| name | multi-character-code-council |
| description | Run a Multi-Character Code Council with four GPT-5.6 Sol reviewer characters and an xhigh skeptical chair. Use only when the user explicitly requests an MCC review, names the Multi-Character Code Council, or directly invokes this skill. Never use for a generic review, second opinion, adversarial review, cross-review, pre-implementation review, or request for multiple perspectives unless the user explicitly asks for MCC. |
Multi-Character Code Council
Run a file-based code review council using GPT-5.6 Sol via the pi CLI with four deliberately incompatible reviewer characters. This is personality diversity, not model diversity: each reviewer owns a different jurisdiction, method, voice, and definition of a valuable finding.
The council runner is a Bun/TypeScript CLI at:
MCC="$HOME/.pi/agent/skills/multi-character-code-council/scripts/mcc"
Invocation gate (strict)
Run this skill only when the user explicitly distinguishes their request as an MCC review. Qualifying requests must directly name one of:
- “MCC” or “MCC review”
- “Multi-Character Code Council” or an unmistakable reference to this specific council
/skill:multi-character-code-council
A generic request for a review, code review, second opinion, adversarial review, cross-review, review before implementation, council/model review, or several engineering perspectives does not qualify. Do not infer that the user wants MCC from the review's complexity or stakes, and do not run the MCC CLI unless the explicit invocation gate is met. Handle non-qualifying requests with the normal review workflow instead.
Once explicitly invoked, the user's request defines the review scope. Do not assume the target is the current git diff. Ask one concise clarifying question only if the target is too ambiguous to review responsibly.
How it works
mcc run creates a global council session under ~/.local/share/pi/multi-character-code-council/.- The runner creates one disposable temp workspace per reviewer, copying the current working-tree state, including tracked and untracked non-ignored files.
- Four reviewer pi processes run in parallel:
- provider:
openai-codex
- model:
gpt-5.6-sol
- thinking:
high
- tools:
read,bash,write,edit,grep,find,ls
- Each reviewer follows a role-specific method and report anatomy, writes its own durable
report.md inside its temp workspace, and stays out of the other reviewers' jurisdictions. The runner copies the report into the council session and deletes the temp workspace.
- After all reviewer processes finish, a skeptical chair pi process runs:
- provider:
openai-codex
- model:
gpt-5.6-sol
- thinking:
xhigh
- The chair reads all reviewer reports, treats repeated same-model claims as correlated evidence rather than votes, cross-examines consensus, preserves dissent, separates facts from intent assumptions, and writes
chair/final.md plus individual issue files.
- The main Pi session reads the chair output, presents issues/proposed solutions/escalations to the human, and asks for approval before editing source code.
Hard-coded reviewer characters
The reviewers are intentionally not full-spectrum clones. Each has exclusive interests, a distinct investigative method, explicit anti-goals, a finding budget, characteristic language, and a role-specific report structure. Out-of-jurisdiction blockers may be noted briefly, but reviewers must not widen into generic review.
ruthless-simplifier — audits abstraction rent, API surface, dependency choices, and existing simpler alternatives; writes blunt KEEP/FLATTEN/MOVE/DELETE judgments and a complexity ledger.failure-mode-red-teamer — runs concrete incident pre-mortems through trigger, propagation, impact, detection, and recovery; ignores aesthetics and unsupported hypotheticals.shipping-intent-advocate — steelmans the intended workflow, protects user value and team velocity, classifies concerns as FIX/ASK/DEFER/REJECT, and pushes back on review-driven overbuilding.contract-prosecutor — derives numbered invariants, constructs minimal counterexamples, and reports only proved contract violations or explicitly unproven obligations.
The chair performs the true adversarial cross-examination because it can see all reports; an isolated reviewer cannot challenge a consensus it has never observed.
Commands
Run a complete council from the target repo or pass --cwd:
"$MCC" run --cwd "$PWD" <<'REQUEST'
<exact user review request>
REQUEST
Useful commands:
"$MCC" status latest
"$MCC" show latest
"$MCC" doctor
"$MCC" doctor latest
"$MCC" cleanup-stale latest
"$MCC" retry latest
"$MCC" latest
Debug option:
"$MCC" run --keep-workspaces --cwd "$PWD" --request "Review ..."
If a terminal disconnect, timeout, or abort interrupts the parent runner, status computes live state from per-process metadata. It marks abandoned IN_PROGRESS sessions as STALE when no reviewer/chair PID is alive and required outputs are missing. Use cleanup-stale to remove known temp workspaces and mark the session ABORTED; use retry to start a new session from the same request.
Environment knobs for the runner itself:
MCC_REVIEW_ROOT="$HOME/.local/share/pi/multi-character-code-council"
MCC_PI_BIN=pi
MCC_PROVIDER=openai-codex
MCC_MODEL=gpt-5.6-sol
The model/provider are explicit so the council does not silently change if the user's interactive pi defaults change.
Review folder
~/.local/share/pi/multi-character-code-council/YYYY-MM-DD-HHMMSS-repo-name/
README.md
request.md
status.txt
run.json # written at start; updated on completion/failure
reviewers/
ruthless-simplifier/workspace.txt
ruthless-simplifier/report.md
failure-mode-red-teamer/report.md
shipping-intent-advocate/report.md
contract-prosecutor/report.md
chair/
workspace.txt
final.md
issues/01-short-title.md
logs/ # stdout/stderr/meta JSON with PIDs and exit codes
prompts/
Non-negotiables
- Pass the user's exact review request/scope to the council.
- Create persistent council artifacts under
~/.local/share/pi/multi-character-code-council/, never inside the project unless explicitly asked.
- Reviewers and chair may use bash only in disposable temp workspace copies.
- Reviewers/chair must not intentionally mutate the real repo, home directory, global config, databases, cloud resources, Docker services, package registries, or network services.
- The main Pi session must not edit source code during the council.
- After the council, read
chair/final.md and relevant chair/issues/*.md, present the findings/proposed fixes/escalations, and ask the human for approval/input before implementation.
Main Pi workflow
- Capture the user's exact review request.
- Run
"$MCC" run --cwd <target-dir>.
- Read
chair/final.md and individual issue files if needed.
- Present:
- top issues
- proposed solutions
- corroboration/which perspectives raised them
- escalated questions or tradeoffs
- verification plan
- review folder path
- Ask for approval before editing code.
Final response pattern
Multi-character code council complete.
Top issues:
- ...
Proposed implementation plan:
- ...
Escalated questions:
- ...
Review folder: <path>
Ready for me to implement the recommended changes?
