honne — 7-Axis Self-Observation

When invoked, execute Step 1 through Step 6 in order immediately. Do not summarize the skill or ask what the user wants — invocation itself is the request. Start by asking the Step 1 question.

Step 1: Scope + Locale HITL

Invoke AskUserQuestion tool with two questions in a single call:

(a) Scope:

• question: "Scan scope?"
• options: [{"label":"repo","description":"current project only"},{"label":"global","description":"all projects"}]

(b) Locale:

• question: "Locale?"
• options: [{"label":"ko","description":"한국어"},{"label":"en","description":"English"},{"label":"jp","description":"日本語"}]

Set SCOPE and LOCALE from the two replies. Do not use plain-text Q&A — arrow-key selection only.

Step 2: Scan

Run: bash "${CLAUDE_PLUGIN_ROOT}/scripts/honne" scan --scope "$SCOPE" --cache ".honne/cache/scan.json" Capture RUN_ID from result: RUN_ID=$(python3 -c 'import json; print(json.load(open(".honne/cache/scan.json"))["run_id"])') Non-zero exit → output stdout+stderr verbatim to user, stop. Do not interpret exit codes.

Step 3: Rejection reframe filter (skip candidate)

For each axis, run: bash "${CLAUDE_PLUGIN_ROOT}/scripts/honne" query --base-dir ".honne" --tag "<axis>" --type rejection --scope "$SCOPE" Before Step 4 records each axis, pipe the candidate through bash "${CLAUDE_PLUGIN_ROOT}/scripts/honne" axis validate --text "$candidate" --locale "$LOCALE" --skip-if-overlaps "$rejection_text" — exit 3 = overlap, skip and log "reframed". 모든 변수는 큰따옴표 인용 필수(공백·특수문자 안전). LLM 호출 없음.

Recording rejections: If the user explicitly says "n" or rejects a candidate claim for any axis, record it as a rejection so Step 3 can filter it in future runs:

bash "${CLAUDE_PLUGIN_ROOT}/scripts/honne" record claim \
  --type rejection --axis "$axis" --scope "$SCOPE" \
  --claim "$CANDIDATE" --run-id "$RUN_ID" \
  --out ".honne/assets/rejections.jsonl"

<!-- TODO(evolutions): evolutions.jsonl cross-run diff tracking is not yet implemented. query --type evolution always returns []. Structural change required. -->

Step 4: Per-axis autonomous record

For each axis from axis list, run each command separately — do NOT bundle into a script file or use heredocs:

bash "${CLAUDE_PLUGIN_ROOT}/scripts/honne" axis run "$axis" \
  --locale "$LOCALE" --scan .honne/cache/scan.json > ".honne/cache/axis-${axis}.json"

python3 -c "import json,sys; d=json.load(open('.honne/cache/axis-${axis}.json')); sys.exit(0 if d.get('insufficient_evidence') else 1)"

If exit 0 → skip this axis (insufficient evidence), continue to next.

python3 -c "import json; print(json.load(open('.honne/cache/axis-${axis}.json'))['candidate_claim'])"

Capture stdout as CANDIDATE.

bash "${CLAUDE_PLUGIN_ROOT}/scripts/honne" record claim \
  --type claim --axis "$axis" --scope "$SCOPE" \
  --claim "$CANDIDATE" --run-id "$RUN_ID" \
  --quotes-file ".honne/cache/axis-${axis}.json" \
  --out ".honne/assets/claims.jsonl"

HARD RULE — execution constraints (test suite enforces):

• Each bash block runs as a direct shell command — no heredocs (<< 'EOF'), no script files, no command bundling.
• No intermediate writes to /tmp — use .honne/cache/ instead. Writing to /tmp is a SKILL.md contract violation.

Step 5: LLM narrative synthesis

Invoke Claude (your own mental reasoning) to synthesize explanations and a one-liner:

(a) Read synthesis prompt: Read "${CLAUDE_PLUGIN_ROOT}/skills/whoami/templates/synthesis_prompt.${LOCALE}.md"

(b) Build USER_PAYLOAD from the claims recorded in Step 4. You already have the AXIS_JSON outputs in memory — construct the payload directly as a JSON object without re-reading files:

USER_PAYLOAD = {
  "locale": "<LOCALE>",
  "claims": {
    "<axis>": {"claim": "<CANDIDATE>", "evidence_count": <len(quotes)>} for each recorded axis,
    "<skipped_axis>": null for each axis that had insufficient evidence
  }
}

Do NOT use python3 << 'PYEOF' or any heredoc to build this payload. Assemble it in your mental context from the Step 4 outputs already known.

(c) Synthesize: Apply synthesis_prompt system instructions to yourself + USER_PAYLOAD as user input. Produce STRICT JSON response.

(d) Resolve the absolute path first:

python3 -c "import os; print(os.path.join(os.getcwd(), '.honne/cache/narrative.json'))"

Capture stdout as NARRATIVE_PATH. Then: Write the JSON response to the resolved path. If JSON parse fails or response is empty, skip saving.

Step 6: Render persona and report

date -u +%Y-%m-%dT%H:%M:%SZ

Capture stdout as NOW.

bash "${CLAUDE_PLUGIN_ROOT}/scripts/honne" render persona \
  --claims .honne/assets/claims.jsonl \
  --scope "$SCOPE" --locale "$LOCALE" --run-id "$RUN_ID" --now "$NOW" \
  --narrative .honne/cache/narrative.json \
  --out .honne/persona.json

bash "${CLAUDE_PLUGIN_ROOT}/scripts/honne" render report \
  --persona .honne/persona.json --locale "$LOCALE" --out docs/honne.md

Completion

Report saved files to .honne/persona.json and docs/honne.md. Use /honne:compare to review past observations.

Output the following next action suggestions to the user:

Next actions

• /honne:persona — generate two personas (antipattern × signature) from this profile
• /honne:crush <topic> — stage a live debate between the two personas