Audit
Audit shortcut — skips task-orchestrator type-classification by pre-fixing task_type=Audit. Same pipeline machinery as /pipeline-orchestrator:pipeline (information-gate, design-interrogator, plan-architect, executor-controller, sanity, Pa de Cal). Invoked manually via `/pipeline-orchestrator:audit [scope]` — never auto-invoked because every audit run produces a structured report (AUDIT_REPORT) that the user must consciously trigger. Variant flags `--light` / `--heavy` route directly to skills/audit-light or skills/audit-heavy with their prescriptive 9-step procedures imported from Pulsar.From its SKILL.md
npx -y skills add fernandoxavier02/Pipeline-Orchestrator --skill auditAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
2 things to look at
- no licenseNo license file was found in the repository. Code published without one is not open source by default, so using it at work is a question for whoever answers licensing questions where you are.
- 1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
SKILL.md
6.6 KB, as published. Nobody here has run it
Audit entry-point (v4.5.0)
You are invoking /pipeline-orchestrator:audit — a thin shortcut that delegates to the same pipeline-controller agent as /pipeline-orchestrator:pipeline, but with task_type pre-fixed to Audit.
Variant override via flag (Slice 3a v4.5.0+)
Before delegating to the controller, inspect $ARGUMENTS for a leading variant-override flag. The check is purely additive — invocations without a flag keep the existing auto-classify behavior.
- If
$ARGUMENTSstarts with--light(with trailing space) OR$ARGUMENTSis exactly--light: strip the--lightprefix and invokeSkill(skill: "pipeline-orchestrator:audit-light")with the remaining$ARGUMENTS(may be empty). - If
$ARGUMENTSstarts with--heavy(with trailing space) OR$ARGUMENTSis exactly--heavy: strip the--heavyprefix and invokeSkill(skill: "pipeline-orchestrator:audit-heavy")with the remaining$ARGUMENTS. - Otherwise (no recognized flag): proceed with the controller dispatch below — auto-classification is unchanged.
The audit-light and audit-heavy skills carry the prescriptive 9-step procedures imported from Pulsar (D:\Projeto Pulsar\.claude\commands\Prompts\Audtiroria\). Phase 0 (information-gate) and Phase 3 (sanity / final-validator / finishing-branch) still wrap them via pipeline-controller when invoked through the auto-classify path.
What this skill does
Spawn the pipeline-controller agent with the user's request prefixed by PRE_CLASSIFIED_TYPE=Audit:
Agent(
subagent_type: "pipeline-orchestrator:core:pipeline-controller",
description: "Orchestrate audit pipeline for the user request",
prompt: "PRE_CLASSIFIED_TYPE=Audit\n\n$ARGUMENTS"
)
The controller returns a PIPELINE COMPLETE block as its tool result. Show it to the user verbatim — do NOT summarize, trim, or paraphrase.
What this skill does NOT do
- Does NOT skip information-gate — gap detection (scope, axes, baseline, stakeholder) is still mandatory.
- Does NOT skip design-interrogator for COMPLEXA audits — scope clarity still applies.
- Does NOT permit any code change. Audit pipelines are REPORT-ONLY by Iron Law. Every audit agent (
audit-intake,audit-domain-analyzer,audit-compliance-checker,audit-risk-matrix-generator) is read-only by frontmatter and prompt. - Does NOT skip per-step evidence requirement — every finding cites file:line or is tagged
[HYPOTHESIS]/[DESIGN]. - Does NOT skip sanity check or Pa de Cal — Phase 3 runs identically (sanity verifies report completeness; Pa de Cal issues GO/CONDITIONAL/NO-GO on the report quality + risk matrix).
- Is NOT auto-invoked.
disable-model-invocation: trueenforces manual-only triggering. Claude will never decide to run this skill on its own — it only runs when the user types/pipeline-orchestrator:audit.
The ONLY phase shortened is Phase 0a (task-orchestrator): the classifier accepts force_type=Audit (via the PRE_CLASSIFIED_TYPE prefix) and skips the type-classification reasoning, but still computes complexity, pipeline_variant, and ssot_status. See agents/core/task-orchestrator.md Step 1a.
Pass-through behavior
The $ARGUMENTS placeholder captures everything the user typed after the skill name. The full string is passed verbatim to the controller, prefixed by PRE_CLASSIFIED_TYPE=Audit\n\n. The controller's Step 1 recognizes the prefix, the task-orchestrator Step 1a strips and consumes it, and the rest of the 4-phase pipeline runs identically to a /pipeline invocation that classified as Audit.
Why this exists
Without /audit, every "audit the auth layer" or "review data integrity" request burns one classification round to deduce type=Audit. With /audit, you tell the controller upfront and it goes straight to scope/baseline gap detection — saves tokens and prevents misclassification as Bug Fix when the user already knows it's a read-only audit.
The variant-override flags (--light / --heavy) are the recommended path when the user knows the depth they want: they skip the full pipeline-controller wrapper and run the prescriptive 9-step skill directly.
Achado #7 GATE_REQUEST handler (2026-05-07+, v5.2.0-rc.2+)
When the dispatched pipeline-controller (or any subagent it transitively dispatches via DISPATCH_REQUEST) returns a tool result containing === GATE_REQUEST v1 ===, === DISPATCH_REQUEST v1 ===, or === PLAN_MODE_REQUEST v1 === blocks AND ends with STATUS: AWAITING_GATE_RESPONSES / AWAITING_DISPATCH_RESULTS / AWAITING_PLAN_MODE_RESULTS, the parent main LLM MUST process them per references/gate-request-protocol.md:
- Parse each block out of the tool result.
- For
GATE_REQUEST: invokeAskUserQuestionwith the parsed question + options. - For
DISPATCH_REQUESTwithtarget_kind: agent: invokeAgent(subagent_type, description, prompt). - For
DISPATCH_REQUESTwithtarget_kind: skill: invokeSkill(skill: target_name). - For
PLAN_MODE_REQUEST: invokeEnterPlanMode, conduct read-only research, exit with plan. - Aggregate responses/results into
GATE_RESPONSES/DISPATCH_RESULTS/PLAN_MODE_RESULTSYAML payloads. - Re-dispatch the SAME subagent with the original prompt PLUS payloads prepended.
- Repeat 1-7 until the subagent emits its terminal block (e.g.,
PIPELINE COMPLETE) without AWAITING_*.
Append every block emission and every response to {PIPELINE_DOC_PATH}/protocol-events.jsonl (NOT gate-decisions.jsonl). Named gates (ADVERSARIAL_GATE, FINAL_ADVERSARIAL_GATE, CLOSEOUT_CONFIRM, TDD_APPROVAL, PLAN_REJECTED, INFO_GATE_BLOCKED) ALSO get a canonical gate-decisions.jsonl entry with decided_by: user referencing the protocol event id. See references/gate-request-protocol.md "gate_id → canonical gate mapping" for the full table.
Never silently default. Malformed blocks → present to user via your own AskUserQuestion ("malformed block — investigate, retry, or abort?"); do NOT guess.