catalyst

Use when executing a spec or multi-task implementation plan through subagents, when writing such a plan, when verifying a completed branch against its original goal, or when a user-ordered AMENDMENT to an approved spec must be routed (any time, including mid-execution or after acceptance). Triggers - implementation plan exists or is requested, work spans multiple tasks, subagent dispatch is available.

Use at the start of work on a new or unfamiliar project - assess agent-readiness, fill tooling gaps, and route to the right Catalyst skill. Triggers - "let's start the project", fresh clone, no build/test/lint commands known.

Use for a program of work spanning many sessions and multiple arcane-mode runs - several phases toward a milestone, "what's next" when a campaign exists under .catalyst/campaign/, or closing out a milestone. Not for a single feature (crucible → arcane-mode), a single foggy decision map (starchart), or resuming ordinary non-campaign work (handoff).

Use when a feature idea, architectural decision, or spec needs hardening before planning - requirements are fuzzy, trade-offs unresolved, or the user wants their thinking stress-tested. Also use before catalyst:arcane-mode when no approved spec exists, and when changing or amending an ALREADY-APPROVED spec that has NO plan or execution yet (the amendment fast path - never hand-edit an approved spec); once a plan, ledger, or campaign phase exists on the spec, the amendment enters via catalyst:arcane-mode's amendment route, which invokes this fast path itself and then reconciles the plan/ledger/campaign state the spec change touches.

Use when facing a bug, test failure, performance regression, or unexpected behavior whose cause is not yet proven - before proposing or applying any fix. Also when a previous fix didn't hold, or debugging spans context resets. Triggers - "why is this failing", "it broke", "it's slow", flaky behavior, a fix that didn't stick.

Use when creating a new skill or agent definition, editing an existing one, or verifying that a skill actually changes behavior - for this family or project-local skills. Triggers - "add a skill", "the skill isn't working", "update the skill", packaging session/spike findings as a skill.

Use when ending a session with unfinished work, before context compaction, or when resuming from a previous session's handoff document. Triggers - session wrap-up, context near limit, "continue from last time".

Use when setting up or repairing the Catalyst/ContinuousClaude toolchain - a fresh machine, missing binaries, binary-dependent features degrading, or "install/update the toolchain". Also the doctor - run it to diagnose which tools actually respond.

Use when a plan or spec is ready and implementation has not started - failure-state projection before work begins. The gate between an approved plan and dispatching Task 1. Also gate RE-RUNS on amended or rejected ground, and REVERSALS of a previously accepted risk (the user withdraws an acceptance) - any time, including mid-execution or after the work shipped (the season rule bars retroactive analysis of shipped work, not re-gating amendments or honoring withdrawals).

Use for open-ended investigation where the answer is not yet known - exploring a domain, comparing approaches, understanding an unfamiliar system. Not for executing toward a known goal.

Use for standalone code review outside the arcane-mode pipeline - "review this code", "review my PR", "check before merge", "what did I break" - and for cleanup requests like "clean this up / remove the AI slop" (the deslop pass). For task review inside the pipeline use arcane-mode's critic flow instead.

Use when an effort is too large for one agent session and the route is still foggy - you can feel the shape of the work but cannot yet write it down as a spec or a plan. Also for re-entering a standing decision map - a committed .catalyst/map/ directory in the repo. Not for well-scoped features.

Use when extending the ouros sandbox with a new external function - "add X search to the sandbox", "database query function from sandbox", "extend sandbox with X". Requires the ContinuousClaude binary layer (tools/ouros_harness.py present in the project).

Use when starting a session or deciding which Catalyst skill a task needs - the family's process-skill routing and its coordination with other libraries. Injected at session start.