claude-tweaks:assess-agent-autonomy

Interaction style: Present single decisions via the AskUserQuestion tool (options with one marked Recommended) instead of a plain-text numbered list. For multi-item decisions, render a batch table with recommended actions pre-filled, then capture the apply-all/override decision via one AskUserQuestion call. Never make more than one AskUserQuestion call per logical decision — resolve each before showing the next. End skills with a ## Next Actions block rendered via AskUserQuestion (context-specific options, one recommended), not a navigation menu.

Assess Agent Autonomy — Content-Aware Trust Verdicts

Three-mode inline helper that replaces mechanical label lookups with judgment read from actual record/diff/failure content. Never invoked directly by a human — always a component step inside /claude-tweaks:triage or /claude-tweaks:dispatch:

/claude-tweaks:triage Step 2        [ grant-check ]  -> RECOMMEND_BUILD / RECOMMEND_MERGE
/claude-tweaks:dispatch Auto-merge  [ merge-check ]   -> VERDICT: auto-merge | needs-human
/claude-tweaks:dispatch Settle      [ failure-check ] -> CLASSIFICATION + NOTIFY_NOW

When to Use

• /claude-tweaks:triage's Step 2 needs a grant recommendation for a worklist record.
• /claude-tweaks:dispatch's Auto-merge gate needs a merge-or-human verdict for a clean, reviewed run.
• /claude-tweaks:dispatch's Settle step needs to classify why a run failed.

Not for: granting auto:build/auto:merge (still /claude-tweaks:triage's human-confirmed job), merging anything itself (/claude-tweaks:dispatch acts on the verdict), or any decision outside these three call sites — this is not a general-purpose risk service.

Input

$ARGUMENTS is {mode} #{n}, where mode is one of grant-check | merge-check | failure-check and #{n} is the record's issue number (used to fetch the record body for grant-check; used for reference/logging in merge-check/failure-check's rendered output).

Invoked inline via the Skill tool — not as a fresh Task-agent dispatch. The calling agent (a human-driven /claude-tweaks:triage session, or dispatch's per-group Task agent running /flow) runs this skill's procedure in its own context and reads the produced report directly; there is no cross-process hand-off.

Mode: grant-check

Called from: /claude-tweaks:triage's Step 2, once per worklist record, every triage session — never pre-filtered to "borderline" records.

Step 1: Gather

gh issue view "$N" --json body,labels -q '{body: .body, labels: [.labels[].name]}' > /tmp/assess-grant-${N}.json

Read the record's full body (Current State / Deliverables / Acceptance Criteria) from the fetched JSON. Extract the current risk:*/effort:* labels, if present:

node -e "const {extractRiskEffort}=require(process.env.CLAUDE_PLUGIN_ROOT+'/bin/lib/issues/tier.js');
  const d=require('/tmp/assess-grant-${N}.json');
  console.log(JSON.stringify(extractRiskEffort(d.labels)))"

Step 2: Judge

Read the body content directly — don't just trust the risk/effort labels as ground truth. Weigh:

• Does the Deliverables/Acceptance Criteria text describe touching authentication, session handling, claim/locking logic, or other structurally sensitive behavior, regardless of what the risk/effort labels say? That's a reason to recommend more cautiously than the labels alone imply.
• Does the record describe creating or substantially editing a file under skills/**/*.md (a new or changed skill)? This includes harness-health:new-skill findings — their body reads "New skill candidate" with a "Proposed new skill" deliverable (see bin/lib/harness-health/issue-payload.js). Recognize this from body content, not from a label — new-skill findings currently carry no risk:*/effort:* labels at all, by design, so labels alone tell you nothing here. A well-specified new-skill proposal can still reasonably recommend RECOMMEND_BUILD: true (drafting the content autonomously is fine — a human still confirms the grant, and reviews again before any merge), but recommend RECOMMEND_MERGE: false — new skill files encode instructions future agents follow, which is high-leverage independent of how small or clean the proposal looks.
• Is the described change actually lower-risk than its labels suggest (e.g. a risk:medium record that turns out to be a pure documentation correction with no behavioral surface)? Judge accuracy, not blanket caution — recommend generously when the content genuinely supports it.
• A missing Current State/Deliverables/Acceptance Criteria section, or an unresolved TBD/TODO/<!-- ambiguity: marker, is not this mode's job to catch — that's /claude-tweaks:triage's own Step 3.5 body-shape re-verification, which runs after this mode regardless of its output.

Step 3: Render

Output ONLY these lines, no preamble:

RECOMMEND_BUILD: true | false
RECOMMEND_MERGE: true | false
RATIONALE: {one paragraph, naming the specific content signal the recommendation is based on}

If nothing in the record's content or scoring supports any recommendation, output RECOMMEND_BUILD: false / RECOMMEND_MERGE: false — triage's Step 2 already treats this the same as today's "flag back (needs scoring)" case; no separate error path is needed here.

Mode: merge-check

Called from: /claude-tweaks:dispatch's Auto-merge gate, replacing layers 2-4 (scoring eligibility, runtime cleanliness, blast radius) entirely. Layer 1 (authorization — auto:merge present on every group member) stays a hard binary gate in dispatch/SKILL.md itself, unchanged.

Step 1: Gather

The calling agent has just finished this run's build, test, and review — the diff and review verdict are already in its own context. Confirm rather than re-derive where possible. $MERGE_BASE is the commit this run's worktree branched from — the same base the pipeline's own build started from — resolvable via git merge-base main HEAD if not already known from context.

git diff --numstat "$MERGE_BASE"..HEAD | node -e "
const fs = require('fs');
let input = '';
process.stdin.on('data', d => input += d);
process.stdin.on('end', () => {
  const files = input.trim().split('\\n').filter(Boolean).map(line => {
    const [additions, deletions, ...pathParts] = line.split('\\t');
    return { path: pathParts.join('\\t'), additions: parseInt(additions), deletions: parseInt(deletions) };
  });
  fs.writeFileSync('/tmp/assess-merge-files-${N}.json', JSON.stringify(files));
});
"

Read this project's own configured merge-sensitive-paths/automerge-max-lines/ automerge-max-files directly — this skill reads its own config, the same way skills/dispatch/SKILL.md's existing Configuration section reads dispatch-retry-ceiling and friends directly rather than expecting a caller to pre-fetch and pass them:

grep -E "^merge-sensitive-paths:|^automerge-max-lines:|^automerge-max-files:" CLAUDE.md .claude-tweaks/policy.yml 2>/dev/null

merge-sensitive-paths is a single line, comma-separated glob list (e.g. merge-sensitive-paths: bin/hooks.js,skills/_shared/*.md,.claude-tweaks/policy.yml) — split on , and trim whitespace; absent entirely defaults to [] (see _shared/work-record.md's Config keys table). automerge-max-lines/automerge-max-files default to 40/2 when absent, matching skills/dispatch/SKILL.md's existing Configuration table.

Then compute the blast-radius summary:

node -e "
  const { classifyDiffFiles, blastRadiusSummary } = require(process.env.CLAUDE_PLUGIN_ROOT + '/bin/lib/issues/blast-radius.js');
  const files = require('/tmp/assess-merge-files-${N}.json'); // [{path, additions, deletions}]
  const sensitivePaths = process.argv[1] ? process.argv[1].split(',').map((s) => s.trim()).filter(Boolean) : [];
  const classified = classifyDiffFiles(files, sensitivePaths);
  console.log(JSON.stringify(blastRadiusSummary(classified)));
" "$MERGE_SENSITIVE_PATHS_CSV" > /tmp/assess-merge-blast-radius-${N}.json

($MERGE_SENSITIVE_PATHS_CSV is the comma-separated value parsed from the grep above, e.g. "bin/hooks.js,skills/_shared/*.md" — passed as a positional arg, not an env var expected from a caller, since this skill reads its own config rather than depending on one.)

Step 2: Judge

• Sensitive-path hit is a hard floor. If sensitiveFilesTouched is non-empty, render needs-human immediately — do not weigh anything else. No content judgment overrides this.
• A new or substantially-edited skills/**/*.md file is needs-human, regardless of size. Generalizes the old harness-health:new-skill exclusion — a skill file shapes future agent behavior, which is high-leverage independent of how small the diff looks.
• Weigh blastRadiusSummary.implLines/implFiles against the project's configured automerge-max-lines/automerge-max-files as one input, not a cutoff — a diff comfortably under the configured guideline (e.g. #18's 33 impl lines under a 40-line guideline) supports auto-merge when review is clean; a diff well past it is a reason to lean needs-human even with a clean review, but is not an automatic disqualifier the way the old mechanical gate was. testLines/testFiles are informational only — never weigh test-file bulk toward risk.
• Weigh the diff's actual content, not just its size or file list: does it touch concurrency, locking, auth, or external API calls in a way that looks structurally sensitive even outside the configured merge-sensitive-paths list? Treat that as elevated risk from content, the same way a human reviewer would flag it on sight.
• Review's findings are a hard input, not advisory: if this run's /claude-tweaks:review pass produced anything at Medium severity or above, render needs-human — this mode never overrides a real review finding.

Step 3: Render

VERDICT: auto-merge | needs-human
RATIONALE: {one paragraph, naming the specific factors weighed}

Mode: failure-check

Called from: /claude-tweaks:dispatch's Settle step, replacing "any failed run unconditionally revokes auto:merge."

Step 1: Gather

gh api "repos/{owner}/{repo}/issues/${N}/comments?per_page=100" > /tmp/assess-failure-comments-${N}.json
node -e "
  const { countFailedAttempts } = require(process.env.CLAUDE_PLUGIN_ROOT + '/bin/lib/issues/retry.js');
  const comments = require('/tmp/assess-failure-comments-${N}.json');
  console.log(JSON.stringify({ priorAttempts: countFailedAttempts(comments) }));
"

Read the actual failure output from the gate that failed (test output, review findings, error logs) — already in the calling agent's context from the run that just failed.

Step 2: Judge

• Transient signatures: gh api rate-limit (HTTP 429) responses, network timeouts, ECONNREFUSED, or a failure in a test this repo already documents as flaky (currently statusline.test.js's render-timing assertion). Classify transient.
• Correctness signatures: a test failure showing an assertion mismatch directly tied to code the record's own diff changed (expected/actual values diverging in logic the change touched). Classify correctness.
• Ambiguous: anything that doesn't clearly match either pattern. Classify ambiguous and handle it exactly like correctness downstream (see Output) — when genuinely unsure, err toward the existing conservative behavior, never toward the new permissive one.
• NOTIFY_NOW: set true when this is the same correctness-class failure recurring verbatim across two or more consecutive attempts (compare this failure's content against priorAttempts' recorded reasons in the comment history) — a signal the agent may be stuck rather than making incremental progress. Otherwise false.

Step 3: Render

CLASSIFICATION: correctness | transient | ambiguous
NOTIFY_NOW: true | false
RATIONALE: {one paragraph}

The caller (dispatch's Settle step) is responsible for acting on CLASSIFICATION — revoking auto:merge for correctness/ambiguous, preserving it for transient — and for the retry-ceiling bookkeeping, which runs unconditionally regardless of this mode's output (see skills/dispatch/SKILL.md's Settle step).

Error Handling

If this skill cannot render a clear verdict for any reason (malformed input, an inconclusive read), default to the conservative outcome for whichever mode was running: grant-check → RECOMMEND_BUILD: false / RECOMMEND_MERGE: false; merge-check → VERDICT: needs-human; failure-check → CLASSIFICATION: correctness. Never resolve ambiguity toward more autonomy — a missed auto-merge costs a human a click; a wrongly-granted one could ship something bad.

Component-Skill Contract

/claude-tweaks:assess-agent-autonomy is always a component skill — it is never invoked directly by a human, and never renders a ## Next Actions block. Its only callers are /claude-tweaks:triage (Step 2, grant-check) and /claude-tweaks:dispatch (Auto-merge gate, merge-check; Settle step, failure-check).

Anti-Patterns

Pattern	Why It Fails
Overriding a review finding at Medium+ severity because the diff otherwise looks safe	merge-check never overrides a real review finding — review's own findings are a hard input, not advisory.
Weighing test-file line count toward risk in merge-check	The entire reason this skill exists is that test-line bulk isn't implementation risk — testLines/testFiles are informational only.
Skipping the sensitive-path hard floor because the content judgment "looks fine"	Sensitive paths are a floor precisely because they're the cases where content judgment alone isn't sufficient signal — never overridden.
Classifying an unclear failure as transient "to be less conservative"	Ambiguity always resolves to correctness's conservative handling — the point of this skill is accuracy, not blanket permissiveness.
Recommending RECOMMEND_MERGE: true for a new-or-changed skills/**/*.md file	Skill files shape future agent behavior — this is a hard needs-human/false case regardless of how clean or small the change looks.
Dispatching this as a fresh Task agent instead of an inline Skill invocation	The calling agent already has the diff/review-findings/failure-output in its own context — a subagent restart only pays to re-derive what's already known.
Writing to decisions.md from inside this skill	This skill doesn't know about run-dir resolution — logging is the caller's job (/claude-tweaks:triage or /claude-tweaks:dispatch), matching every other auto-decision log entry in this codebase.

Relationship to Other Skills

Skill	Relationship
/claude-tweaks:triage	Calls grant-check once per worklist record in Step 2 — the output becomes the batch table's Recommended column directly. Triage still renders the human batch-confirm exactly as before; only what generates the suggestion changed.
/claude-tweaks:dispatch	Calls merge-check in the Auto-merge gate (replacing layers 2-4) and failure-check in the Settle step (replacing the old unconditional-revocation rule). Dispatch still owns layer 1 (authorization) and all label/claim mechanics directly.
bin/lib/issues/blast-radius.js	Pure module supplying merge-check's one genuinely mechanical input — test-exclusion-aware diff sizing. This skill never computes blast radius itself.
bin/lib/issues/tier.js	extractRiskEffort's surviving colon-form reader supplies grant-check's current-label input. recommendGrants/recommendTier are retired — this skill replaces them as triage's recommendation signal.
bin/lib/issues/retry.js	countFailedAttempts supplies failure-check's retry-history input.
_shared/work-record.md	Taxonomy home — the merge-sensitive-paths config key this skill's merge-check mode reads as a hard floor.
docs/superpowers/specs/2026-07-15-assess-agent-autonomy-design.md	The full design rationale, motivation (the #18/#19 evidence), and calibration examples this skill's judgment procedures are anchored against.