ccc-safety · diff
git:20260728.38c269b to git:20261013.fbe187d
66 added, 34 removed. Audit A to A.
---
name: ccc-safety
- description: "Safety deck: what Commander's hooks caught for you — dangerous actions blocked, auto-fixes applied, and where your tools fail most. A self-contained artifact."
+ description: "Safety deck: dangerous actions Commander's hooks blocked, auto-fixes applied, and the tools that fail most, narrated here or published as a private snapshot page after consent. Use when the user asks what the guardrails caught, why tools fail, or types $ccc-safety."
allowed-tools:
- Read
- Bash
- Write
- AskUserQuestion
- argument-hint: "[open | status]"
+ - Artifact
+ argument-hint: "[status | open | snapshot]"
+ model: sonnet
+ effort: low
---
- # $ccc-safety — Safety
+ # $ccc-safety — what the guardrails caught
- The guardrails deck. CC Commander's hooks already journal every permission decision (`~/.claude/commander/analytics/permission-gate.jsonl`) and every tool failure (`~/.claude/commander/tool-failures.jsonl`). This turns them into one page: how many dangerous actions got blocked, how many were auto-fixed, and which tools/errors trip you up most (a Sniffly-style breakdown).
+ **CC Commander** · $ccc-safety · blocked actions, auto-fixes and failure hotspots from local logs.
- > 🔒 Privacy: read locally. Publishing uploads the artifact (counts, tool names, redacted error signatures) to your private claude.ai artifact URL — private to your account, but it leaves the machine. Always ask before publishing; never publish automatically. Error text is redacted (sk-/Bearer/token patterns → [redacted]) before it ever reaches the page.
+ Sources: `~/.claude/commander/analytics/permission-gate.jsonl` (every permission decision) and `~/.claude/commander/tool-failures.jsonl` (every tool failure). Error text is redacted (`sk-`, `Bearer`, token patterns become `[redacted]`) before it reaches any output.
- ## How it routes
+ ## When to use / when not to
- On `$ccc-safety` with no argument, open a click-first picker:
+ - Use for "what did the hooks stop?", "why does Bash keep failing?", or a shareable guardrail page.
+ - Use `$ccc-console safety` for the same tab inline; `$ccc-doctor` when hooks seem not to be running at all.
- ```
- AskUserQuestion:
- question: "Safety — how do you want it?"
- options:
- - 🛰️ Publish the deck — a living Safety artifact: blocked actions, auto-fixes, failure hotspots
- - 🗣️ Plain-English status — narrate the headline numbers here, no jargon
- ```
+ ## What happens
- `$ccc-safety open` (or `snapshot`) → publish the deck. `$ccc-safety status` → narrate.
+ ### 1. Pick the output
- **Zero-state:** if `permission-gate.jsonl` and `tool-failures.jsonl` are both missing/empty, say so plainly — "No guardrail activity logged yet — the hooks fill this in as you work." Never render an error for an empty deck.
+ With no argument, ask with AskUserQuestion (Codex: a plain-text question):
- ### 🛰️ Publish the deck
+ - 🟢 Plain-English status — the headline numbers here, nothing leaves the machine
+ - Publish the deck — a private claude.ai page; asks before uploading
+ - Show it inline — runs `$ccc-console safety`
- **LIVING PATTERN:** always render to the SAME file path (`scratchpad/ccc-safety-live.html`), then republish that same path with the Artifact tool — same path, same URL, one living page. **Do not change this path** — it is the identity of every bookmark anyone already has for this deck.
+ `status` goes to step 2. `open` or `snapshot` goes to step 3.
- Since v7.4.0 this page is the **Safety tab of the Commander Console**, published on its own — one builder renders every deck and the console, so the numbers can't disagree:
+ If both logs are missing or empty, say "No guardrail activity logged yet. The hooks fill this in as you work." and stop. That is not an error.
+ ### 2. Plain-English status
+
```bash
- mkdir -p scratchpad
- if [ -f "${CLAUDE_PLUGIN_ROOT}/scripts/build-console.mjs" ]; then
- node "${CLAUDE_PLUGIN_ROOT}/scripts/build-console.mjs" \
- --surface artifact --tab safety --out scratchpad/ccc-safety-live.html
- else
- node commander/cowork-plugin/scripts/build-console.mjs \
- --surface artifact --tab safety --out scratchpad/ccc-safety-live.html
- fi
+ node --input-type=module -e "
+ const { readSafetyModel } = await import('${CLAUDE_PLUGIN_ROOT}/lib/safety-snapshot.js');
+ const m = await readSafetyModel({});
+ console.log(JSON.stringify({ decisions: m.decisions.counts, failures: m.toolFailures.total, byTool: m.toolFailures.byTool.slice(0, 3) }));
+ "
```
- The page is unchanged apart from one added line pointing at `$ccc-console` (the inline all-in-one view).
+ Narrate it in two sentences: "Commander stopped N risky actions and auto-fixed M. Bash fails most, usually on timeouts." Sum the `blocked` kinds for N and the `autofixed` kind for M.
- **First publish this session** → ask before publishing (counts, tool names, redacted error signatures leave the machine for your private artifact URL). **Every later run of `$ccc-safety`** → invoking the skill again IS the refresh consent: republish the same file path to the same URL without re-asking. Publish `scratchpad/ccc-safety-live.html` with the Artifact tool (favicon 🛡️). End with the artifact title + the headline: "blocked N dangerous actions, auto-fixed M." To refresh later: run `$ccc-safety` again — same URL updates in place; viewers reload.
+ ### 3. Publish the deck (consent-gated)
- ### 🗣️ Plain-English status
+ 1. First publish this session: ask with AskUserQuestion.
+ - Question: "Publish the Safety deck as a private claude.ai page? Decision counts, tool names and redacted error signatures leave this machine."
+ - 🟢 Publish — create or update the private page
+ - Narrate here instead — nothing uploads
+ - Cancel
+ Publish only on an explicit yes. Running `$ccc-safety` again later in this session is the refresh consent.
+ 2. Build the Safety tab of the console onto this deck's fixed path:
+ ```bash
+ mkdir -p "<scratchpad>"
+ if [ -f "${CLAUDE_PLUGIN_ROOT}/scripts/build-console.mjs" ]; then
+ node "${CLAUDE_PLUGIN_ROOT}/scripts/build-console.mjs" \
+ --surface artifact --tab safety --out <scratchpad>/ccc-safety-live.html
+ else
+ node commander/cowork-plugin/scripts/build-console.mjs \
+ --surface artifact --tab safety --out <scratchpad>/ccc-safety-live.html
+ fi
+ ```
+ Exit 0 and `build-console: wrote …` on stderr mean the page is ready. The page carries the Commander decks strip and a line pointing to `$ccc-console`.
+ 3. Publish with the Artifact tool:
+ - Already published from this path in this session: publish the same `file_path`; omit `icon`.
+ - Otherwise: `action: "list"`, find "Commander Safety", `action: "read"` its `url`, then publish `file_path` with that `url`. If none exists, publish `file_path` with `icon: "shield"` and `description: "Actions Commander's hooks blocked or auto-fixed, and where tools fail most."`.
+ - On Codex or without the Artifact tool: give the user the HTML file path.
+ 4. Reply with the page title, the URL, and the headline "blocked N risky actions, auto-fixed M".
- Run `readSafetyModel({})` and read the decision counts + top failing tool, then narrate caveman-simple: "Commander stopped N risky actions and quietly fixed M. Your Bash commands fail most — usually timeouts." No jargon.
+ ## Done when
- ## The deck also links the others
+ - Status: the numbers you narrated came from the step 2 command output in this turn.
+ - Publish: the build exited 0 and the Artifact tool returned a URL, which you reported.
- The published artifact carries the shared **Commander decks** strip at the top — one click (copy) away from the Cockpit, Mission Control, and Usage & Cost decks. Users always know the other decks exist. It also carries one line pointing at `$ccc-console`, the inline console that shows all of these tabs at once (and has a prompt bar, which a published page cannot).
+ ## Guardrails
- > **On Codex:** some panels may be empty on Codex today — Safety's failure hotspots, Mission Control's agent roster, and Usage's savings hero are fed from Claude-only hooks right now. A follow-up workstream wires the matching Codex telemetry.
+ - Never publish without the step 3 consent, and never on a schedule or as a side effect.
+ - Never quote unredacted error text; use the model's redacted samples only.
+ - Keep the path `<scratchpad>/ccc-safety-live.html`; a new path is a new URL.
+ - The published page is a snapshot: it changes only when republished and cannot send prompts into the session.
+ - On Codex, failure hotspots can be thin: they are fed by Claude-side hooks.
+
+ ## See also
+
+ - `$ccc-console` — this tab inline, with the prompt bar
+ - `$ccc-usage` — what the work cost
+ - `$ccc-doctor` — verify the hooks are wired
---
> ⚙️ **Fable contract:** plan before build · verifier ≠ worker · prove before alarm · loops need gates · leave durable state — `rules/fable-method.md`
> (On Codex, present these options as a numbered list and ask the user to reply with a number — AskUserQuestion is Claude-only.)