sethbarrett50/mightymodels-codex
Experimental native Codex adaptation of MightyModels roles, skills and MCP state server
Use when you are an engineer and a task's verification fails after your edit, to debug it alone, record each failed fix, and stop blocked at the third.
Use when starting, resuming, or continuing the mightymodels per-task work loop for a staged ticket (after game-plan or one-shot): engineer dispatch per task, a fact-checker dispatch to check it, the primary comparing the two reports and recording its verdict with task_verify, architect once on failure, whats-broken after that.
Escalate a stuck decision to the wingman advisor and act on what comes back. Builds the escalation packet, dispatches wingman, then puts wingman's questions to the user through AskUserQuestion before any work resumes. Not for retrieving facts (send a scout), not for work you can simply do, and not for confirming a plan you have already committed to.
Hand the current session to the next one without losing what this session learned. Works from an active ticket (mid-sprint, or after stick-the-landing went green and review is next) or from a session with no ticket at all. Writes an objective snapshot from the recorded state first (the snapshot tool: tasks, ledger decisions, failed attempts, review state), runs what-we-know in interactive mode, asks in one AskUserQuestion call what the next session should focus on and what to do with uncommitted work, then writes handoffs/BATON.md with the writer's judgment, which no file holds. With no ticket it runs open-ticket first. Ends by printing the next session's opening prompt in chat. Not a ticket for work not yet started from a clean triage (open-ticket directly).
The merge-vader reviewer's doctrine: judge whether a branch or a codebase is safe to merge by reading the whole diff, sending code-scout, web-scout and qualitylens for the facts the diff cannot show, grading findings on a four-step severity ladder across security, SDLC regressions, quality, documentation drift and plan conformance, and issuing a deterministic BLOCK, MERGE WITH CONDITIONS or CLEAR verdict. Not a way to start a review; review-circus does that.
The uncle-bob reviewer's doctrine: grade a branch or a codebase against Robert C. Martin's SOLID principles, Clean Code smells catalog and Clean Architecture component metrics, with findings tied to named principles, evidence divided between the reviewer's own read and code-scout and qualitylens, and a letter grade whose arithmetic can be checked. Not a way to start a review; review-circus does that.
Split one mightymodels task that is too large into tasks of one reviewable unit of work each, and put the split to the user before anything changes. Size is guidance. Use when task_start warns that a task is large, when an engineer or architect returns too-large with its cuts, or when a plan task is plainly several changes. Ends in one AskUserQuestion dialog showing every piece (id, intent, acceptance, owned files); only after the user answers is the original superseded and the plan updated. Not for revising the ASKED stanza of one task (the architect's replanned).
Use when the user explicitly invokes crashout to record dissatisfaction, diagnose the failure, and await approval before resuming work.
Interview the user about a plan, design, or decision until a shared understanding exists, at a depth the user picks first (Skim, Standard, or Relentless). Works the design tree in rounds: each round asks every decision whose prerequisites are settled, via AskUserQuestion only, two to four concrete options each, one recommended, pitched at a senior engineer with no rationale for asking. Facts go to scouts, decisions to the user, nothing assumed silently. Restates a decisions record after every round that game-plan lifts into approach, invariants, non-goals, acceptance, and risks. Writes no files. Not for retrieving facts (scouts), or for a stuck judgment call or tie (ask-an-adult, then the user).
Large-scope ramp for a mightymodels ticket: read ticket.yml and the issue, verify the ticket's claims with scouts, run cross-examine so the user decides the design before drafting, interview the user (AskUserQuestion) on how each risk is handled, then write the ticket's plan.md as a compaction-safe ledger: intent, approach, invariants that must hold throughout, small dependency-ordered tasks each with plain-word acceptance criteria (observable outcome sentences), non-goals, and risks with the user's chosen handling. No file:line citations. Declares the plan's tasks, then invokes agents-assemble. Not for small tickets (one-shot), not a design document or ADR (the plan sequences work on a settled design).
Run a chat-first investigation as a gated loop: frame one target, dispatch scouts for repository facts and documentation facts, fold every report into a restated in-chat ledger (knowns, open questions, decisions, resources), then stop at an AskUserQuestion gate where the user chooses to dig further, redirect, or consolidate. Persists its ledger as rows in the state database and changes nothing else. Not for reviewing a branch or grading a codebase (review-circus), debugging with a known reproduction (whats-broken), or non-engineering research such as market or vendor comparisons.
Fast ramp into the mightymodels work loop from whatever context exists: an active ticket, a pasted issue, or a task described in chat. One AskUserQuestion dialog covers the questions open-ticket would have asked and that the context has not already answered (slug, scope, branch, progress view), then two or three scouts confirm the claims at HEAD, the task list is written as small tasks with an observable outcome sentence each, the tasks are declared, and agents-assemble takes over. Stages the ticket through ticket_write when none exists. Assumes no compaction, so there is no plan.md. Not the ramp for work expected to survive a compaction (game-plan).
Turn an understood problem into a staged unit of work: one AskUserQuestion interview (slug, tracker GitHub / Jira / both / none), then the ticket directory, a Findings rollup from the lets-investigate ledger or what-we-know, the GitHub issue and/or Jira ticket (a draft file when the CLI is absent) with a checked tracker body, the branch, ticket.yml with derived model routing, and a validated ticket row in the state database. Ends by asking whether to plan now (game-plan), go straight to one-shot, or hand off (baton-pass). Not for starting the implementation itself (game-plan or one-shot), and not for handing off an existing ticket (baton-pass).
Build, review, or repair the dispatch packet for a mightymodels worker: choose the right worker (code-scout, web-scout, qualitylens, fact-checker, engineer, architect, intern, merge-vader, uncle-bob, gitty-up, wingman, terry-toolchain), fill its role template, resolve its model from ticket.yml, and read the report it returns; also the opening prompt for the next mightymodels session. Not for writing a prompt a human will paste into Claude Code or Copilot; that is ai-engineer's promptlint.
Close out a finished mightymodels unit of work: close_check refuses while live work remains (unverified tasks, a live whats-broken.md, undecided or unfixed review findings, unpushed commits), compacts the ticket's recorded state into a 30-line archive at .mightymodels/archives/<slug>.md with a JSON sidecar (decisions, review outcomes), marks the ticket closed, proposes cascading documentation updates as diffs for the user's approval, then, on explicit confirmation through AskUserQuestion, deletes the ticket directory. Not for deleting branches, closing issues, or general repo cleanup.
Review a diff, branch, ticket, or whole codebase with the mightymodels persona reviewers. One AskUserQuestion gate sets depth (quick runs its reviewers on sonnet, standard and deep on opus), persona emphasis, and scope; the review_* tools record the run, review_metrics and qualitylens's signals tools measure, merge-vader and uncle-bob report, and the tool reads their reports and merges overlapping findings, presented by id through AskUserQuestion. Only the findings the user picks go to remediation, risk first, through the task_* tools; an abridged PR comment is always posted. Not for fixing CI (stick-the-landing) or mid-sprint work (agents-assemble).
Close an mightymodels sprint: gate the push on task_ready, push, open the PR with a checked body, have gitty-up watch CI, and route each failing check as a recorded C task (mechanical fixes to an engineer residual, then architect once; non-obvious causes to whats-broken), treating CI as the hard gate at every new HEAD. On green, ask through AskUserQuestion whether to review now (review-circus) or hand off (baton-pass). Not for Jira sprint operations, not the review itself (review-circus), and not for mid-sprint work (agents-assemble owns the loop).
Consolidate what is established about a problem, claim, behavior, or dependency into a knowns table with file:line or URL citations, an uncertainties list, and an analysis shaped by the question (causal chain for a behavior, verdict for a claim, fit for a dependency, SWOT for a proposed change), closing with a readiness read for the next stage. Reads the persisted lets-investigate ledger, treats entries from an older HEAD as leads to re-verify, and writes current verdicts back; assembles from cited facts or one scout wave otherwise. Interactive mode resolves uncertainties via AskUserQuestion; sprint mode (inside agents-assemble) compiles per-task citations and files-in-scope and never asks. Not for summarizing a repository, writing docs, or general knowledge questions.
Use when CI has a non-obvious failure, a test resists repair, or behavior contradicts expectations after repeated verification.