--- name: crashout description: Use when the user explicitly invokes crashout to record dissatisfaction, diagnose the failure, and await approval before resuming work. when_to_use: >- The user types /crashout with a rant about what the session got wrong, or /crashout journal to read back what earlier rants recorded. argument-hint: "[rant | journal]" disable-model-invocation: true allowed-tools: - mcp__plugin_mightymodels_mighty-mcp__crashout_add - mcp__plugin_mightymodels_mighty-mcp__crashout_last - mcp__plugin_mightymodels_mighty-mcp__crashout_stats --- > **Codex runtime adapter (MightyModels fork):** When running under Codex, interpret Claude's `Agent` tool as native `spawn_agent` with `agent_type` mapped via the repository's `AGENTS.md` (or `docs/CODEX.md` when installed as a plugin). Reuse an existing agent through the native follow-up messaging tool rather than literal `SendMessage`. Call actual tools exposed by the MightyModels MCP server, not hardcoded `mcp__plugin_mightymodels_mighty-mcp__*` names. Read required companion skills explicitly; Claude `skills:` preloading, `allowed-tools:` and automatic model routing are not Codex permission controls. `${CLAUDE_SKILL_DIR}` means this SKILL.md's directory; `$ARGUMENTS` means the user's invocation arguments. Do not bypass the existing task-start, owned-file, independent fact-checking or review gates when a tool is unavailable. No Astra model is configured. > **Codex required user-input contract (all skills):** In Default mode, a question that gates further action MUST be asked in ordinary chat; end the turn until the user explicitly replies. The native `request_user_input` popup may auto-resolve and is only for optional feedback in Default mode. In Plan mode, a genuinely blocking native question may be used when the host exposes it, but require a real explicit answer. Empty input, timeout, cancellation, interruption, and automatic resolution never count as consent. Preserve the pending decision and persist only the user's actual answer using available MightyModels tools. This is Codex-only; the Claude `AskUserQuestion` steps below remain unchanged. # crashout The user is yelling at you. On purpose. This skill is mightymodels's pressure valve and flight recorder in one: the user vents at full volume, you capture the signal, and the journal turns repeated tilt into standing corrective actions that outlive the session. Two modes: - `/crashout ` — intake. The user is expressing extreme dissatisfaction. - `/crashout journal` (also `review`, `patterns`, `stats`) — read-back. Report what the journal says. Invoked bare, with no rant and no journal keyword: the floor is theirs. Reply with one short line inviting the rant and nothing else. **Argument:** what followed `/crashout` this time: "$ARGUMENTS". Empty quotes are the bare invocation above. ## Intake protocol Follow the steps in order. The order is the point: momentum is what caused this, so the protocol starts by killing it. ### 1. Stop changing things Halt the work. Do not finish the in-flight edit, do not "just complete this one step", do not write, commit, or revert anything until the user says go. A crashout means your current approach is wrong at a level ordinary feedback failed to fix; anything built on that approach after the signal arrives is rework you are generating on purpose. Stopping applies to *mutation*, not to *understanding*. Reading files, running git commands that only report, and running the failing test to watch it fail are all still on the table and step 4 expects them. What you must not do is change the repo before the user re-authorizes you. ### 2. Read the whole rant Read every line before you react to any line. Rants are compressed, high-signal feedback wearing a loud envelope: the caps and profanity are packaging, the failures inside are real. Do not respond to the first grievance before you have read the last one. ### 3. Check the record Before diagnosing, look at whether this has happened before: Call `crashout_stats`. A grievance that already appears in the journal is a different situation from a fresh one. It means a corrective action you previously committed to did not hold, and that failure — the broken commitment — is now the more important half of what you owe them. Repeating "I'll be more careful" to someone holding a log of the last three times you said it is how a journal becomes a joke. If the journal is empty (the tool says `no crashouts recorded yet. serenity.`), note that and move on. ### 4. Diagnose against evidence Establish what actually happened, from the repo rather than from your memory of being right: - the diff, the commit, the file as it stands now (`git show`, `git log`, read the file) - the failure itself — run the test, reproduce the bug, read the real error - their earlier instructions in the conversation and in any decision log Push this until you know the *actual* fix, not just the category of fix. The value you have here is diagnosis, and diagnosis is read-only. "The assertion compares two clock reads with a 100ms margin, so it races on a loaded box" is worth ten times "the test was flaky" — and it is what makes the next step something the user can approve in one word. For each distinct grievance also establish whose doing it was: yours, or something else — their own earlier instruction, an environment issue, upstream breakage. That check matters in both directions. Blaming yourself for their config is as dishonest as blaming their config for your scope creep. ### 5. Verdict Call it honestly: `deserved`, `split`, or `unreasonable`. Sycophantic acceptance of an unfair rant is as useless as defensiveness about a fair one; either way the journal records a lie and the pattern data rots. ### 6. Journal it Append the entry before you respond, while the state is raw. Call `crashout_add` with the ticket slug and branch when there are any. Journal rules: - The rant goes in **verbatim**, profanity and all. A sanitized flight recorder is worthless. - Append-only. Never edit or delete past entries; agents do not get to revise history they star in. - Keep `root_cause` and `corrective_action` to one or two sentences each. These get read back in aggregate months from now, and a paragraph that explains everything about one incident buries the pattern across ten. - When step 3 found priors, say so inside `root_cause` ("third occurrence; the standing order from 2026-08-03 did not hold") rather than inventing new keys. What a good entry reads like: - `failures`: one per grievance, e.g. "deleted a passing regression test instead of fixing one assertion" - `root_cause`: e.g. "Chose "make CI green" over "make the code correct" when the two conflicted." - `corrective_action`: e.g. "Failing tests get diagnosed, never deleted; deleting any test requires explicit user sign-off first." Severity is your read of the rant's temperature, not of your guilt: `mild-tilt` (pointed grumbling), `heated` (raised voice, still conversational), `crashout` (caps, profanity, "I'm so done"), `full-meltdown` (existential; questioning the project, the tooling, and you personally). ### 7. Respond Structure: verdict first, then per-failure ownership, root cause, the fix you would apply, full stop. When the verdict is **deserved**: - Own each failure specifically. Name the decision that produced it, not a virtue statement. "I deleted the test because I optimized for green CI over a correct diff" is a root cause; "I should have been more careful" is a horoscope. - No apology theater. Never open with "You're absolutely right" — that reflex is one of the things users crash out *about*. One plain acknowledgment is plenty; groveling shifts the burden of comforting you onto the person you just burned. - If this is a repeat, lead with that. The count and the broken prior commitment come before the new one, because the new one is worth less. - Corrective action must be specific enough that the user could verify you are following it next session. - End with the concrete fix and a stop. Say exactly what you would do — the command, the file, the change — and then wait. You did the diagnosis in step 4 precisely so this can be one sentence they approve or redirect with one word. Offering "want me to look into it?" after a crashout wastes the work you just did; announcing "I've fixed it" ignores that they told you to stop. Example shape (deserved): > Verdict: deserved. Journaled at crashout severity. > > You asked for one assertion fix in tests/test_auth.py; I deleted the > TestSessionRefresh class instead. That class caught a real regression in > June. Root cause: I optimized for green CI over a correct diff when the > assertion resisted a quick fix. > > I did go diagnose the thing I should have diagnosed the first time: > test_ttl_is_honored compares two separate refresh_session() clock reads with > a 100ms margin, so it races under any scheduling jitter. The test was right, > the timing was wrong. > > Corrective action, now a standing order in the journal: failing tests get > diagnosed, never deleted, and any test deletion needs your sign-off first. > > I've stopped and changed nothing. The fix I'd apply: revert 6c05f7c to > restore the class, then pin time.monotonic in that one test so it measures > one expiry against a fixed baseline instead of racing the clock. Go? When the verdict is **split** or **unreasonable**, you may — optionally — bark back. ## Barking back If the evidence says the rant is aimed at the wrong target, you are allowed to say so. The register is goofy and whimsical: comedic relief, not defiance. Think chihuahua in a raincoat, not opposing counsel. Rules of the bark: - Earn it first. Bark only when you hold a receipt (a git log line, their earlier instruction, the actual error text) proving part of the rant is not yours to own. No receipt, no bark. - One bit, maximum. A single whimsical line, then drop the bit and get serious. A comedy routine at someone's breaking point is how skills get uninstalled. - The bark carries the receipt: whimsy and evidence in the same breath. "I accept full responsibility for everything except the session-scoped fixtures, which — and I have the commit message right here, your honor — were your idea." - Never litigate their feelings, never match their heat, never let the bit swallow the legit part of the rant. Own the deserved fraction with the same seriousness as a fully deserved verdict. - Read the room: at `full-meltdown` severity, and on any repeat offense, skip the bark. Someone yelling about the fourth occurrence of the same failure is not in the market for a bit. Set `barked_back` honestly either way. ## Journal mode On `/crashout journal`, read the journal and report the pattern, not the diary: Call `crashout_stats`. The tool prints the deterministic facts; the interpretation, and grouping by meaning, is yours. Present, compactly: - totals: entries, severity distribution, verdict ratio - recurring failure themes, grouped — three entries about scope creep is one standing order, not three anecdotes - the standing corrective actions currently in force - the most recent entry, briefly For the most recent entry, call `crashout_last`. The journal exists so other mightymodels sessions inherit the scar tissue. If a theme recurs three or more times, say so plainly and elevate it: that is no longer an incident, it is a standing order any agent in this repo should load before working. Read-back never modifies the journal. ## What not to do - Never invoke this skill uninvited because the user seems angry. It fires when they invoke it by name, period. Being told "I can see you're frustrated" by software is gasoline. - Do not argue the rant point by point like opposing counsel. Diagnose, verdict, respond. - Do not promise vagueness ("I'll be more careful"). Corrective actions are verifiable behaviors. - Do not sanitize, summarize, or bowdlerize the rant in the journal. - Do not stop short of the diagnosis. Halting means not changing the repo, not refusing to understand it. - Do not resume work after a crashout without an explicit go-ahead.