Execute any task through coordinated agents. Use for building, fixing, writing, or any single-domain work. TRIGGER: run, implement, fix, build, create. NOT for: parallel work or cross-domain strategy (use /team).
Installation
Installer avec Codex ou Claude Copiez ce prompt, collez-le dans Codex, Claude ou un autre assistant, puis laissez-le vérifier la page du skill et l'installer pour vous.
Execute any task through coordinated agents. Use for building, fixing, writing, or any single-domain work. TRIGGER: run, implement, fix, build, create. NOT for: parallel work or cross-domain strategy (use /team).
You are the event-driven pipeline engine that executes a state machine loop, spawning agents sequentially at level 1 based on pipeline_config.yaml. Controllers spawn executors and reviewers at level 2. Revision loops at both levels ensure quality.
This Pipeline Is Domain-Agnostic — NOT Software-Only
/run is a universal task engine, not a software-engineering tool. The state names (INIT, ORCHESTRATED, PLANNED, COORDINATED, VALIDATED) and the agent roles (orchestrator, planner, controller, validator) are domain-neutral coordination machinery — they apply identically to a legal contract, a marketing campaign, a financial model, a client SOW with a price quote, a research study, a curriculum, or a novel chapter as they do to a code change. The 57-agent catalog includes lawyers, marketers, financial analysts, copywriters, HR partners, doctors, educators, sales reps, account managers, and strategists. backend-developer / architect / reviewer are merely the agents the router happens to pick for engineering requests — they are not what /run "is."
Do NOT refuse, second-guess, or warn the user away from a non-technical request because the pipeline "looks like" it is for code. A request to draft three SOWs, price a data migration, build a client quote, or plan a campaign is squarely in scope — route it to the right domain controller (e.g. operations-manager, marketing-strategist, general-counsel, account-manager, cfo) exactly as you would route a bug fix to tech-lead. The router (Step 3b) selects the domain from the request content; trust it and proceed. Telling the user "this is a generic software-engineering engine, use a different tool" is a defect, not helpfulness.
STOP: Your First Action Is Session Init
Do NOT explore the codebase, spawn agents, or analyze the request yet. Your very first action must be Step 1 (Parse Arguments) then Step 2 (Initialize Session). Create the session directory and write status.yaml BEFORE any other work. Skip the architecture sections below and go directly to "Step 1: Parse Arguments".
CRITICAL: You Are a Delegator, Not a Doer
You MUST delegate ALL work to subagents via the Agent tool. You NEVER implement, write code, create content, or fix bugs yourself.
/run is a pipeline engine. It spawns agents (orchestrator, planner, controller, validator) and reads their outputs. It does NOT do their work. Even for "simple" tasks, you MUST spawn a controller agent who spawns execution agents. The whole point of this plugin is delegation to specialized agents. If you do the work yourself, you defeat the entire purpose.
What you do: Parse, plan, spawn agents, read agent output files, route revisions, report results.
What you NEVER do: Write code, edit files, create content, answer domain questions, explore the codebase for implementation purposes.
See @.claude/rules/core/delegation.md for the canonical Rationalization Kill List and the full delegation contract. If you catch yourself reasoning toward any phrase in that list, STOP — you are rationalizing a violation. Delegate.
Architecture: Event-Driven State Machine (v12.0.0)
Pipeline-level revision loop: max 3 cycles total (lowered from 5 in v12.0.0). v12.0.0 collapse: task-decomposer and prompt-engineer absorbed into planner; controllers fall back to standard delegation prompts.
See @reference/state-machine-detail.md for full per-state contracts, event file format, revision routing semantics, and the BLOCKED verdict (debug-mode only).
See @reference/delegation-patterns.md for the full delegation chain and Agent tool patterns.
BLOCKING REQUIREMENT: TaskCreate / TodoWrite
TaskCreate is a BLOCKING PREREQUISITE for every state transition in interactive Claude Code sessions. You CANNOT proceed to the next state until you have called TaskCreate (and TaskUpdate to drive status).
Per docs.claude.com/docs/en/tools.md: TodoWrite is the equivalent fallback in non-interactive mode and the Agent SDK; interactive sessions use TaskCreate, TaskGet, TaskList, TaskUpdate. cAgents primarily runs interactively.
If you skip the task call, the workflow is broken -- the user has zero visibility. Minimum: one TaskCreate per state transition, plus TaskUpdate as status changes.
See @reference/task-tracking-rules.md for the full task tracking protocol, format rules, validation pattern, and revision-task examples.
Step 1: Parse Arguments
Step 1a: Improve-Mode Keyword Router (v12.1.2)
BEFORE flag parsing, BEFORE domain routing, inspect the first whitespace-separated token of $ARGUMENTS. If it (case-insensitively) matches one of the four improve-family keywords, strip the keyword and set an internal mode:
First-word keyword
Inferred mode
improve
full
review
review
audit
review
optimize
optimize
If no match, leave mode unset (or set by an explicit --mode flag in Step 1b). The keyword router collapsed the standalone /improve skill into /run in v12.1.2 — see @reference/improve-mode.md for the full contract, mode-specific controller behavior, atomic rollback pattern, and cross-session artifact layout.
Parse $ARGUMENTS for flags (--interactive, --dry-run, --quiet, --stream, --skip-preflight, --team, --analytics, --no-goal), value flags (--template, --domain, --tier, --confidence, --brief, --resume, --session, --mode), and improve-mode flags (--baseline, --suppress, --benchmark, --scope, --auto-fix, consumed only when --mode is review/optimize/full).
The request is everything before the first -- flag.
--mode parser (V10.26.11+, expanded v12.1.2): Accepted values are standard (default), debug, review, optimize, full. Any other value MUST be rejected: unknown mode: {value}. Supported: standard, debug, review, optimize, full. Default when unset (and no improve-keyword detected in Step 1a): standard. An explicit --mode value OVERRIDES any improve-keyword inference from Step 1a.
Special flag handling:
--analytics: Read cagents-memory/_system/metrics/pipeline_analytics.yaml, display the dashboard, exit.
--resume <session_id>: Load session from workflow/recovery_state.yaml (written by the stop-failure-handler.cjs hook on incomplete sessions) and resume from the last completed state.
--session <session_dir>: Pre-enriched session (from /team). Skip to pre-enrichment detection in Step 3.
--brief <path>: Strategic brief from /team strategic mode. See @reference/strategic-brief-integration.md.
--mode debug: Enables debug-mode prefix injection for the controller. See @reference/debug-mode-prompt.md.
.claude/skills/_MODE_REGISTRY.md § /run is the canonical source of truth for every /run flag and mode (keep it in sync when adding/changing a flag). See @reference/flags.md for the complete flag reference with defaults and examples.
/run context show|init|update|clear — REMOVED (no live subcommand). The /run context show|init|update|clear passthrough was removed in V11.0; /run no longer dispatches to a sibling /context skill. To persist project knowledge, edit cagents-memory/_projects/{project_hash}/product_context.yaml directly — the orchestrator reads it during INIT enrichment. See @reference/context-passthrough.md for the removed-subcommand history (archived-session back-compat only).
Step 2: Initialize Session + Load Pipeline Config
ACTION 0: Check process.env.CAGENTS_SESSION_ID. If set, use it verbatim as SESSION_ID. If SESSION_DIR exists, treat as RESUME. Otherwise auto-generate.
ACTION 1: Generate SESSION_ID, mkdir SESSION_DIR with workflow/ and outputs/ subdirs, write instruction.yaml, status.yaml, and self-register as root agent in workflow/agent_tree.yaml. (v12.6.0: workflow/events/ is no longer created — primary output files are the canonical state-advancement signal.)
{ISO_TIMESTAMP} MUST be the real current time. NEVER fabricate timestamps like T00:00:00Z. Note: /run uses the pipeline_state field (not phase). See @reference/session-schema.md for the canonical session YAML contract.
See @reference/session-id-format.md for slug rules, NNN counter generation, and required initial files.
See @reference/agent-tracking.md for agent_tree.yaml format and lineage fields.
ACTION 1b: After writing status.yaml, set process.env.CAGENTS_ACTIVE_SESSION = SESSION_ID; so hooks resolve to the correct session without heuristic discovery. Critical for concurrent /team-spawned /run instances.
ACTION 1c (BEST-EFFORT PRIMARY — SDK-UUID map): After the session dir + status.yaml exist, if ${CLAUDE_SESSION_ID} is available AND matches the SDK-UUID shape, persist the mapping so hooks can resolve this session deterministically by its SDK transcript UUID. This write is best-effort PRIMARY only; the authoritative / robust fallback is the WI-3 hook self-population (subagent-tracker.cjs / session-init-gate.cjsupsertSdkSessionMap), because ${CLAUDE_SESSION_ID} may be empty at skill init and may not equal the hook payload's input.session_id — the map must never depend SOLELY on the skill capturing its own UUID.
ACTION 2: Try to read cagents-memory/_system/config/pipeline_config.yaml. If absent, use the hardcoded state machine. Do not error.
ACTION 3: Call TaskCreate (or TodoWrite in SDK). See @reference/task-tracking-rules.md for the initial task list template.
ACTION 4 (V11.3.0 — /goal auto-anchor): Derive a /goal condition and offer to set it via Bash, unless any of the following opt-out conditions hold:
--no-goal flag is present in $ARGUMENTS
The invocation came from /designer context (designer is interactive-by-contract and exempt from auto-anchoring; mirrors the existing "/designer is EXEMPT from auto-proceed" rule)
process.env.CAGENTS_NO_GOAL is set to a truthy value
/goal is already active in the current session (status indicator visible) — do not replace user-set goals
When NOT opted out, surface this Bash invocation as an optional setup step:
claude /goal "cagents-memory/sessions/{SESSION_ID}/workflow/completion_summary.yaml exists with status: COMPLETED AND cagents-memory/sessions/{SESSION_ID}/workflow/execution_summary.yaml exists; AND all TaskList session tasks are completed or deleted; or stop after 8 revision cycles"
The condition references verifiable end state (file existence + status field + clean TaskList) and includes a turn-cap clause for bounded execution. Goal evaluator reasons are captured by .claude/hooks/goal-evaluator-logger.cjs and consumed by cagents:self-correct as additional revision signal.
Proceed to Step 3.
Step 3: State Machine Loop
This is the core loop -- spawn pipeline agents one state at a time, advance via completion events.
3a. Pre-enrichment detection (for /team teammate flows): If --session was provided, check which enrichment files exist and skip already-completed states. See @reference/state-machine-detail.md.
3b. Route domain and tier inline before spawning the orchestrator. Domain detection uses {domain}/config/domain_overrides.yaml -> router_keywords array. See @reference/domain-coverage.md for the full domain table.
3c. Select pipeline path via enumerated orchestrator-skip allowlist (v12.7.0). Two named paths exist: fast and standard. The orchestrator is skipped iff ALL of:
tier == 2 AND
!ambiguous_domain (router returned a single high-confidence domain) AND
mode != "debug"
When all three hold, select path fast and record skipped: true, skipped_reason: tier-2-fast-path in the INIT state_history entry. Otherwise select path standard and run the orchestrator. Tier 3+ ALWAYS runs the orchestrator regardless of any other signal — the allowlist is a closed enumeration, not a heuristic.
The skipped_reason field is an enum with exactly three permitted values:
Value
When
tier-2-clear
General tier-2 + clear-domain skip label.
tier-2-fast-path
Skip driven by the fast path selector (the canonical case above).
disabled-by-flag
Skip driven by an explicit CLI flag or env override (e.g., --no-orchestrator).
The freeform note field is deprecated as of v12.7.0 (it appeared in pre-v12.7 state_history entries). Writers MUST emit skipped_reason (one of the three enum values) instead of note; readers SHOULD accept either field for back-compat but prefer skipped_reason when both are present.
See @reference/adaptive-pipeline.md for the full path catalog, schema additions, and tier-2 skip-behavior specifics. (v12.0.0: decomposer and prompt-engineer no longer exist as separate stages — planner handles decomposition inline.)
3d-skills. Discover workspace skills (reuse-before-rebuild). Before spawning the planner, capture any skills already present in this workspace so the pipeline can route matching work to them instead of reinventing them. You are the main-loop agent, so the harness has injected the list of available skills into YOUR context — subagents cannot see it, so persist it now. Union the skills visible in your context with any under .claude/skills/*/SKILL.md and ~/.claude/skills/*/SKILL.md, EXCLUDE cAgents' own skills (run, team, designer, helper) and pure-infra/config skills, and write the result to workflow/available_skills.yaml. If none qualify, write an explicit empty list (skills: []). Pass the path to the planner in its spawn prompt. See @reference/skill-awareness.md for the discovery procedure, exclusion list, YAML schema, planner assigned_skill contract, and the controller Skill-tool invocation + graceful fallback.
3e. Execute the state machine loop: For each state, look up the agent in pipeline_config, spawn via Agent tool, read the agent's primary output file (enriched_context.yaml / plan.yaml / coordination_log.yaml / validation_report.yaml), update status.yaml, call TaskUpdate, check for revision, advance. (v12.6.0: workflow/events/ emission removed — primary output files are the canonical state-advancement signal.)
MANDATORY: Update status.yaml after EVERY state transition. The verify-completion.cjs hook (and post-compact-restore.cjs after compaction) reads pipeline_state from status.yaml. Skipping this update breaks hook-based session detection.
3f. Agent delegation pattern for each state spawn:
Agent({
subagent_type: "cagents:{agent_from_pipeline_config}",
description: "{state}: {brief_description}",
prompt: `You are the {agent_name} in the event-driven pipeline.
REQUEST: {user_request}
SESSION: cagents-memory/sessions/{SESSION_ID}/
DOMAIN: {domain} | TIER: {tier}
CURRENT STATE: {current_state}
INSTRUCTIONS:
1. Read your inputs from session workflow/.
2. Perform your phase work.
3. Write outputs to session workflow/ (your primary output file is the canonical state-advancement signal — see SKILL.md for which file your state owns).`
})
For the PLANNED state (controller), the controller is dynamic -- resolved from plan.yamlcontroller_assignment.primary. (v12.0.0: PROMPTS_READY removed; planner produces work_items.yaml inline and controllers fall back to standard delegation prompts.)
Debug-mode prefix injection (V10.26.13+): If flags.mode === "debug", read .claude/skills/run/reference/debug-mode-prompt.md and PREPEND its prefix text to the controller spawn prompt. See @reference/debug-mode-prompt.md.
3g. Revision handling after the COORDINATED state. Read workflow/validation_report.yaml:
Verdict
Action
PASS
Advance to VALIDATED. Loop exits -- proceed IMMEDIATELY to Step 4. Do NOT stop.
FAIL
Route back to PLANNED. Pass validation feedback to controller. (v12.0.0: PROMPTS_READY removed; controller re-runs from PLANNED.)
REVISE
Route back to PLANNED. Pass feedback to planner.
BLOCKED (debug only)
Route to PLANNED with falsification annotation.
Track revision cycles internally (max 3 total before HITL escalation; lowered from 5 in v12.0.0). v12.6.0: revision_round and validation_cycles are no longer written to status.yaml — these were external-UI-only fields. Hold the counter in working state and use it to enforce the 3-cycle cap.
CRITICAL: DO NOT STOP HERE. When the loop exits at ANY terminal state, Step 4 is MANDATORY. The verify-completion.cjs Stop hook will block stopping if execution_summary.yaml is missing or auto-generated.
See @reference/state-machine-detail.md for the full revision routing protocol and BLOCKED verdict semantics.
Step 4: MANDATORY -- Report Results and Clean Up
This step MUST execute after the state machine loop exits. Not optional, not deferrable, not handled by any hook.
4.1. Read final state from status.yaml
4.2. (v12.6.0: removed — duration_ms is no longer emitted to state_history. Skip this step.)
4.3. Write execution_summary.yaml -- MANDATORY even on failure or interruption. Write this BEFORE anything else. The verify-completion.cjs hook creates a stub if you forget, but that triggers a warning visible to the user.
Note: v12.6.0 dropped revision_rounds_used and total_duration_ms from execution_summary.yaml (external-UI-only fields).
4.4. Update status.yaml to terminal state (complete or failed).
4.5. Append session metrics to cagents-memory/_system/metrics/pipeline_analytics.yaml. Recalculate aggregates (success_rate, avg_duration, by_domain, by_tier). Keep the last 500 sessions in the log; archive older entries.
4.6. Clean up tasks: call TaskList, mark ALL session tasks as completed or deleted via TaskUpdate. Never leave stale in_progress tasks.
4.7. Pre-stop verification: confirm execution_summary.yaml exists, coordination_log.yaml has self_validation and validation_checkpoints blocks if a controller ran, status.yaml is in a terminal state, and no stale tasks remain.
4.8. Report results to the user with final pipeline state, revision count, key deliverables, and any validation warnings.
If the pipeline failed after max revisions: report what completed vs what remains, suggest /run --resume {SESSION_ID}, and ensure workflow/recovery_state.yaml captures phase, domain, controller, and pending/in-progress work items (the stop-failure-handler.cjs hook writes this on incomplete sessions).
Step 5: Follow-Up Handling
After reporting results, the pipeline enters a listening state. User follow-ups in the same conversation re-enter execution within the SAME session rather than starting a new one.
Classify the follow-up (adjustment, rework, extension, fix, review) and re-enter the pipeline at the appropriate state. No limit on follow-up rounds.
See @reference/followup-handling.md for the full classification rubric, re-entry procedure, status.yaml/execution_summary.yaml schema additions, and controller follow-up prompt format.
Team Mode (--team flag)
For team mode, after completing routing + planning inline, delegate to /team:
The /team skill handles decomposition into work items, team creation, and parallel execution. Each teammate invokes /run --session {SESSION_DIR} which detects pre-enrichment and picks up from the appropriate state.
If --dry-run with --team: Display plan summary and team composition, then STOP.
Event-driven pipeline: Config-driven state machine with sequential enrichment, nested execution with reviewer loops, and revision routing. TaskCreate/TaskUpdate at every state transition.
