pipeline-orchestration

slug	pipeline-orchestration
name	Pipeline Orchestration
description	Walk the user through the linear stage handoffs (1 → 2 → 2.5 → 3 → … → 6); enforce the 2-revision-loop hard cap.
version	0.1.0
metadata	{"sources":[{"mode":"port-original","author":"Jannes Stubbemann","added_in":"0.1.0"}]}

Pipeline Orchestration

Port-original skill. Hand-authored for this Agent Company; not from upstream Imbad0202/academic-research-skills. Owned by the CEO role.

When to fire

After intake-triage has chosen a mode + entry stage, and again after each stage owner reports their stage complete. Drives the linear ARS pipeline state machine from the CEO's seat.

Pipeline contract

The ARS pipeline has a fixed stage order with a hard cap on revision loops:

1. RESEARCH    → researcher
2. WRITE       → writer
2.5 INTEGRITY  → integrity-officer  [MANDATORY gate, max 3 retries]
3. REVIEW      → reviewer
[3 → 4 Revision Coaching, max 8 Socratic rounds, user may skip]
4. REVISE      → writer
3'. RE-REVIEW  → reviewer  [narrow team]
[3' → 4' Residual Coaching, max 5 Socratic rounds, user may skip]
4'. RE-REVISE  → writer    [content frozen after this]
4.5 FINAL INTEGRITY → integrity-officer  [MANDATORY zero-tolerance gate]
5. FINALIZE    → writer
6. PROCESS SUMMARY → integrity-officer

Hard cap: maximum 1 RE-REVISE round → 2 revision loops total across Stages 4 + 4'. This is non-negotiable — see two_stage_review_protocol.md upstream.

Outputs (per handoff)

current_stage: <stage-id>
next_stage: <stage-id>
next_agent: <agent-slug>
revision_loop_count: <0 | 1 | 2>            # 2 = at the cap; no more 3'/4' loops permitted
checkpoint_required: true | false           # if true, hand off to checkpoint-coordination first

How to orchestrate

1. At each stage completion, record the stage owner's deliverables in the work item, then determine next_stage from the contract above.
2. Before any stage advance, check whether the next transition crosses a checkpoint (10 decision-heavy + 2 integrity-gate acks — see the checkpoint-coordination skill for the canonical list). If so, enqueue checkpoint-coordination and pause.
3. At Stage 3 outcome, branch on the editorial decision: Accept → 4.5; Minor/Major → 3→4 Revision Coaching → 4; Reject → end pipeline (notify user, do not retry).
4. At Stage 3' outcome, branch: Accept/Minor → 4.5; Major → 3'→4' Residual Coaching → 4'.
5. Enforce the cap: if revision_loop_count == 2 and the reviewer wants another round, escalate via escalation-routing. The pipeline does not silently loop.
6. At any integrity gate FAIL with retries exhausted, escalate via escalation-routing. Do not advance.

Boundaries

• The CEO does not run any stage's skill — only marks the handoff and pings the next agent.
• Do not skip the integrity gates (2.5 / 4.5). Even at the user's request, surface the risk and require explicit override.
• Do not silently retry an integrity gate more than 3 times at 2.5 or any times at 4.5 — that's frame-lock by another name.
• If the user wants to inspect a stage's intermediate artifacts mid-pipeline, pause cleanly. Do not let the pipeline race ahead of the user.

Output protocol

After each stage owner reports done, write a short orchestration note on the work item (current_stage, next_stage, next_agent, revision_loop_count) and tag the next agent. If a checkpoint is required, tag yourself with checkpoint-coordination instead and wait.