Check Consistency

check-consistency

Scan ytstack's own repo for drift between its docs (README, concept, DECISIONS, QUICKSTART, CLAUDE.md) and its implementation (skills, hooks, artifacts). Emit one structured report at .ytstack/CONSISTENCY-REPORT.md.

README.md is the authoritative source of truth; every other file is compared against it. When a skill, hook, or workflow is mentioned in a doc but does not exist on disk -- or when it exists on disk but no doc references it -- this skill surfaces the gap. It does not silently fix either side.

Anti-Pattern: "docs look fine, skip the audit"

Doc drift accumulates silently: a skill gets renamed, a hook is added for a hotfix, a README bullet is written aspirationally ahead of the code, a concept paper is generated in one session and forgotten in the next. By the time an end-to-end smoke-test reveals the mismatch, the agent has been primed with false claims for weeks. This skill runs in seconds. Run it.

Checklist

You MUST create a TodoWrite task for each item and complete them in order:

1. Run preamble -- locate the plugin root and collect on-disk inventory
2. HARD-GATE -- abort if the plugin root is missing core files
3. Read authoritative docs -- README.md, QUICKSTART.md, CLAUDE.md, docs/concept.md, docs/methodology.md, .ytstack/DECISIONS.md
4. Read implementation inventory -- skills/, hooks/hooks.json, agents/, bin/
5. Run the nine consistency checks (below)
6. Emit the report to .ytstack/CONSISTENCY-REPORT.md
7. Print a one-screen summary of findings to the user
8. Return control -- do NOT auto-fix anything

Process Flow

digraph check_consistency {
    "Run preamble" [shape=box];
    "Plugin root OK?" [shape=diamond];
    "Abort with instructions" [shape=doublecircle];
    "Read docs + inventory" [shape=box];
    "Run nine checks" [shape=box];
    "Write report" [shape=box];
    "Print summary" [shape=box];
    "Return" [shape=doublecircle];

    "Run preamble" -> "Plugin root OK?";
    "Plugin root OK?" -> "Abort with instructions" [label="no"];
    "Plugin root OK?" -> "Read docs + inventory" [label="yes"];
    "Read docs + inventory" -> "Run nine checks";
    "Run nine checks" -> "Write report";
    "Write report" -> "Print summary";
    "Print summary" -> "Return";
}

Preamble

_PLUGIN_ROOT="${CLAUDE_PROJECT_DIR:-$PWD}"
cd "$_PLUGIN_ROOT" 2>/dev/null || { echo "ERR: cannot enter plugin root $_PLUGIN_ROOT"; exit 1; }

_HAS_README=$([ -f README.md ] && echo yes || echo no)
_HAS_CONCEPT=$([ -f docs/concept.md ] && echo yes || echo no)
_HAS_DECISIONS=$([ -f .ytstack/DECISIONS.md ] && echo yes || echo no)
_HAS_QUICKSTART=$([ -f QUICKSTART.md ] && echo yes || echo no)
_HAS_CLAUDE_MD=$([ -f CLAUDE.md ] && echo yes || echo no)
_HAS_HOOKS_JSON=$([ -f hooks/hooks.json ] && echo yes || echo no)
_HAS_WRITING_STYLE=$([ -f docs/ux/writing-style.md ] && echo yes || echo no)

_SKILLS_ON_DISK=$(ls -d skills/*/ 2>/dev/null | sed 's|skills/||;s|/$||' | sort)
_SKILL_COUNT=$(echo "$_SKILLS_ON_DISK" | grep -c . || echo 0)
_HOOK_FILES=$(ls hooks/ 2>/dev/null | grep -v '^hooks\.json$' | sort)
_HOOK_COUNT=$(echo "$_HOOK_FILES" | grep -c . || echo 0)

_TIMESTAMP=$(date -u +%Y-%m-%dT%H:%M:%SZ)
_REPORT_PATH="$_PLUGIN_ROOT/.ytstack/CONSISTENCY-REPORT.md"

echo "PLUGIN_ROOT: $_PLUGIN_ROOT"
echo "HAS_README: $_HAS_README"
echo "HAS_CONCEPT: $_HAS_CONCEPT"
echo "HAS_DECISIONS: $_HAS_DECISIONS"
echo "HAS_QUICKSTART: $_HAS_QUICKSTART"
echo "HAS_CLAUDE_MD: $_HAS_CLAUDE_MD"
echo "HAS_HOOKS_JSON: $_HAS_HOOKS_JSON"
echo "HAS_WRITING_STYLE: $_HAS_WRITING_STYLE"
echo "SKILL_COUNT: $_SKILL_COUNT"
echo "HOOK_COUNT: $_HOOK_COUNT"
echo "TIMESTAMP: $_TIMESTAMP"
echo "REPORT_PATH: $_REPORT_PATH"
echo "--- skills on disk ---"
echo "$_SKILLS_ON_DISK"
echo "--- hook files ---"
echo "$_HOOK_FILES"

Procedure

Step 2: HARD-GATE

If HAS_README=no OR HAS_CONCEPT=no OR HAS_DECISIONS=no OR HAS_HOOKS_JSON=no -- abort with:

"check-consistency requires README.md, docs/concept.md, .ytstack/DECISIONS.md, and hooks/hooks.json. Missing one of these means either the working directory is wrong or ytstack's own scaffolding is broken. Fix that first, then re-run."

Step 3: Read authoritative docs

Use the Read tool to load into working memory:

• README.md (full)
• QUICKSTART.md (full, if present)
• CLAUDE.md (full)
• docs/concept.md (full)
• docs/methodology.md (full)
• .ytstack/DECISIONS.md (full)
• docs/ux/writing-style.md (full, for em-dash rule and any residual style requirements)

Step 4: Read implementation inventory

For every directory in _SKILLS_ON_DISK, Read skills/<name>/SKILL.md frontmatter (name + description + triggers fields). Record in memory.

Read hooks/hooks.json (full).

Read skills/using-ytstack/SKILL.md (full) -- the meta-directive is there. Per DECISIONS 2026-04-24 "Skill selection is semantic, not keyword-based", it no longer contains a trigger-map.

Step 5: Run the nine consistency checks

For each check, record: check name, result (PASS / WARN / FAIL), findings list.

Check 1: Skill count drift

• Authoritative count: parse README §Status ("Ships: N skills ...") and extract N.
• Actual count: _SKILL_COUNT from preamble.
• If N != actual: WARN. Findings: "README claims N skills, disk has M."

Check 2: Wrapped-skill claims vs on-disk

• From docs/concept.md §3.1 table, extract the list of wrapped skill names.
• For each, verify skills/<name>/SKILL.md exists. Missing -> FAIL.
• From the README comparison table (the markdown table under §"Compared to using each framework alone"), cross-check each "via wrapper" row resolves to a skill that exists. Missing -> FAIL.

Check 3: Native-skill claims vs on-disk

• From docs/concept.md §3.4 native-skill list, verify every named skill exists in skills/.
• Missing -> FAIL. Extra on disk but not in §3.4 -> WARN "undocumented native skill".

Check 4: Explicit-skip claims

• README §"ytstack is the curation" text: regex-extract any skill names called out as "duplicates" or "conflicts with" or "skipped".
• Cross-check with docs/concept.md §3.2 table rows. Each README-skip must have a corresponding §3.2 row. Each §3.2 row must cite README. Mismatch -> WARN.

Check 5: Skill descriptions cover README examples semantically

• From skills/using-ytstack/SKILL.md, verify there is NO trigger-map table (per DECISIONS 2026-04-24 "Skill selection is semantic"). If one is found -> FAIL with "trigger-map reintroduced; skill selection must be description-based".
• From README.md §"What a session feels like", extract the bulleted "Say X -> skill Y" examples.
• For each README phrase, the target skill's description: field must semantically cover the intent (not via keyword list; via situational / contextual when-to-use prose). Manual review: read each target skill's description and confirm the phrase would semantically route there. WARN on any phrase whose target description does not obviously cover the intent -- the fix is to sharpen the description, never to re-add a keyword list.

Check 6: Hook inventory

• From hooks/hooks.json, list every registered hook.
• From _HOOK_FILES, list every script file in hooks/.
• Each JSON-registered hook must have a matching script file (or vice versa). Mismatch -> FAIL.
• docs/concept.md §3.4 hook list must include all 8 hooks (or whatever the actual count is). Drift -> WARN.

Check 7: Artifact list vs init-project

• From docs/concept.md §3.4 artifact list, extract the six-file list (PROJECT / DECISIONS / KNOWLEDGE / STATE / RUNTIME / PREFERENCES).
• Grep skills/init-project/SKILL.md for actual Write-tool templates. Each named artifact must have a template in init-project. Missing -> FAIL.

Check 8: Workflow skill references

• From docs/concept.md §5 (all three workflows), extract every ytstack skill name mentioned.
• Each name must resolve to skills/<name>/. Dangling reference -> FAIL.
• From QUICKSTART.md do the same.

Check 9: Em-dash punctuation hygiene

• Grep README.md, docs/concept.md, QUICKSTART.md, CLAUDE.md, .ytstack/PROJECT.md for the literal em-dash character (—, U+2014).
• Hits -> WARN with line reference. The em-dash is a concrete macOS autocorrect hazard (not a style preference) per docs/ux/writing-style.md; use -- instead.
• Meta-reference files that DESCRIBE the em-dash rule are exempt: docs/ux/writing-style.md, .ytstack/KNOWLEDGE.md, .ytstack/REVIEW-NOTES.md, .ytstack/DECISIONS.md.
• NO banned-vocabulary scan (rolled back 2026-04-24, see DECISIONS).

Step 6: Emit the report

Write .ytstack/CONSISTENCY-REPORT.md with this structure:

---
generated: {TIMESTAMP}
generator: ytstack:check-consistency v0.1.0
authoritative_source: README.md
---

# ytstack Consistency Report

## Summary

- Total checks: 9
- PASS: N
- WARN: N
- FAIL: N
- Overall: OK | DRIFT DETECTED | BLOCKED

## Check 1: Skill count drift

Status: PASS | WARN | FAIL

Findings:
- <bullet>
- <bullet>

(... one section per check ...)

## Actions suggested (human decides)

For each FAIL / WARN, a suggested reconciliation path. Do NOT auto-apply.

<list>

## Next steps

- Review this report
- Pick reconciliations
- For any architectural resolution, land a DECISIONS entry BEFORE editing docs or code
- Re-run `/ytstack:check-consistency` after reconciliation to verify

Step 7: Print a one-screen summary

To the user, print:

check-consistency run at {TIMESTAMP}. Report: .ytstack/CONSISTENCY-REPORT.md

• {N} PASS
• {N} WARN
• {N} FAIL

Top findings (first 3 FAILs or top 3 WARNs if no FAILs):

1. {finding}
2. {finding}
3. {finding}

Overall: {OK | DRIFT DETECTED | BLOCKED}. Open the report for the full list.

Step 8: Return control

Do NOT auto-invoke any reconciliation skill. Do NOT edit README, concept.md, DECISIONS.md, or any SKILL.md. The report is a reading surface; the human chooses what to fix and files the DECISIONS entries.

Terminal State

Return after writing the report and printing the summary. Valid next actions:

• User reviews the report
• User reconciles specific findings with explicit edits + DECISIONS entries
• User re-runs this skill to verify

Do NOT invoke /ytstack:plan-milestone, /ytstack:plan-task, or any other skill from here.

Why no auto-fix

Each drift finding could be caused by (a) a doc lagging behind code, (b) code lagging behind an approved doc change, or (c) an aspirational doc that was never going to ship as written. The skill cannot tell which. Auto-fixing either side without human judgment risks deleting work or silently locking in the wrong direction. Report-only, always.