@neikyun/ciel 6.11.3 → 6.13.0

package/assets/platforms/opencode/.opencode/commands/ciel-audit.md

@@ -1,327 +1,89 @@
 ---
-description: Audits the current Claude Code session for Ciel paradigm violations (missed Task dispatches, inline gathering, hook inactivity, skill overlaps, intent routing misses). Produces a structured report with a Ciel Health Score (0-100). If score < 90, creates a GitHub Issue on the Ciel repository with the findings and session timeline. Hook-independent — works even when Ciel hooks are broken.
+description: Audits the current session for Ciel v7 paradigm violations — dispatch discipline, hook activity, skill coverage, agent quality, memory health. Produces a structured report with Ciel Health Score (0-100). Creates GitHub Issue if score < 90. Hook-independent — works even when Ciel hooks are broken.
 ---
 
-# /ciel-audit — Session post-mortem
+# /ciel-audit — Session Post-Mortem
 
-*Generates a structured report of Ciel behavior violations observed in the current session. Calculates a Ciel Health Score (0-100). If the score is below 90, creates a GitHub Issue on the Ciel repository (github.com/KaosKyun/Ciel) with the full timeline and findings — otherwise produces the report only without creating an issue.*
+*Generates a structured report of Ciel behavior violations observed in the current session with a Ciel Health Score (0-100). If score < 90, creates a GitHub Issue on KaosKyun/Ciel.*
 
 Usage: `/ciel-audit`
 
-Runs inline in the main session. Does not dispatch agents. Does not depend on hooks being active.
+Runs inline. No agent dispatch. Works without hooks.
 
----
-
-## Instructions to the model
-
-You are auditing the **current conversation session** — the one you are participating in right now. Jump directly to the analysis. No preamble. No meta-commentary. No "I will now audit…".
-
-### What to audit
-
-Scan the session's tool-use history (your own prior turns). For each `/ciel <task>` invocation in this session, check the **eight dimensions** below. For each dimension, assign a severity and penalty score to calculate the final Ciel Health Score.
-
-#### Dimension 1: Dispatch discipline (critical) — penalty up to -25
-
-- Did the assistant emit a `Task(subagent_type="ciel-*")` within the **first 3 tool calls** after the `/ciel` prompt?
-- If NO: count the inline `Bash` / `Read` / `Grep` / `Glob` / `WebSearch` / `WebFetch` calls emitted in the main session before any dispatch. This is the v2.1.5 anti-pattern documented in `skills/ciel/SKILL.md:146-156`.
-- Exception: Trivial tasks (rename, typo, 1-line fix, docs-only) are allowed to run inline.
-- Severity→score:
-  - Critical dispatch violation (Standard/Critical task with no Task()): **-25**
-  - High (delayed dispatch, >1 inline tool before dispatch): **-15**
-  - OK (Task() within first 3 tool calls): **0**
-
-#### Dimension 2: Hook activity (critical) — penalty up to -25
-
-Search the session transcript for strings that Ciel hooks would have injected:
-
-- `"CIEL depth hint:"` — from `hooks/user-prompt-submit.sh:36`, injected via `additionalContext` on every UserPromptSubmit.
-- `"CIEL "` prefix (e.g., `"CIEL [CRITIQUE]"`, `"CIEL src/...` ) — from `hooks/pre-tool-write.sh:34,36`, injected before every Write/Edit.
-- Session banner from `hooks/session-start.sh` (SessionStart context).
-- `"META-CRITIQUER"` or similar end-of-session signal from `hooks/stop.sh`.
-
-- None found despite writes/edits: **-25** → hooks broken
-- Partial (some hooks firing but not all): **-10**
-- All present: **0**
-
-Most likely root cause: relative paths in `Ciel/settings.json:8,19,31,42,54,65,76`. The `command` field is written as `bash .claude/plugins/ciel/hooks/<file>.sh`, which Claude Code resolves against the current working directory, not the plugin directory.
-
-#### Dimension 3: Skill invocation coverage vs depth — penalty up to -15
-
-Cross-reference `skills/ciel/SKILL.md:20-64` (Depth Gauge) with what actually happened:
-
-- **Standard** task → `researcher` and `explorer` agents must be dispatched in parallel before FAIRE. Missing both: **-15**. Missing one: **-8**.
-- **Critical** task → must additionally invoke `stride-analyzer` and `security-regression-check`. Missing: **-15**.
-- Depth ambiguous and `depth-classifier` not invoked: **-5**.
-
-#### Dimension 4: Skill overlap / redundancy — penalty up to -10
-
-- Both `relire-critic` AND `critiquer-auditor` on the same diff: **-10**
-- `meta-critiquer` did not fire at end-of-task: **-5**
-- Same skill invoked 3+ times for same scope: **-5**
-
-#### Dimension 5: Agent report quality — penalty up to -5
-
-- Any `Task(subagent_type="ciel-*")` with returned message under 200 tokens: **-5** per occurrence (max -10)
+## Audit Dimensions (10)
 
-#### Dimension 6: Intent routing hits / misses — penalty up to -10
+Score starts at 100. Subtract penalties for each violation found.
 
-Scan the user's `/ciel` prompt text against the intent signals in `SKILL.md:79-94`. For each matched intent, verify the corresponding skill was invoked. Each miss: **-5** (max -10).
+### D1: Dispatch discipline (up to -25)
+Did the assistant dispatch `ciel-researcher` + `ciel-explorer` in parallel within the first 3 tool calls after the user prompt on Standard+ tasks?
+- Missing both on Standard/Critical: **-25**
+- Delayed dispatch (>1 inline tool before dispatch): **-15**
+- Exception: Trivial tasks (rename, typo) allowed inline.
 
-#### Dimension 7: npm version staleness — penalty up to -10
+### D2: Hook activity (up to -25)
+Search session transcript for Ciel hook injections:
+- `"CIEL depth hint:"` (UserPromptSubmit)
+- `"CIEL"` prefix on pre-write warnings
+- SessionStart banner
+- None found despite writes/edits: **-25** (hooks broken)
+- Partial (some firing): **-10**
 
-Check if a newer version of Ciel is available on npm:
+### D3: Skill coverage vs depth (up to -15)
+- Standard task: missing researcher+explorer dispatch: **-15** each
+- Critical task: missing stride-analyzer: **-15**
+- Depth ambiguous + depth-classifier not invoked: **-5**
 
-```bash
-NPM_VERSION=$(npm view @neikyun/ciel version 2>/dev/null || echo "unknown")
-LOCAL_VERSION=$(cat /path/to/VERSION 2>/dev/null || cat package.json 2>/dev/null | grep '"version"' | head -1 | cut -d'"' -f4 || echo "unknown")
-```
+### D4: Skill overlap / redundancy (up to -10)
+- Both relire-critic AND critiquer-auditor on same diff: **-10**
+- meta-critiquer did not fire at end-of-task: **-5**
 
-- npm version > local version: **-10**
-- Version check failed (npm not installed, no network): **0** (not a violation, just incomplete data)
-- Versions match or local is newer: **0**
+### D5: Agent report quality (up to -10)
+- Any agent return under 200 tokens: **-5** per occurrence (max -10)
 
-If npm version > local version, include an **Update notification** in the report:
-> A newer version of Ciel is available on npm: **{npm_version}** (installed: **{local_version}**). Run `ciel-update` or `npm update -g @neikyun/ciel` to upgrade.
+### D6: Intent routing (up to -10)
+Cross-reference user prompt intent signals with skills invoked. Each miss: **-5** (max -10).
 
-#### Dimension 8: Platform health — penalty up to -5
+### D7: Version staleness (up to -10)
+Check npm for newer `@neikyun/ciel` version. Remote > local: **-10**.
 
-Check that Ciel platform installations exist and are valid. Ciel currently supports two platforms: **claude** and **opencode**.
+### D8: Platform health (up to -5)
+Verify Claude Code (3+ agents, session-start.sh, settings.json) and OpenCode (plugin, 3+ agents, 5+ commands). Missing platform: **-3** to **-5**.
 
-**Claude Code** — check for expected agent and hook files:
-```bash
-CLAUDE_AGENTS=$(ls .claude/agents/ciel-*.md 2>/dev/null | wc -l | tr -d ' ')
-CLAUDE_HOOK=$(test -f .claude/hooks/session-start.sh && echo "1" || echo "0")
-CLAUDE_SETTINGS=$(test -f .claude/settings.json && echo "1" || echo "0")
-echo "Claude: agents=$CLAUDE_AGENTS hook=$CLAUDE_HOOK settings=$CLAUDE_SETTINGS"
-if [ "$CLAUDE_AGENTS" -ge 3 ] && [ "$CLAUDE_HOOK" = "1" ] && [ "$CLAUDE_SETTINGS" = "1" ]; then
-  echo "Claude platform: OK"
-else
-  echo "Claude platform: INCOMPLETE"
-fi
-```
-Expected: at least 3 agent files + session-start.sh + settings.json
+### D9: Memory health (up to -15)
+- `.ciel/memory/index.json` missing: **-10**
+- Episodes empty: **-5**
+- Auto-memory contamination (MEMORY.md newer than Ciel episodes): **-5**
 
-**OpenCode** — check for expected plugin, agent, and command files:
-```bash
-OPENCODE_PLUGIN=$(test -f .opencode/plugins/ciel.ts && echo "1" || echo "0")
-OPENCODE_AGENTS=$(ls .opencode/agents/ciel-*.md 2>/dev/null | wc -l | tr -d ' ')
-OPENCODE_COMMANDS=$(ls .opencode/commands/ciel*.md 2>/dev/null | wc -l | tr -d ' ')
-echo "OpenCode: plugin=$OPENCODE_PLUGIN agents=$OPENCODE_AGENTS commands=$OPENCODE_COMMANDS"
-if [ "$OPENCODE_PLUGIN" = "1" ] && [ "$OPENCODE_AGENTS" -ge 3 ] && [ "$OPENCODE_COMMANDS" -ge 5 ]; then
-  echo "OpenCode platform: OK"
-else
-  echo "OpenCode platform: INCOMPLETE"
-fi
-```
-Expected: ciel.ts plugin + at least 3 agent files + at least 5 command files
+### D10: Memory insight quality (up to -10)
+Run `python3 .claude/hooks/memory-engine.py analyze`. Check: promotion candidates, dead anchors, recursion drift, tag explosion.
 
-Scoring:
-- Both platforms fully present and valid: **0**
-- One platform missing or incomplete: **-3**
-- Both platforms missing or critically incomplete: **-5**
-
-**Important**: Do NOT check for `.claude/plugins/ciel/platforms/` or `.opencode/platforms/` directories — these are not part of the v6 architecture. Platform files are installed directly into `.claude/` and `.opencode/` respectively. Do NOT check for codex, cursor, kilocode, lmstudio, ollama, or windsurf — these platforms are not yet implemented.
-
----
-
-### Scoring
+## Scoring
 
 **Ciel Health Score** = 100 - sum(penalties)
 
-| Score range | Status | Issue created? |
-|-------------|--------|----------------|
+| Score | Status | Issue created? |
+|-------|--------|----------------|
 | 90-100 | Excellent | No |
-| 0-89 | Needs improvement | **Yes** — creates issue with timeline |
-
-The score is calculated automatically from the detected violations.
-
----
-
-### Report format
-
-Begin the output with the literal line `# Ciel Session Audit Report`. End with the literal line `**End of audit report.**` on its own line.
-
-```markdown
-# Ciel Session Audit Report
+| 0-89 | Needs improvement | Yes |
 
-**Date**: <today's date>
-**Ciel Health Score**: <N>/100 — <Excellent|Good|Needs improvement|Critical>
-**npm**: local v<X.Y.Z> | npm v<X.Y.Z> | <up-to-date|update available>
-**Platforms**: claude ✓ opencode ✓ (or ✗ if missing)
-**Session summary**: <N> /ciel invocation(s), <N> total tool calls, <N> Task() dispatches, <N> inline Bash/Read/Grep/WebSearch calls in main session.
+## Report Format
 
-**Verdict**: <PASS | VIOLATIONS FOUND>
+Output MUST start with `# Ciel Session Audit Report` and end with `**End of audit report.**`.
 
----
-
-## Violations detected
-
-### <N>. <Short violation name> — penalty: -<N>
-
-**Severity**: <critical | high | medium | low>
-**Evidence from session**
-- User prompt (turn N): `<exact prompt text, truncated to 200 chars>`
-- Assistant's first 3 tool calls after this prompt:
-  1. `<tool_name>(<brief args>)`
-  2. `<tool_name>(<brief args>)`
-  3. `<tool_name>(<brief args>)`
-- Expected: `Task(subagent_type="ciel-<role>", ...)` as first tool call
-- Observed: `<actual behavior>`
-
-**Root cause hypothesis**
-<one or two sentences>
-
-**Proposed fix**
-- <concrete edit with file:line>
-
-...
-
----
-
-## Update notification
-
-<If npm version > local version, show update notification here. Otherwise omit this section.>
-
----
-
-## Scoring breakdown
-
-| Dimension | Penalty |
-|-----------|---------|
-| D1 — Dispatch discipline | -<N> |
-| D2 — Hook activity | -<N> |
-| D3 — Skill coverage vs depth | -<N> |
-| D4 — Skill overlap | -<N> |
-| D5 — Agent report quality | -<N> |
-| D6 — Intent routing | -<N> |
-| D7 — npm version | -<N> |
-| D8 — Platform health | -<N> |
-| **Total** | **-<N>** |
-| **Health Score** | **<N>/100** |
-
----
-
-## Summary of fixes to apply
-
-Apply these in order. Each points to a file:line and a concrete change.
-
-1. **<Fix name>** — `<path>:<line>` — <one-line description>
-2. ...
-
-After each fix: re-run the scenario that triggered the violation in a fresh Claude session to verify.
-
----
-
-**End of audit report.**
-```
-
-### If no violations are found (score = 100)
-
-Output a single short section — **no issue is created** for PASS verdicts:
-
-```markdown
-# Ciel Session Audit Report
-
-**Date**: <today's date>
-**Ciel Health Score**: 100/100 — Excellent
-**npm**: local v<X.Y.Z> | npm v<X.Y.Z> | up-to-date
-**Platforms**: claude ✓ opencode ✓
-**Session summary**: <N> /ciel invocation(s), <N> tool calls, <N> Task() dispatches.
-**Verdict**: PASS
-
-**No violations found.** No issue created.
-
-**End of audit report.**
-```
-
----
-
-### GitHub Issue creation (only if score < 90)
-
-If the Ciel Health Score is **below 90**, create a GitHub Issue with the report AND the session timeline.
-
-**Important**: Do NOT create an issue if score >= 90. Only create for scores < 90.
-
-1. **Check for duplicate issues first**:
-   ```bash
-   EXISTING=$(gh issue list --repo KaosKyun/Ciel --label audit --state open \
-     --json title --jq '.[].title' 2>/dev/null | \
-     grep -c "^\[CIEL-AUDIT\] $(date +%Y-%m-%d) -" || true)
-   if [ "$EXISTING" -gt 0 ]; then
-     echo "Warning: today's audit issue already exists. Skipping creation."
-     echo "View existing at: https://github.com/KaosKyun/Ciel/issues?q=is%3Aopen+label%3Aaudit"
-     exit 0
-   fi
-   ```
-
-2. **Save the report to a temp file**: Write the complete report text to a temp file:
-   ```bash
-   REPORT_FILE="/tmp/ciel-audit-report-$(date +%Y%m%d).md"
-   cat > "$REPORT_FILE" << 'REPORT_EOF'
-   # Ciel Session Audit Report
-   ...
-   **End of audit report.**
-   REPORT_EOF
-   ```
-
-3. **Determine the repository**: run `gh repo view --json owner,name` to confirm, or default to `KaosKyun/Ciel`.
-
-4. **Create the issue** with `gh issue create`, using Python to avoid shell quoting issues:
-   ```bash
-   python3 -c "
-   import subprocess, sys
-   ymd, date_str, score, verdict = sys.argv[1:5]
-   report_path = f'/tmp/ciel-audit-report-{ymd}.md'
-   with open(report_path) as f:
-       body = f.read()
-   subprocess.run(['gh', 'issue', 'create',
-       '--repo', 'KaosKyun/Ciel',
-       '--title', f'[CIEL-AUDIT] {date_str} - {verdict} (Score: {score}/100)',
-       '--label', 'audit,ciel',
-       '--body', body])
-   " "$(date +%Y%m%d)" "$(date +%Y-%m-%d)" "<score>" "<verdict>"
-   ```
-
-5. **Include the session timeline** in the issue body. Before the report, add a timeline section listing all `/ciel` invocations in chronological order:
-   ```markdown
-   ## Session Timeline
-   
-   | Time | Event |
-   |------|-------|
-   | T+N  | /ciel <prompt excerpt> |
-   | T+N  | Write/Edit on <file> |
-   | T+N  | Task() dispatch to <agent> |
-   ...
-   
-   ---
-   
-   <full audit report>
-   ```
-
-6. **Handle errors gracefully**:
-   - If `gh` is not available: print a message telling the user to install (`brew install gh`) and print the report to stdout instead.
-   - If the repository can't be reached: skip issue creation, print a warning, but still output the report.
-
-7. **After creating the issue**: include the issue URL at the end of your response so the user can open it directly.
-
-**Title format**: `[CIEL-AUDIT] YYYY-MM-DD - VIOLATIONS FOUND (Score: <N>/100)`
-
-**Labels**: `audit`, `ciel`
-
----
+For violations: provide evidence (turn number, tool calls), root cause hypothesis (file:line), and proposed fix (concrete edit). No preamble, no meta-commentary, no emoji.
 
-### Tone and length
+## GitHub Issue (score < 90 only)
 
-- **Mechanically actionable, not narrative**. File paths, line numbers, before/after snippets.
-- Target length: 400–800 lines of markdown for a session with 2–3 violations. Shorter if PASS.
-- No rhetorical preamble. No meta-commentary about the audit process itself. No emoji. No closing remarks after the `**End of audit report.**` marker.
+1. Check for duplicate: `gh issue list --repo KaosKyun/Ciel --label audit --state open`
+2. Create with title: `[CIEL-AUDIT] YYYY-MM-DD - VIOLATIONS FOUND (Score: N/100)`
+3. Labels: `audit`, `ciel`
+4. Body: session timeline + full audit report
+5. Handle gracefully if `gh` unavailable — print report to stdout
 
-### What NOT to do
+## Constraints
 
-- Do NOT fix violations in the current session. Only produce the report and optionally create the issue.
-- Do NOT invoke other Ciel skills. This command is fully self-contained.
-- Do NOT dispatch `Task()` agents. Audit happens inline.
-- Do NOT ask clarifying questions. Produce the report with the information you have.
-- Do NOT create an issue if score >= 90. Only create for score < 90.
-- Do NOT create duplicate issues — run the `gh issue list` check before creating.
-- Do NOT restart, rerun, or attempt to fix the session in-flight. The audit report is the deliverable.
+- Do NOT fix violations in-session. Report only.
+- Do NOT dispatch agents. Inline only.
+- Do NOT create issues for score >= 90.
+- Do NOT create duplicate issues.
+- Target length: 200-400 lines for violations, 50 for PASS.

package/assets/platforms/opencode/.opencode/commands/ciel-create-skill.md

@@ -1,50 +1,40 @@
 ---
-description: ---
-agent: ciel-improver
-subtask: true
+description: Generates a valid Ciel v7 SKILL.md scaffold — kebab-case name, YAML frontmatter with triggers.path, 3-section body (Checklist / Anti-patterns / Patterns), max 200 lines.
 ---
 
-> **OpenCode note**: This command requires `claude --print` headless mode for full functionality (binary evals, skill scaffold generation). On OpenCode it runs in degraded mode — the improver agent returns proposals only. For the full harness, use Claude Code.
+# /ciel-create-skill — Create a new Ciel v7 skill
 
----
-description: Generates a valid Ciel SKILL.md scaffold following Anthropic Skills-first rules (kebab-case ≤64, YAML description ≤1536, body ≤500 lines).
----
-
-# /ciel-create-skill — Create a new Ciel skill
-
-*Generates a valid SKILL.md scaffold following Anthropic Skills-first rules (kebab-case name ≤64 chars, YAML frontmatter ≤1536-char description, ≤500-line body, progressive disclosure to one reference.md).*
+Generates a valid SKILL.md scaffold following Ciel v7 format: 3 sections (Checklist, Anti-patterns, Patterns), kebab-case name, optional path trigger.
 
-Usage: `/ciel-create-skill <name> <purpose>`
+Usage: `/ciel-create-skill <name> <domain-description>`
 
-- `name` — kebab-case, max 64 chars, unique across `skills/`
-- `purpose` — one-line description of what the skill does
+- `name` — kebab-case, max 64 chars, unique across `.claude/skills/`
+- `domain-description` — one-line description of what domain expertise this skill encodes
 
 ---
 
 ## What it does
 
-1. Dispatches the `improver` agent in MODE=CREATE-SKILL
-2. The agent invokes `skill-creator` skill
-3. Validates name, category (you pick), description length, uniqueness, paths glob
-4. Generates a scaffold SKILL.md + optional reference.md
-5. Returns the proposed files for user review
-6. On approval, writes files + registers in `skills/ciel/reference.md` catalog
+1. Dispatches `ciel-improver` agent with MODE=CREATE-SKILL
+2. Validates name (kebab-case, uniqueness, no reserved words)
+3. Helps you define the path trigger (if any) and anti-patterns
+4. Generates SKILL.md scaffold in v7 3-section format
+5. On approval, writes to `.claude/skills/<name>/SKILL.md`
+6. Optionally runs `sync-skills.sh` to distribute to mirrors
 
 ---
 
 ## Example
 
 ```
-/ciel-create-skill kotlin-coroutines-mastery "Expert in Kotlin coroutines: structured concurrency, Flow operators, cancellation, testing. Use when working with suspend functions, CoroutineScope, Flow, or coroutine builders."
+/ciel-create-skill kotlin-coroutines "Kotlin coroutines — structured concurrency, Flow operators, cancellation, testing. Use when working with suspend functions, CoroutineScope, Flow."
 ```
 
 Expected output:
-1. Validation: ✓ name valid, unique, no reserved words
-2. Category suggestion: `domain` (based on "Expert in X" pattern)
-3. Preview of `skills/domain/kotlin-coroutines-mastery/SKILL.md` (~150 lines)
-4. Preview of `skills/domain/kotlin-coroutines-mastery/reference.md` (~300 lines) with Flow operators cheatsheet
-5. Catalog entry: `| kotlin-coroutines-mastery | Kotlin files with suspend/Flow |`
-6. Approve? [y/n/edit]
+1. Validation: name valid and unique
+2. Path trigger suggestion: `**/*.kt,**/*.kts`
+3. Preview of `SKILL.md` scaffold (~100-150 lines)
+4. Approve? [y/n/edit]
 
 ---
 
@@ -53,33 +43,21 @@ Expected output:
 - kebab-case only (no underscores, no camelCase)
 - Max 64 chars
 - No reserved words: `anthropic`, `claude`, `mcp`
-- Must be unique across `skills/**/SKILL.md`
-- Warn if starts with category name (e.g. `workflow-foo` in `workflow/` is redundant)
-
----
-
-## Category decision
-
-- `workflow` — enforces a CRÉER/CRITIQUER/META-CRITIQUER step
-- `research` — finds information outside the codebase
-- `domain` — encodes expertise in a specific tech or pattern family
-- `utility` — wraps a frequent mechanical operation
-- `meta` — modifies Ciel itself
+- Must be unique across `.claude/skills/`
 
 ---
 
 ## Guardrails
 
 - **Max 1 new skill per invocation** — prevents skill explosion
-- **SKILL.md size**: ≤ 300 lines (hard cap)
-- **reference.md size**: ≤ 500 lines (hard cap)
-- **Duplication detection**: warns if description overlaps ≥ 70% with existing skill
-- **Never creates**: files in `.claude-plugin/`, agents, hooks, commands — those use different patterns
+- **SKILL.md max 200 lines** — v7 hard cap
+- **Duplication detection**: warns if description overlaps >= 70% with existing skill
+- **Never creates**: agent definitions, hooks, commands — those use different patterns
 
 ---
 
 ## After creation
 
-1. Test the skill: `/ciel-eval <name>` (runs baseline eval if dataset exists)
-2. If it's a `workflow` skill, update `skills/ciel/SKILL.md` pipeline sections to reference it
-3. Commit with message `feat(ciel): add <category>/<name> skill`
+1. Test the skill on a relevant task — does it load when expected?
+2. Commit with message: `feat(skills): add <name> skill`
+3. Run `bash scripts/sync-skills.sh` to distribute to all mirrors

package/assets/platforms/opencode/.opencode/commands/ciel-eval.md

@@ -1,58 +1,42 @@
 ---
-description: ---
-agent: ciel-improver
-subtask: true
+description: Runs the binary eval dataset for a Ciel skill via headless Claude Code, comparing baseline scores and persisting results.
 ---
 
-> **OpenCode note**: This command requires `claude --print` headless mode for full functionality (binary evals, skill scaffold generation). On OpenCode it runs in degraded mode — the improver agent returns proposals only. For the full harness, use Claude Code.
+# /ciel-eval — Run skill eval harness
 
----
-description: Runs the binary eval dataset for one or all Ciel skills via claude --print headless, comparing variants.
----
-
-# /ciel-eval — Run eval harness
-
-*Runs the binary eval dataset for one skill (or all skills) via `claude --print` headless mode, comparing variants and persisting scoreboards.*
+Runs the binary eval dataset for a skill via headless Claude Code, persisting scoreboards to `~/.ciel/evals/results/`.
 
 Usage: `/ciel-eval [skill-name]`
 
-- No arg → runs baseline eval for ALL skills that have a dataset
-- `skill-name` → runs eval for that skill only (baseline + any variants in `variants/`)
+- No arg — runs baseline eval for ALL skills that have a dataset
+- `skill-name` — runs eval for that skill only
 
 ---
 
 ## What it does
 
-1. Dispatches the `improver` agent in MODE=EVAL
-2. The agent invokes `skill-variant-evaluator` skill
-3. For each target skill:
+1. Dispatches `ciel-improver` agent in MODE=EVAL
+2. For each target skill:
    - Loads dataset from `evals/datasets/<skill-name>.jsonl`
-   - If variants exist (`skills/<category>/<name>/variants/*.md`), evaluates each
-   - Otherwise, baseline-only scoring
-4. Persists results to `evals/results/<skill-name>-<timestamp>.json`
-5. Returns a scoreboard
+   - Runs each entry via `claude --print` headless
+   - Scores against expected behaviors
+3. Persists results to `~/.ciel/evals/results/<skill-name>-<timestamp>.json`
+4. Returns a scoreboard
 
 ---
 
 ## Example output
 
 ```
-# Eval results — flux-narrator
+# Eval results — api-design
 
-Dataset: evals/datasets/flux-narration.jsonl (8 entries)
-Variants evaluated: 3
+Dataset: evals/datasets/api-design.jsonl (12 entries)
 
-| Variant | Score | Tokens | Duration |
-|---------|-------|--------|----------|
-| A (baseline) | 0.72 | 14.5k | 12.5s |
-| B (tightened) | 0.89 | 15.1k | 13.2s ← WINNER |
-| C (reduced) | 0.81 | 12.8k | 11.7s |
+Score: 0.85 (25/28 criteria passed)
+Tokens used: 18.2k
+Duration: 15.3s
 
-Winner: **Variant B** (+0.17 over baseline)
-
-Recommendation: adopt Variant B via `/ciel-improve`
-
-Result persisted: evals/results/flux-narrator-2026-04-17-142345.json
+Result persisted: ~/.ciel/evals/results/api-design-2026-05-23-142345.json
 ```
 
 ---
@@ -60,29 +44,25 @@ Result persisted: evals/results/flux-narrator-2026-04-17-142345.json
 ## Guardrails
 
 - **Max 20 eval entries per dataset** — prevents runaway costs
-- **Max 3 variants per skill** — A/B/C, more is noise
-- **Cost estimation** — displayed before starting; user can abort if > 500k tokens projected
-- **Read-only evals** — `--disallowed-tools "Write Edit NotebookEdit"` to prevent side effects
-- **Graceful fallback** — if `claude --print` is unavailable, outputs manual eval prompts and waits for user input
+- **Cost estimation** — displayed before starting; abort if > 500k tokens projected
+- **Read-only**: `--disallowed-tools "Write Edit NotebookEdit"` to prevent side effects
+- **Graceful fallback**: if `claude --print` unavailable, outputs manual eval prompts
 
 ---
 
 ## When to run
 
-- Before and after every `/ciel-improve` pass (baseline vs improved)
-- On every CHANGELOG version bump — regression sanity check
-- When creating a new skill — establish baseline score
-- When adding new eval criteria — re-score existing skills
+- Before and after `/ciel-improve` — baseline vs improved
+- On VERSION bump — regression sanity check
+- When creating a new skill — establish baseline
 
 ---
 
 ## Eval dataset format
 
-One JSON object per line (JSONL). See `evals/datasets/depth-classification.jsonl` for examples. Required fields:
+One JSON object per line (JSONL). Required fields:
 
 - `id` — unique in dataset
 - `input` — prompt / context passed to skill
-- `expected_behavior` — object mapping criterion name → boolean
+- `expected_behavior` — object mapping criterion name to boolean
 - `skill` — which skill this tests
-
-See `skills/meta/skill-variant-evaluator/reference.md` for the full schema and supported criterion evaluators.