs and on-hold items is guidance to /intent-writeback only; the final updates are done by writeback.
• The absence of .intent/ guides the user through setup and stops. The absence of mode.md does not stop; continue with the standard default and announce it.