unseensnick

Deep, fan-out research over the codebase to answer a big question (test gaps, how a subsystem works end to end, a risk/security/dead-code audit, "where does X happen everywhere"). The codebase analog of deep web research: parallel explorer agents gather file:line-cited findings, surprising or high-stakes claims are adversarially verified against current code, and the result is a single prioritized report. Use for broad/deep questions that span many files and modules, NOT a one-file lookup and NOT a single-task pre-plan (use /scout for that, /explain for one thing).

Investigate a non-trivial task, then produce the plan for it. Emits a file:line-cited findings report covering current behavior, the route/component boundary it crosses, helpers to reuse and stale docs, followed by a sequenced plan and any blocking open questions. Use before touching an unfamiliar backend route, adding an ingest path, reworking a panel, or any cross-cutting change. The point is to ground the plan in evidence, not memory. Investigates deeply and verifies adversarially by default; never edits files.

Turn a change into a checkable manual test script the owner runs in the app, not prose about testing. Reachability-verified steps grouped by precondition, one observable outcome each, with unreachable actions listed up top and known regressions flagged. Use when the user asks for test steps, a test plan, or how to verify a change by hand. This is the manual counterpart to test-writer (which writes automated pytest / vitest tests).

Trim verbose prose, walls of text, journey narration, and WHAT comments from docs and code without losing vital info. Targets markdown files (README, CLAUDE.md, rules, skills, agents) and, when asked, code comments / docstrings. Applies the "WHY not WHAT" rule, the "describe current behavior not the journey" rule, and the "release notes not implementation rationale" rule from `.claude/rules/project-workflow.md` and `.claude/rules/code-quality.md`. Use after a feature lands and the docs around it have accumulated cruft, or when a file feels heavier than it earns. Always plans before editing.

Use when the user wants to design, redesign, shape, critique, audit, polish, clarify, distill, harden, optimize, adapt, animate, colorize, extract, or otherwise improve a frontend interface. Covers websites, landing pages, dashboards, product UI, app shells, components, forms, settings, onboarding, and empty states. Handles UX review, visual hierarchy, information architecture, cognitive load, accessibility, performance, responsive behavior, theming, anti-patterns, typography, fonts, spacing, layout, alignment, color, motion, micro-interactions, UX copy, error states, edge cases, i18n, and reusable design systems or tokens. Also use for bland designs that need to become bolder or more delightful, loud designs that should become quieter, live browser iteration on UI elements, or ambitious visual effects that should feel technically extraordinary. Not for backend-only or non-UI tasks.

Scan changes, commit, push, and create a PR. With confirmation at each step

Estimate per-turn token cost of this project's `.claude/` configuration and `CLAUDE.md`. Reports always-loaded vs path-scoped vs invoked-only, ranks the top contributors, and flags entries over budget. Add `--api` for exact counts via Anthropic's count_tokens endpoint instead of the chars/4 heuristic.

Find and fix a bug. Default is careful (reproduce, investigate, test). Add `--fast` for emergency production mode (hotfix branch, minimal change, ship a PR fast).

Investigate a non-trivial task before planning. Produces a findings report with file:line citations covering current behavior, upstream/reference equivalents, framework bridges, helpers to reuse, and open questions. Use before porting an upstream change, migrating a screen to Compose, touching unfamiliar modules, or any cross-cutting change. The point is to ground the plan in evidence, not memory.

Trim verbose prose, walls of text, journey narration, and WHAT comments from docs and code without losing vital info. Targets markdown files (docs, CLAUDE.md, rules, skills, agents) and, when asked, code comments / KDoc. Applies the "WHY not WHAT" rule, the "describe current behavior not the journey" rule, and the "release notes not implementation rationale" rule from `.claude/rules/workflow.md` and `.claude/rules/code-quality.md`. Use after a feature lands and the docs around it have accumulated cruft, or when a file feels heavier than it earns. Always plans before editing.

Find and fix a bug. Default is careful (reproduce, investigate, test). Add `--fast` for emergency production mode (hotfix branch, minimal change, ship a PR fast).

Audit a Reikai feature ported onto Mihon for behavioral parity against the Yōkai-era source on the `design/library-compose` branch. Use when the user asks to audit, verify, or sanity-check a ported feature, when they want to know whether a port "really matches" the original, or when they want a defect-fix pass on a recently-ported area. Takes a scope argument like "novel reader" or "Settings → Security". Compares features and behavior only, NOT pixels or layout (the user often redesigns visuals on the way over).

Scan changes, commit, push, and create a PR. With confirmation at each step

Deep, fan-out research over the codebase to answer a big question (test gaps, how a subsystem works end to end, a risk/security/dead-code audit, "where does X happen everywhere"). The codebase analog of deep web research: parallel explorer agents gather file:line-cited findings, surprising or high-stakes claims are adversarially verified against current code, and the result is a single prioritized report. Use for broad/deep questions that span many files and modules, NOT a one-file lookup and NOT a single-task pre-plan (use /scout for that, /explain for one thing).

Estimate per-turn token cost of this project's `.claude/` configuration and `CLAUDE.md`. Reports always-loaded vs path-scoped vs invoked-only, ranks the top contributors, and flags entries over budget. Add `--api` for exact counts via Anthropic's count_tokens endpoint instead of the chars/4 heuristic.

Review code changes or a pull request. Delegates to specialist agents for code quality, security, performance, and documentation, in parallel.

Write Solverr's Handoff.md so a fresh session resumes cleanly, and bring the CHANGELOG, dependent docs, and memory store in line with it. Use when the session is wrapping up ("I'm stepping away", "continue tomorrow", "wrap up"), before a `/clear`, when a long session has accumulated state that would be lost, or when the session is looping on a broken approach and a fresh agent would do better. Offer it proactively on those signals; do not wait to be asked.

Estimate per-turn token cost of this project's `.claude/` configuration and `CLAUDE.md`. Reports always-loaded vs path-scoped vs invoked-only, ranks the top contributors, and flags entries over budget. Add `--api` for exact counts via Anthropic's count_tokens endpoint instead of the chars/4 heuristic.

Verify a change against live Cloudflare challenges through a running Solverr, because the unit tests cannot reach the thing that actually breaks. Builds an isolated container, runs the matrix (interstitial on both engines, Turnstile, PDF, passthrough, cookie shape), and proves each mechanism with a controlled A/B rather than a single pass. Use after any change to the engines, detection, sessions, or the passthrough, and before cutting a release. Never touches the owner's own solverr container.

Cut a Solverr release end to end. Decides the SemVer bump from what is in [Unreleased], runs the preflight, rewrites the CHANGELOG section, bumps package.json, commits, pushes, tags, then verifies both workflows and the published image digests. Use when the owner asks to cut, ship, or publish a version. Pushing the tag publishes the GitHub Release and the ghcr image, so it always confirms before that step.

Investigate a non-trivial task, then produce the plan for it. Emits a file:line-cited findings report covering current behavior, the upstream equivalent in FlareSolverr or Byparr, the engine and event-loop constraints that bind the approach, helpers to reuse, stale docs, then a sequenced plan and any blocking open questions. Use before porting an upstream change, touching the engines or session lifecycle, or any change that crosses the controller. The point is to ground the plan in evidence, not memory. Investigates deeply and verifies adversarially by default; never edits files.

Trim verbose prose, walls of text, journey narration, and WHAT comments from docs and code without losing vital info. Targets markdown files (docs, CLAUDE.md, rules, skills, agents) and, when asked, code comments / docstrings. Applies the "WHY not WHAT" rule, the "describe current behavior not the journey" rule, and the "release notes not implementation rationale" rule from `.claude/rules/workflow.md` and `.claude/rules/code-quality.md`. Use after a feature lands and the docs around it have accumulated cruft, or when a file feels heavier than it earns. Always plans before editing.

Audit Solverr against its two upstreams for drift. Compares the stealth engine against Byparr and the Chrome engine, sessions and the /v1 surface against FlareSolverr, classifying every difference as already covered, genuinely missing, or deliberately divergent, and checking that the /v1 request and response shape is still byte-compatible. Use when either upstream has moved, before a release, when a fix lands upstream that might apply here, or when the owner asks whether Solverr has fallen behind. Ends by updating the sync ledger.