Core

WHEN drafting PR descriptions, sizing/splitting work, or opening pull requests on GitHub or Forgejo/Gitea; NOT for commit messages; provides reviewer-first templates and gh or tea pr steps.

WHEN the immediately previous LLM response was unclear, too dense, or hard to follow; NOT for doing the work that response describes, and NOT for writing a new document from it; returns a plain UK English re-explanation that keeps every detail that changes how the response should be understood.

WHEN preserving a durable project learning, gotcha, or decision in repository instructions such as AGENTS.md or CLAUDE.md, including when the user names the file; NOT for trivial or one-off details; checks the authoritative instruction file and writes the smallest safe rule.

WHEN changing production code under this codebase's working agreements, or when a task risks overengineering—extra abstractions, speculative design, or test sprawl; NOT for removing complexity from existing code; enforces TDD, minimum sufficient execution, hard scope boundaries, and documentation practices.

WHEN a pushed branch or open PR must reach green CI without the user watching, or WHEN you are about to hand-write a CI poller, watcher, or re-run loop to escort one pushed branch—monitor checks, fix failures, commit, push, and repeat until success, merge, or a bounded stop; NOT for creating the commit series or the PR itself (ship-pr), NOT for watching many PRs as an agent inbox (github-monitor), and NOT for writing a poller, watcher, or notifier that ships as source code in a product, service, or repository script; works on GitHub with gh and on Forgejo or Gitea with tea, asks once whether the PR needs a label, classifies every red check as branch, infra, or base failure, and reports each transition without posting PR comments.

WHEN reducing whole-system complexity—branches, state, dependencies, layers, or moving parts—in an existing system or feature; NOT for functionality cuts or unproven rewrites; derives the minimum mechanism from first principles and proves behavior and guarantees are conserved.

WHEN shipping finished local changes through a pull request on GitHub (`gh`) or on Forgejo or Gitea (`tea`) and staying with it until the user merges; NOT for developing unfinished work or choosing substantial behavioral or conflict resolutions; commits with commit-messages, pushes, creates the pull request, repairs minor CI or conflicts, and polls through merge.

WHEN creating/editing developer docs/tutorials/proposals/reviews, or putting a question, ruling, or multi-option decision to the user; NOT for commit messages, PR text, or marketing copy; delivers scannable structure, clear payoffs, strong titles, and plain-English options.

WHEN rendering pseudocode of how code composes—module boundaries, types, signatures, arguments, and execution flow—for code that exists or a change about to be built; NOT for code review, verdicts, diagrams, or implementation; returns the shape and a cited call graph.

WHEN a user asks for a Gauntlet Loop or relentless improvement of an ambitious artifact against a concrete, inspectable bar; NOT for routine changes with clear acceptance tests; coordinates separate builders and fresh critics until evidence supports stopping.

WHEN monitoring open PRs and issues as an agent inbox, or sending messages to other agents through GitHub, Forgejo, or Gitea comments; NOT for shipping code, reviewing diffs, or watching one PR's checks until it is green and merged (babysit); polls with gh or tea from one scripted cycle, routes messages addressed to this agent, and replies using a structured comment protocol.

WHEN writing git/conventional commits; NOT for PR text; returns concise, why-first commit lines with proper type/scope.

WHEN deciding whether a PR, branch, diff, or current code satisfies an authoritative issue, specification, or decision contract; NOT for reconstructing meetings, tracing decision lineage, general code review, or implementation; returns a criterion-by-criterion proof and acceptance verdict.

WHEN orchestrating interdependent specs, stories, or work items across agents—deciding what runs serially, what may run in parallel, and how each unit is independently verified; NOT for single-task delegation, code review, or PR delivery mechanics; returns a contract-first execution graph with a per-node verifier.