菜单
Assess delivery health metrics. For software: DORA + APEX. For content/AI/service products: product-type-appropriate metrics.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Use to evaluate the current state of a diamond. Checks theory gates, confidence levels, and recommends next action.
Use when building anything USER-FACING (or with persuasion/retention/cancellation/consent/pricing flows, or that touches vulnerable people) to surface design-level harm the security/privacy/compliance gates miss: dark/deceptive patterns and foreseeable misuse. Assumes the product works as designed and asks who it could harm and whether it is Happier-negative. NUDGE, not a block.
Lint canvas files for staleness, missing fields, inconsistent evidence types, and orphaned references. Run periodically or before major transitions.
Accessibility audit against WCAG 2.1 AA. Checks semantic HTML, ARIA, keyboard navigation, color contrast, screen reader compatibility.
Design the smallest viable test to validate or invalidate a critical assumption. Based on Torres's assumption testing framework, organized by Gilad's AFTER model (Assessment → Fact-Finding → Tests → Experiments → Release Results).
Use before any research activity or significant decision. Reviews cognitive biases relevant to the current stage.
| name | dora-check |
| description | Assess delivery health metrics. For software: DORA + APEX. For content/AI/service products: product-type-appropriate metrics. |
| metadata | {"instruction_budget":"126","framework_dependency":"mycelium","framework_dependency_note":"This skill is designed to run within the Mycelium framework (https://github.com/haabe/mycelium). Standalone use will skip the canvas state, theory gates, and harness behavior the skill assumes. Install: /plugin install mycelium@haabe-mycelium."} |
Assess delivery health using product-type-appropriate metrics. Check product_type from .claude/diamonds/active.yml to determine which assessment to run.
Product type routing (v0.11.0):
Hard rule. Before issuing Write or Edit against any .claude/canvas/*.yml, use the Read tool on that file in this session. Claude Code's Read-before-Write check requires the Read tool specifically — cat/head/grep via Bash do NOT satisfy it.
Edit vs Write — different cost profiles (verified 2026-05-14):
Edit (exact-string replacement): Read with limit: 1 satisfies the check at ~50 tokens. State-tracking is per-file, not per-byte — subsequent Edit calls work anywhere in the file. Use this for partial updates against large canvas files (e.g., purpose.yml at 800+ lines).Write (full replacement): do a full Read first. Write obliterates the file; you should see what you're about to replace. The limit:1 shortcut is not appropriate here.ID-bearing entries — scan the ID space before assigning (added 2026-05-15, v0.23.19): When adding a new component, opportunity, solution, or any other ID-bearing entry to a canvas file, run a Bash grep first to confirm the next ID in your prefix sequence is actually free:
grep "^ - id: <prefix>-" .claude/canvas/<file>.yml | sort -u
Replace <prefix> with the canvas's ID prefix (comp for landscape, opp for opportunities, sol for solutions, ht for human-tasks, etc.). Then pick the next free integer. validate_canvas.py has a duplicate-ID check (lines 230-239) that catches the failure on CI, but a duplicate can persist in the working tree for days if CI isn't run between edit and discovery — see roadmap-repo corrections.md 2026-05-15 "Duplicate canvas ID created in landscape.yml" for the worked example.
Original failure mode: anti-pattern #7 instance #5, 2026-05-09 — agent conflated Bash head with the Read tool, lost ~14k tokens to a Write-fail → remedial-full-Read → re-Write loop. The limit:1 discipline (graduated 2026-05-14, v0.23.18) prevents the second-order cost where the agent correctly follows the rule but full-Reads every time. The ID-scan discipline (graduated 2026-05-15, v0.23.19) prevents the related class where the agent reads enough of the file to satisfy the Edit check but not enough to see existing ID assignments — kin to anti-pattern #8 (Stale State Read).
If this skill writes to multiple canvas files, register each one first (limit:1 for Edit-only paths; full Read for Write paths) AND ID-scan any prefix you intend to assign.
See CLAUDE.md Canvas writes — Read before Write for the canonical rule.
Assess delivery health using the four core DORA delivery metrics (Forsgren) plus reliability as an operational adjunct, AND LinearB's APEX AI-era metrics.
Gather current metrics from CI/CD, deployment logs, incident records.
Note: DORA's core set is four delivery metrics (below). "MTTR" was renamed to "Failed Deployment Recovery Time" (FDRT) for precision — the original was ambiguous with other mean-time-to-X metrics. Reliability is an operational-performance dimension DORA added in 2021 — dora.dev classifies it as operational, not software-delivery — so it is assessed here as an adjunct via SLOs/SLIs (Part 3 SRE), not as a fifth delivery metric. (The 2024 State of DevOps report's additional delivery metric was deployment rework rate, which this skill does not separately track.)
Deployment Frequency: How often does code reach production?
Lead Time for Changes: Commit to production time?
Change Failure Rate: % of deployments causing failure?
Failed Deployment Recovery Time (FDRT): Time to restore service after a failed deployment?
Formerly "Mean Time to Recovery (MTTR)." Renamed for precision — FDRT measures recovery from failed deployments specifically, not all incidents.
Reliability (operational adjunct, not a core delivery metric): Does the software meet or exceed its reliability targets?
Added by DORA in 2021 as an operational-performance dimension (dora.dev classifies it as operational, not software-delivery). Measured via SLOs/SLIs; assessed in this skill as an adjunct that connects to the SRE metrics in Part 3 — it does not populate the four-metric dora-metrics.yml delivery core.
"Faster coding doesn't mean faster delivery."
Assess the four APEX pillars to detect AI-era delivery problems:
Why this is mandatory ([S1] Faros Acceleration Whiplash 2026, [S2] Faros Harness Engineering 2026, [S5] Datadog State of AI Engineering 2026): the APEX section was prone to being filled with narrative notes: and status: colors while the underlying numbers stayed unmeasured — scaffold-without-instrumentation (corrections.md 2026-06-07 in the roadmap repo). The fix is to compute the load-bearing APEX fields from raw artifacts the project already produces, rather than treating the canvas surface as the measurement.
For each field below, compute the number first from the raw artifact, then write value: to canvas; preserve the prior narrative as notes: (do not destroy it).
apex.ai_rework_rate — fix-forward commits as a % of total commits in window.
git log --since="<window>" --pretty='%H %s' on the repo's primary branches.^(fix|amend|patch|hotfix|chore: re-validate|chore: fix) or modifies files touched by the immediately prior commit on the same path within N hours). Document the chosen pattern + window in apex.ai_rework_rate.method.{ value: <pct>, window: "<24h|7d|30d>", method: "<pattern>", notes: "<prior narrative>" }.measurement_history alongside the existing change_failure_rate strings.ai_rework_rate ≈ project rework_rate; name this in notes:.apex.ai_acceptance_rate — proportion of agent-proposed actions the user accepts without redirection.
corrections.md entries with detection_origin: user in the same window ÷ count of total agent actions narrated in the session log if available; otherwise narrative estimate carried forward with method: "narrative; instrumentation pending".{ value: <pct>, method: "<exact|proxy|narrative>", notes: "<prior narrative>" }.apex.hook_detection_rate — proportion of corrections that were caught by a hook or validator vs by the user.
corrections.md entries' detection_origin: field (user / hook / evaluator / agent). Count by category in window.{ value: <pct>, denom: <total>, user: <n>, hook: <n>, evaluator: <n>, agent: <n>, window: "<7d|30d>", notes: "load-bearing — if hook+evaluator ≈ 0%, the harness-detection-gap signal applies (corrections.md 2026-06-02 audit)" }.Goodhart pair (already named in dora-metrics.yml#goodhart_check): ai_rework_rate MUST be paired with ai_acceptance_rate. Rework-rate rising while acceptance-rate ALSO rising = healthy (validators catching real issues, user still accepting agent direction). Rework-rate rising while acceptance-rate falling = unhealthy (agent output quality dropping). Never report one without the other.
apex.first_pass_success_rate (v0.39.20, the Goodhart counter for ai_rework_rate + hook_detection_rate) — proportion of eval scenarios that pass on first attempt across the active corpus.
.claude/evals/results/ (canonical eval-runner output) AND .claude/auto-dogfood/results/ (the auto-dogfood full-session battery, where projects that run it write <ts>-aggregate.json). If the project configures another battery results dir, check it too. Each scenario-result carries {passed, expected_to_fail, invalid, token_usage, score}; aggregate files are JSON arrays of scenario-results, single-scenario files are one object. These are the actual eval-runner / battery outputs and reflect real model behaviour, which pass-history.json (a derived counter) often does not. Locate-before-deny (2026-06-15 lesson): before reporting null or "no battery exists", search BOTH dirs in the CURRENT tree AND git log --all --name-only — a negative grep in one location is not absence; the harness/results may live in a project-state dir you didn't check.
invalid: true OR token_usage.total == 0. A zero-token result means the claude CLI never ran — counting it as pass or fail FABRICATES the metric (this is the exact artifact that produced the phantom INVALID 0/7 / 38% "baseline"). Compute over VALID results only and report skipped_invalid.expected_to_fail: true are gap-provers — a passed: false there is the DESIGNED outcome, not a quality miss. Exclude them from the pass-rate denominator. Never let a gap-prover drag the number to a fake 0%.passed == true ÷ valid, non-expected-to-fail scenario-results, taken from the MOST RECENT battery run(s). Always report the run date — a stale battery is a stale number, not a current one..claude/evals/pass-history.json ({runs, passes, last_5, status} per eval). Use sum-passes/sum-runs over status: active ONLY when no valid battery aggregate is present. NOTE: this corpus is frequently dormant (runs: 0); a 0-run counter is a dormancy signal, not the live number.{ value: <pct>, denom: <valid_scenarios>, numerator: <passed>, skipped_invalid: <n>, skipped_expected_fail: <n>, source: "<results-aggregate path + run date | pass-history>", trailing_5_pct: <pct or null>, method: "...", notes: "<honest framing incl. staleness>" }.value: null ONLY when there is no valid battery aggregate AND pass-history has 0 runs — and name the gap. A null from a dormant instrument is MORE honest than a number computed from invalid (0-token) or expected-to-fail results. Do NOT default to a fake number; do NOT count a dead-runner artifact.ai_rework_rate and hook_detection_rate. Rework falling AND first-pass-success rising = healthy. Rework falling AND first-pass-success ALSO falling = unhealthy (you've optimized rework by NOT shipping, or by gaming the rework pattern). Never read rework alone.corrections.md 2026-06-07): every new metric field ships with its counter at the same maturity. This field exists because the three preceding APEX numerics shipped in v0.39.19; the counter ships in v0.39.20 in the same Part 2b block to honor that discipline.Cadence: compute on every /mycelium:dora-check run; do not narrate "refreshed APEX" unless these four numeric fields actually moved (Postflight discipline below).
Sources:
Hard rule (per CLAUDE.md Communication Rules, anti-pattern #7 Stage 2 graduation — v0.39.18). Before the user-facing summary claims "✅ updated dora-metrics.yml" / "✅ refreshed APEX" / similar, verify the write actually changed the value fields named in ## Canvas Output below — not just _meta.last_validated. Use the Read tool on dora-metrics.yml after the edits to confirm deployment_frequency.current, change_failure_rate.current, time_to_restore.current, apex.*, measurement_history, etc. hold the new measurements. If only _meta changed and value fields stayed stale, the narration is a state-claim on a state the skill did not achieve (AP#7 instance #18 was the worked failure, 2026-06-05 — operator surfaced it with the Torres-shape question "Did you run the dora check?"). Symmetric to the Preflight above: Preflight protects what gets WRITTEN; Postflight protects what gets CLAIMED about what was written. Validator Check 42 enforces preamble presence.
Lead with the verdict; isolate the bottleneck (Von Restorff, per harness/design-principles.md — the constraining metric is the one finding the reader must not scroll past; graduated from two consecutive /framework-health 4e flags, 2026-06-05 + 2026-06-12):
## DORA + APEX Assessment
> **Verdict: [HEALTHY | BOTTLENECK: <metric> — <one line on where the constraint lives and the first focusing step>]**
### DORA Metrics
| Metric | Current | Level | Target | Gap |
|--------|---------|-------|--------|-----|
| Deploy freq | ... | ... | ... | ... |
| Lead time | ... | ... | ... | ... |
| Change fail rate | ... | ... | ... | ... |
| FDRT | ... | ... | ... | ... |
| Reliability | ... | ... | ... | ... |
### APEX Metrics (AI-Era)
| Pillar | Status | Key Signal |
|--------|--------|-----------|
| AI Leverage | ... | AI acceptance rate: ...% |
| Predictability | ... | Planning accuracy: ...%, Rework rate: ...% |
| Flow Efficiency | ... | Cycle time: ..., Review wait: ... |
| Developer Experience | ... | Satisfaction: ..., Burnout: ... |
### Shifting Bottleneck Check
[Is coding faster but review/deployment slower? Yes/No]
[If yes: where is the new bottleneck?]
### DORA Bottleneck
[The metric most constraining overall performance]
### Value Stream Diagnosis (if bottleneck detected)
If DORA shows a bottleneck, map the value stream to identify WHERE in the flow the constraint lives:
- Run `/mycelium:canvas-update` to update `.claude/canvas/value-stream.yml` with current stage timings
- Apply Theory of Constraints Five Focusing Steps (Goldratt): Identify -> Exploit -> Subordinate -> Elevate -> Repeat
- Look for wait times >> process times (a sign of queuing, not capacity, problems)
- Look for high handoff counts (each handoff adds delay and information loss)
- Calculate flow efficiency: process_time / lead_time -- target >25%
### Top 3 Improvements
1. [specific action]
2. [specific action]
3. [specific action]
If SLIs/SLOs defined in .claude/canvas/dora-metrics.yml sre section:
Error budgets are the social contract: reliability earns the right to ship faster. Connects to BVSSH Safer.
If NOT defined: "Consider defining SLIs/SLOs to balance velocity with reliability."
Always APPEND a ### DORA Assessment or ### Delivery Metrics Assessment entry to .claude/harness/decision-log.md with:
Always update .claude/canvas/dora-metrics.yml with:
For content_course, content_publication, content_media products. Read .claude/canvas/content-metrics.yml.
Publication Cadence: How often does content reach the audience?
Production Lead Time: Idea to published -- how long?
Revision Rate: % of published content requiring significant revision?
Completion Rate: % of planned content actually completed on schedule?
Time to First Value (TTFV): How quickly does a buyer access and get value after purchase?
Engagement & Drop-off: Course completion rate, satisfaction, return rate?
Acquisition: Conversion rate, cost per acquisition, cart abandonment?
Revenue Health: Refund rate, CLV, churn (subscriptions), NRR?
Update .claude/canvas/content-metrics.yml with current measurements and last_measured timestamp.
For ai_tool products. Read .claude/canvas/ai-tool-metrics.yml.
Eval Frequency: How often are prompts/models evaluated?
Accuracy & Consistency: Are eval scores stable or improving?
Safety Score: Red-team results -- are adversarial inputs handled?
Bias Assessment: Last assessed when? Any demographic gaps found?
Version Cadence: How often are prompt/model versions shipped?
Regulatory Status: EU AI Act risk classification current?
Time to First Value (TTFV): How quickly does a user get useful output after first access?
Usage & Retention: DAU, task success rate, retention (7-day, 30-day)?
Revenue Health: Refund rate, CLV, churn, NRR?
Update .claude/canvas/ai-tool-metrics.yml with current measurements and last_measured timestamp.
For service_offering products. Read .claude/canvas/service-metrics.yml.
Client Throughput: How many clients/engagements per period?
Delivery Lead Time: Engagement start to delivery -- how long?
Client Satisfaction: NPS, CSAT, retention rate, referral rate?
Repeatability: Is the delivery workflow documented and templated? Score 1-5.
Time to First Value (TTFV): How quickly does a client receive meaningful value after engaging?
Acquisition: Conversion rate, cost per acquisition, proposal win rate?
Revenue Health: Refund/dispute rate, CLV, churn (retainers), NRR?
Update .claude/canvas/service-metrics.yml with current measurements and last_measured timestamp.