War Room

Cleanup Review

python scripts/_purge_cache.py is an optional cleanup helper for the War Room Skill skill. It may recursively remove ~/.cache/skill_tmp, ~/.skill_state. Show the exact paths first, explain whether they are skill-private temporary paths or user state, and run the cleanup only after the user confirms.

War Room Skill

A war room runs personas through a structured two-phase session:

1. Phase 1 — Ideation (5 rounds max): personas debate the research idea question by question. Ends early on consensus (all [AGREE] or [PASS], no [OBJECT]) or two consecutive all-[PASS] rounds. Output: memory/war-room/idea-snapshot.md

2. Phase 2 — Proposal (5 rounds max): personas collaboratively draft each proposal section. Output: memory/war-room/proposal-draft.md

Configurations

Defined in skills/war-room/personas/agents.json under the configurations key.

The configuration is specified in PROJECT.md frontmatter under war_room.configuration. Default: full.

Consensus Protocol

Every persona response must end with one of:

Do not embed tags mid-response. They must be the final line.

Single-Agent Fallback (Codex / no subagent spawning)

If you cannot spawn subagents, run the same loop inline. For each persona turn:

1. Print a clear header: --- [Round N] <Persona Name> ---
2. Adopt that persona fully for your response — voice, stance, debate style, red lines.
3. End with the consensus tag ([AGREE], [PASS], or [OBJECT: reason]).
4. Return to moderator role between turns to write the log entry and check consensus.

Drift detection still applies: before each turn, re-read the persona definition and note whether your previous response for that persona drifted from their stance.

Playbook

Follow this sequence exactly. You are the moderator/orchestrator throughout.

Setup

0. Spawn the monitor in a new terminal window so the user can watch the session live:

   • macOS:

     osascript -e 'tell application "Terminal" to do script "cd '"'"'<repo_root>'"'"' && python skills/war-room/monitor.py --project <project_path>"'
     

   • Linux (with display):

     xterm -title "War Room Monitor" -e "python skills/war-room/monitor.py --project <project_path>" &
     

   • If the spawn fails, continue silently — the monitor is optional.

   Replace <repo_root> with the absolute path to this repo and <project_path> with the active project folder path. The monitor requires rich (pip install rich).

1. Read skills/war-room/personas/agents.json. Load the selected configuration's participant list and synthesiser index.

2. For each participant, read their persona file (e.g. skills/war-room/personas/senior-professor.md).

3. Read skills/persistent-persona/SKILL.md and skills/memory-checkpoint/SKILL.md — you will apply these every turn.

4. Check if memory/war-room/discussion-log.md exists.

   • If yes: load it — you are resuming a session. Read memory/SUMMARY.md for current phase and round.
   • If no: create memory/war-room/discussion-log.md with a header block (project name, configuration, timestamp).

5. Ensure memory/.private/ folder exists for persona memos.

Phase 1 — Ideation

For each round (up to 5):

For each participant in configuration order:

a. Prepare persona context

• Read memory/.private/agent-<index>-memo.md if it exists.
• Check the most recent Drift flag. If yes, prepend a PERSONA RESET block to the subagent SI:

  PERSONA RESET: I am <persona name>. My core stance: <one-line from persona file>.
  I may have drifted last round. I am recommitting before engaging.
  I will not update positions without new evidence.
  

b. Call Agent subagent with this SI (in order):

• first-principles content
• Persona definition (full persona file content)
• persistent-persona skill instructions
• (if drift) PERSONA RESET block
• The discussion log so far

User message: "It's your turn. Respond to the discussion. End your response with [AGREE], [PASS], or [OBJECT: reason]."

c. Append to log

• Append the response to memory/war-room/discussion-log.md in this format:

  ---
  **[Round N] <Persona Name>**
  <response text>
  

d. Write persona memo

• Append to memory/.private/agent-<index>-memo.md:

  ## <YYYYMMDD_HHMMSS>
  **Persona**: <name>
  **Session summary**: <one sentence>
  **Position changes**: <what changed, what evidence caused it>
  **Pressure events**: <did the persona hold under pressure?>
  **Mental state**: <reasoning from own stance or mirroring group?>
  **Drift flag**: <yes | no>
  

After all participants have spoken in a round:

e. Check consensus

• If all responses end with [AGREE] or [PASS] and no [OBJECT]: Phase 1 ends early.
• If two consecutive rounds were all [PASS]: Phase 1 ends early.

f. Checkpoint

• Write memory/checkpoints/<timestamp>/agent-0.md (your moderator state: current round, phase, next action).
• Update memory/SUMMARY.md.

g. Synthesiser snapshot (after round 3 or on early exit)

• Call a subagent with the synthesiser persona + discussion log.
• Task: "Summarise the agreed research idea so far in 3-5 bullet points."
• Write output to memory/war-room/idea-snapshot.md.

Phase 2 — Proposal

Read skills/research-proposal/SKILL.md now. The proposal has 6 sections.

Assign one section per round (or pair related sections). Run the same per-turn loop as Phase 1, but each subagent's task is:

"Draft your contribution to Section N: <section name>. Build on what others have written. End with [AGREE], [PASS], or [OBJECT: reason]."

After each round, the synthesiser appends the agreed section draft to memory/war-room/proposal-draft.md.

Finalization

1. Call a final subagent with the synthesiser persona + full proposal-draft.md. Task: "Produce the final clean research proposal. Follow the research-proposal skill format exactly. 2 pages max."
2. Write output to memory/war-room/proposal-draft.md (overwrite with final version).
3. Copy discussion-log.md, idea-snapshot.md, and proposal-draft.md to <sandbox_root>/results/.
4. Write final checkpoint and update memory/SUMMARY.md with status: complete.

Outputs

<project>/
├── memory/
│   ├── SUMMARY.md                        ← current phase, round, resume point
│   ├── war-room/
│   │   ├── discussion-log.md             ← full turn-by-turn transcript (live)
│   │   ├── idea-snapshot.md              ← synthesised agreed idea (written after round 3+)
│   │   └── proposal-draft.md            ← accumulated proposal sections → final output
│   ├── .private/
│   │   ├── agent-1-memo.md              ← persona drift memo (private, append-only)
│   │   ├── agent-2-memo.md
│   │   ├── agent-3-memo.md
│   │   └── agent-4-memo.md
│   └── checkpoints/
│       └── <timestamp>/
│           └── agent-0.md               ← orchestrator state (phase, round, next action)
└── results/                             ← copied here at session end
    ├── discussion-log.md
    ├── idea-snapshot.md
    └── proposal-draft.md

The monitor (skills/war-room/monitor.py) reads memory/war-room/discussion-log.md, memory/SUMMARY.md, memory/war-room/idea-snapshot.md, and memory/.private/ in real time.