Agents
Accelerator skills delegate exploratory work to subagents defined in
agents/*.md. Each agent runs in an isolated context, does its reading or
searching there, and returns only a summary — keeping the main
conversation’s context small. See Philosophy for why.
The locator/analyser split
Section titled “The locator/analyser split”Most agents come in pairs across three domains — codebase, documents, and browser. Locators answer where: they enumerate files, documents, routes, or screens, and deliberately have no Read tool (or, for the browser, take no screenshots), so they cannot be tempted into deep analysis and their context stays bounded. Analysers answer how: they read (or inspect) a focused set of targets in depth and return distilled findings.
| Domain | Locator (WHERE) | Analyser (HOW) |
|---|---|---|
| Codebase | codebase-locator | codebase-analyser |
| Documents | documents-locator | documents-analyser |
| Browser | browser-locator | browser-analyser |
Orchestrating skills typically fan out locators first, then spawn analysers on the narrowed target list.
Codebase agents
Section titled “Codebase agents”codebase-locator
Section titled “codebase-locator”Finds WHERE code lives. Searches for files, directories, and components relevant to a feature or task and organises the results by purpose — without analysing their contents.
Tools: Grep, Glob, LS (no Read)
codebase-analyser
Section titled “codebase-analyser”Explains HOW code works. Reads specific files, traces data flow and
method calls, and reports implementation details with precise
file:line references.
Tools: Read, Grep, Glob, LS
codebase-pattern-finder
Section titled “codebase-pattern-finder”A hybrid: locates similar implementations, usage examples, and established patterns, and — unlike the locator — returns concrete code extracts that can serve as templates for new work.
Tools: Grep, Glob, Read, LS
Documents agents
Section titled “Documents agents”documents-locator
Section titled “documents-locator”Discovers relevant documents in the configured document directories
(meta/ by default) and categorises them, without deep content
analysis. Preloads the paths skill so it searches the directories the
project actually configures.
Tools: Grep, Glob, LS (no Read)
documents-analyser
Section titled “documents-analyser”Extracts high-value insights from specific documents: decisions, conclusions, actionable recommendations, constraints. Filters aggressively so only relevant material returns to the caller.
Tools: Read, Grep, Glob, LS
Browser agents
Section titled “Browser agents”Both browser agents drive a running web application through the
Playwright executor (run.sh), whose path is injected via the preloaded
browser-executor skill. They are used by the
design convergence workflow.
browser-locator
Section titled “browser-locator”Enumerates WHERE things appear in the rendered UI: routes, screens, and DOM-level component presence, using navigation and accessibility-tree snapshots — no content analysis, state extraction, or screenshots.
Tools: Bash (Playwright executor only)
browser-analyser
Section titled “browser-analyser”Captures HOW a focused set of screens behaves: state variants (loading, empty, error, success), interactions, screenshots, and computed style and layout values.
Tools: Bash (Playwright executor only)
Other agents
Section titled “Other agents”reviewer
Section titled “reviewer”Generic review agent behind the multi-lens
Review System. Review orchestrators
(review-pr, review-plan, review-work-item) spawn one instance per
lens, injecting the lens skill and output-format specification at spawn
time; the agent reads both, explores the artefact, and returns a
structured JSON review.
Tools: Read, Grep, Glob, LS
web-search-researcher
Section titled “web-search-researcher”Researches questions on the public web: analyses the query, searches, fetches promising sources, and returns findings with citations. Used when a task needs current information beyond the codebase.
Tools: WebSearch, WebFetch, TodoWrite, Read, Grep, Glob, LS