browzer

Code review specialist for browzer-indexed repos. Operates as one of four mandatory review lanes (senior-engineer, software-architect, qa, regression-tester) or as a domain specialist discovered via find-skills. Always read-only — cannot modify files. Receives REVIEWER BRIEF (diff + browzer deps + changed symbols inlined verbatim) and writes one CODE_REVIEW.<lane>.md per dispatch.

Implementation specialist for browzer-indexed repos. Executes a single closed-prompt TASK_NN.md dispatched by execute-task — loads each `task.skillsFound[]` skill via Skill(...), edits files in `task.scope.files[]`, verifies `task.invariants[]`, and returns a structured `## Subagent report` block. Never reads PRD.md or EXPLORATION.md — TASK_NN.md is a closed prompt by contract.

Documentation patching specialist for browzer-indexed repos. Patches host markdown docs whose accuracy depends on just-changed code. Dispatched by update-docs Phase B. Reads DOC_PATCHES.md discovery receipts and applies targeted inline edits without rewriting docs wholesale.

RAG discovery specialist for browzer-indexed repos. Maps files, dependencies, symbol mentions, and domain context via browzer explore / search / deps / mentions. Dispatched by scope-feature, code-review (REVIEW_CONTEXT pre-render), receiving-code-review (rare). Always read-only — never modifies files. Returns structured JSON receipts at /tmp/<phase>-<featId>-<noun>.json paths.

Post-review fix specialist for browzer-indexed repos. Consumes per-finding FIX BRIEF from receiving-code-review, applies fixes through the 7-step model-escalation ladder (sonnet → retry → research → opus → retry → research → tech-debt), and writes staging/fixes/F-NNN.completed.md (success) or staging/fixes/F-NNN.tech_debt.md (exhausted). Haiku is forbidden.

Pre-flight probe for any Browzer agent loop — confirms login, workspace binding, gateway/organization via `browzer status`. Use first in any session before explore/search/sync; for install/login itself use `use-rag-cli`. Triggers: browzer status, browzer pre-flight, browzer auth check, browzer context, browzer health check, browzer diagnose, BROWZER_SERVER check, 'am I logged in to browzer', 'which browzer workspace', 'what server is browzer pointing at'.

Interactive clarification before generate-prd — when the operator's request lacks persona, success signal, or scope. Asks one grounded question at a time, informed by browzer explore/search on the host repo, optionally dispatches parallel research agents (WebFetch, WebSearch, Firecrawl, Context7) for unknowns, proposes search-trigger expansions for uncovered domain concepts, hands off to generate-prd by writing BRIEF.md. Use proactively whenever a request names a capability but omits who benefits, what success looks like, or what's out of scope. Triggers: brainstorm, help me think about, walk me through an idea, spec this with me, sanity check an idea, rough idea, sketch this out, I want to add, what if we, how could we.

One-shot Browzer onboarding for an existing repo. Reads every markdown doc (README, CLAUDE.md / AGENTS.md / CONTRIBUTING.md, docs/**, runbooks, ADRs), reconciles each claim against real code via `browzer explore`, DELETES stale/duplicate/incorrect content outright, shows the diff for approval, then runs `browzer init` + `browzer sync` so the index reflects the cleaned docs. Idempotent. Triggers: bootstrap rag on this repo, browzer bootstrap, onboard this project into browzer, sync the docs against the code, clean up the docs and rag this repo, rag steroids, claude rag onboarding, full rag onboarding, one-shot rag setup.

Post-implementation team review of a feature's diff. Spawns 4 mandatory parallel reviewer lanes (senior-engineer, software-architect, qa, regression-tester) plus optional domain specialists discovered via /find-skills. Each lane writes CODE_REVIEW.<lane>.md; the dispatcher aggregates into CODE_REVIEW.md frontmatter via scripts/aggregate-findings.mjs. Read-only — receiving-code-review applies fixes next. Triggers: code review, review this feature, audit my changes, review the diff, post-implementation review, team review, peer review, find issues in this PR.

Write a Conventional Commits v1.0.0 message mirroring the host's last 5 commits, stamp the on-behalf-of: @<org> trailer (per GitHub's organization-commit convention — orgs link via on-behalf-of, NOT Co-authored-by which is for human collaborators) when the host has configured an org, and run git commit. Hard veto when invoked with a feature id: ACCEPTANCE.md.verdict must be accepted AND README.md must exist AND no TASK_*.failed.md present. Reports the SHA. Does NOT push.

20 hooks across 6 events