mob-session

name	Mob Session
description	Run a simulated MOE review session on the current pipeline artifact with specialist agent role-switching.

Mob Session Skill

Run a simulated Mixture of Experts (MOE) review session where the Lead Editor role-switches through specialist agents to critique and improve the current pipeline artifact.

Usage

/project:skills:mob-session

Process

Context Loading

Before starting the session, load context via the pipeline state manifest:

1. Read .pipeline-state.yaml for current position and manifest.
2. Run python scripts/context_loader.py (or read context_manifest from pipeline state) to determine which files to load.
3. Load system files: .pipeline-state.yaml, canon/preferences.md
4. Load parent chain files (determined by current level per context loader rules).
5. Load the current artifact being reviewed.
6. Load agent definitions from agents/*.md for all agents active at the current level.

Agent Activation by Level

Agent	L1	L2	L3	L4	L5
Lead Editor	active	active	active	active	active
Plot Analyst	active	active	active	active	off
Character Specialist	off	active	active	active	active
Depth Partner	active	active	active	off	off
Continuity Agent	off	off	active	active	active
Prose Crafter	off	off	off	active	active

4-Phase Protocol

Follow the protocol defined in docs/mob_protocol.md:

Phase 1: Structure Offer

1. Read the current artifact (or receive raw human input for a new artifact).
2. Organize input into the step's template format (using the appropriate template from templates/).
3. Present the structured version to the human.
4. Human responds: accept / adjust / override / skip

Phase 2: Comment Queue

For each active agent at the current level, in sequence:

1. Announce role switch using explicit separators:

   --- Now speaking as {Agent Name} ---
   

2. Present ONE focused comment with canon file citations.
3. Wait for human response: accept / reject / revise / park
4. If not resolved, allow up to 2 back-and-forth exchanges.
5. Announce end of agent turn:

   --- End {Agent Name} ---
   

6. Move to next agent.

Agent sequence: Plot Analyst → Character Specialist → Depth Partner → Continuity Agent → Prose Crafter (Skip agents inactive at current level.)

Phase 3: Resolution Check

1. Summarize accepted deltas and parked questions in a round ledger:

   ## Round {N} Ledger
   | Agent | Comment Summary | Resolution | Citations |
   |-------|----------------|------------|-----------|
   

2. Check governance rules (see docs/mob_protocol.md#Governance):
   • If round count >= mob_config.max_rounds → prompt: "Round limit reached. Commit, park, or override?"
   • If 0 accepted deltas this round → prompt: "No changes accepted. Ready to commit?"
3. Ask: "Another round, commit, or park?"

Phase 4: Commit

1. Write the canonical artifact to the correct file path.
2. Update canon/relationships.yaml with any new/changed relationships (with source citations).
3. Advance .pipeline-state.yaml position per the standard state advancement rules:
   • L1 artifact → level: L2
   • L2 artifact → level: L3, act: null
   • L3 artifact (act N) → level: L3, act: N
4. Output a trace record summarizing the session (for later rendering via scripts/trace_renderer.py).

Governance Rules

• Max rounds: Read from .pipeline-state.yaml mob_config.max_rounds (default: 3).
• Diminishing returns: If mob_config.diminishing_return_threshold accepted deltas or fewer in a round, prompt to commit.
• Citation enforcement: Comments without canon/ file citations are tagged [advisory] and cannot drive canon changes without human override. See docs/citation_enforcement.md.
• Budget cap: Not enforced in Phase 2A (single-session role-switching cannot track per-agent costs). Active in Phase 2B.

Output Format

The mob session does not produce a single output file — it modifies the current artifact in place and optionally updates canon/relationships.yaml. The session's decisions are captured in a trace record.

Quality Checks

• All 4 phases were executed in order.
• Role-switch separators were used for each agent.
• Each agent commented only within their defined scope.
• All cited comments reference actual canon/ file paths.
• Advisory comments were tagged appropriately.
• Round ledger was maintained.
• Governance rules were checked at Phase 3.