agent-profile

Use when the user says they are stepping away and the agent should continue without interactive approvals

Use when running e2e tests, debugging test failures, or fixing flaky tests. Covers failure taxonomy, fix rules, and workflow. Never changes source code logic or API without spec backing.

Use when auditing how well the shared agent instructions (AGENTS.md, skills) hold up in real sessions — sampling transcripts via recall, scoring them against the gap rubric, and turning findings into ratified amendments

Use when preparing clean, logical git commits from an existing working tree

Use when syncing a feature branch onto the latest origin base branch via git rebase.

Fetch latest from origin, prune remote-tracking refs, delete stale local branches and worktrees, and fast-forward important branches. Use when tidying up a worktree-based repo layout.

Use when preserving enough task context for another agent or future session to continue without the current conversation

Use when a non-trivial changeset is ready for a human to review in hunk (the terminal diff viewer), or the user is reviewing your changes in a live hunk session.

Use when operating on live systems - cutovers, migrations, production config changes, or incident triage

Use when the user asks to rebuild a feature branch into a clean, reviewable git history

Use when a SPEC.md exists or is needed and the user wants gaps found through an interview before implementation

Use when checking deployment health, investigating errors, reading logs, or working with Tiltfiles. Queries Tilt resource status, logs, and manages dev environments.

Use when starting tilt, debugging Tiltfile errors, or bootstrapping a dev environment. Starts Tilt in zmx, monitors bootstrap to healthy state, fixes Tiltfile bugs without hard-coding or fallbacks.

Use when requirements are clear enough to plan and the work spans multiple steps, files, or verification stages

Use when creating or editing a skill and you need it to be discoverable, concise, and native to the target harness

Use when starting dev servers, watchers, tilt, or any process expected to outlive the conversation. Provides zmx session management patterns for long-lived processes.

Use when creating, reviewing, or updating a BRIEF.md (the quality law for a surface), defining what "good"/shippable means, or setting up a verified autonomous loop.

Use when creating commits, managing branches, opening PRs, or rewriting history. Not for non-git implementation tasks or repo-specific release policy decisions.

Use when creating, reviewing, or updating SPEC.md files or entering the ADF SPEC gate.

Use when designing tests, writing test cases, planning test strategy, or attributing test failures. Covers unit, integration, and e2e layering plus verifier discipline (flake attribution, base-commit repro, broken-verifier handling).

Use when a feature branch or campaign needs an autonomous multi-session loop — the user asks for a LOOP.md or a self-driving plan to finish a branch, says they are stepping away with work in flight, or a session must run or resume a campaign that has a LOOP.md (legacy path .claude/loop.md).

Use when reading or writing Rust files (.rs, Cargo.toml).

Use when reading or writing Zig files (.zig, build.zig, build.zig.zon).

Patterns for invoking the GitHub CLI (gh) from agents. Covers structured output, pagination, repo targeting, search vs list, gh api fallback.

Use when working with Nix flakes, overlays, shell.nix, or flake.nix files.

Use when reading or writing Go files (.go, go.mod).

Use when reading or writing Python files (.py, pyproject.toml, requirements.txt).

Use when reading or writing TypeScript or JavaScript files (.ts, .tsx, .js, tsconfig.json).

Fetches web content as clean markdown by preferring markdown-native responses and falling back to selector-based HTML extraction. Use for documentation, articles, and reference pages at http/https URLs.