recall-scout

name	recall-scout
description	Scout an unfamiliar codebase — memory-first repo map, key paths, tests, risks, and next steps, with a strict sensitive-data boundary
allowed-tools	Bash(git log -10 --oneline), Bash(git status -s | head), Bash(codegraph status --json)

Produce a scout report for the current repository: a fast, structured orientation for whoever is about to work in this codebase. An optional focus argument narrows the scout to a subsystem, path, or question (e.g. auth, the extraction pipeline, src/db). With no focus, scout the repo as a whole.

This is the canonical scouting workflow; platform guides reference it and only remap tool names. Do not hand-copy these steps into FOR_CLAUDE.md, FOR_PI.md, FOR_OPENCODE.md, or opencode/recall-memory.md — those files point here and supply only their own tool-name mapping.

Live context

Three bounded orienters, each well under 1 KB of output. On Claude Code the ! prefix inlines their output here at prompt load; on hosts without the ! mechanic, run the same three commands manually as your first step.

• Recent commits: !git log -10 --oneline
• Working tree: !git status -s | head
• CodeGraph probe: !codegraph status --json

Workflow

1. Memory first. Before reading git history or source files, search Recall for what is already known about this repo and the focus. Use keyword search for names and hybrid search for natural-language questions, scoped to this project only. Recall's structured decisions, learnings, and session summaries are richer than commit messages — start here, not with git log.
2. CodeGraph capability ladder. Read the probe (codegraph status --json — it exits 0 even on an unindexed repo) and branch on initialized:
   • Indexed (initialized: true): check freshness — if pendingChanges is non-zero or reindexRecommended is true, run codegraph sync -q (cheap, safe) before querying.
   • CLI installed, repo unindexed (initialized: false): offer indexing, don't run it — "this repo isn't indexed; codegraph init takes ~2 s per 200 files, writes only a self-gitignored .codegraph/, and codegraph uninit reverses it — want it?" Run codegraph init only on an explicit user yes; indexing is the user's decision, never auto-run it. Without a yes, continue on the backstop.
   • No CLI (the probe errors with command-not-found): skip CodeGraph entirely and produce the full report via the grep/tree-walk backstop.
   • CodeGraph is an enhancement, never a hard dependency: every report section below must be producible with zero CodeGraph. (Hosts with the CodeGraph MCP server may use its single tool, codegraph_explore, wherever codegraph explore appears — identical output; everything else on this page is CLI-only.)
3. Repo map. Build a high-level map: languages, entry points, top-level layout, build/test commands. With a live index, start from the orientation bundle — codegraph status --json (languages, symbol-kind profile, file/node counts; already inlined above) plus codegraph files --max-depth 2 --no-metadata (the repo tree; add --filter <dir> when the focus names an area). The bundle costs ~600 tokens and replaces a blind tree walk. Then read the declared ground truth directly — these files are small and never in the index: README, AGENTS.md/CLAUDE.md, the dependency manifest (package.json including its scripts, or pyproject.toml / Cargo.toml / go.mod), the compiler/tooling config (tsconfig.json or equivalent), and CI workflow names. Without CodeGraph, the same declared reads plus a bounded tree walk produce the map.
4. Tech stack. From manifests only: languages, runtime, frameworks and major libraries with their declared versions, build tooling, test framework, and package manager — read from the manifest's declared ranges (package.json, pyproject.toml, Cargo.toml, go.mod). Lockfiles stay behind the sensitive-data boundary: map their existence, never their contents, so version ground truth is the declared range, not the resolved lockfile pin. CodeGraph contributes the languages list and symbol-kind profile from status --json; it does not know dependencies or versions.
5. Key paths. Identify the handful of files that matter most for the focus — entry points, core modules, config, and the seams where changes usually land. Cite them as path:line where useful. With a live index, surface them with at most two narrow codegraph explore calls (see the query discipline below) rather than guessing from the tree: one explore answers "what calls this seam", "what does it depend on", and "trace a path between two symbols" in a single query. Fall back to grep when CodeGraph isn't available.
6. Decisions & learnings tied to the key files. For each key path, search Recall (keyword + hybrid, scoped to this project) for the decisions and learnings that reference the file or the symbols it defines — why a hot file is the way it is. This focuses the memory-first search from step 1 on each key path.
7. Conventions. How the house writes code, from three legs: declared — the rules files already read in step 3 (AGENTS.md/CLAUDE.md/CONTRIBUTING.md) plus lint/format configs (eslint/prettier/.editorconfig, tsconfig strictness); observed — one CodeGraph style probe over a representative module (codegraph explore of that module, or cheaper, codegraph node --file <path> --symbols-only) for naming patterns, module seams, and test-file conventions — an explore used here counts against the two-explore cap; learned — Recall learnings, which often encode "wrong → right" convention corrections.
8. Tests. Locate the test suite, the runner, and how to invoke it. Note coverage gaps relevant to the focus.
9. Risks. Surface what would bite someone working here — fragile areas, undocumented invariants, conventions that are easy to violate (pull these from memory on the hot files and from project rules files).
10. Next steps. Recommend the concrete first moves for the focus, ordered, each tied to a path or command.

CodeGraph query discipline

CodeGraph replaces the expensive grep-and-read loop, but explore pastes verbatim source and has no output cap — discipline keeps the scout token-frugal:

• Narrow, named queries only. Point explore at a symbol, a file, or a seam — never a vague topic. Off-target matches cost the same tokens as on-target ones.
• Soft cap: at most two explore calls per scout run, including the Conventions style probe.
• Never re-paste explore output into the report. The report wants a one-line "why it matters" per key path and path:line cites — not quoted source blocks.
• Prefer cappable forms for edge questions: codegraph callers <symbol> --limit 5, codegraph callees <symbol> --limit 5, codegraph query <term> --limit 3 --kind <kind>, codegraph node --file <path> --symbols-only.
• Backstop means fallback, not second pass: if CodeGraph answered a question, skip grep and the tree walk for that same question.
• Symbol misses exit 0. query/node/callers print "not found" and still exit 0 — check the output text for misses, don't exit-check.

Report format

Output the scout report in chat with these sections, in order:

• Memory summary — what Recall already knew (decisions, learnings, breadcrumbs, prior sessions). Say so explicitly if memory was empty.
• Repo map — languages, layout, build/test commands.
• Tech stack — languages/runtime, frameworks and major libraries with declared versions, build/test tooling, package manager (from manifests; lockfiles excluded).
• Key paths — the files that matter, with one line each on why (grounded in the code graph where available).
• Decisions & learnings — per key file, the memory tied to its code (from Recall search on the file and its symbols). Omit the section if nothing was found.
• Conventions — the house style: declared rules, observed patterns, and Recall-learned corrections.
• Tests — suite location, runner, how to run, notable gaps.
• Risks — what would bite someone working here.
• Next steps — ordered, concrete, path- or command-anchored.

Sensitive-data boundary (MANDATORY)

Never read, echo, summarize, or write any of the following — if the focus would require crossing this boundary, stop and say so instead:

• Secrets / credentials — .env*, *.key, *.pem, tokens, API keys, anything matching a credential pattern. Do not print their values even if asked to "summarize config."
• Generated / vendored files — node_modules/, dist/, build/, .git/ internals, lockfile contents. Map their existence, never their contents.
• The live database — never open, query, or dump ~/.agents/Recall/recall.db (or any production DB) directly. Use the Recall CLI and search tools, which read the database for you, scoped to this project — never raw SQL against the file.
• Cross-project memory — scope every Recall search to this repository's project. Do not surface, quote, or leak another project's memory into this report.

Artifacts (opt-in)

The scout report is chat-only by default — write nothing to disk. Persist an artifact only when the repo endorses it: a .agents/atlas/artifacts/ directory already exists, or the user explicitly asks for a saved report. When endorsed, write only to .agents/atlas/artifacts/ as YYYY-MM-DD-scout-<focus>.md. Never write scout artifacts to .agents/atlas/handoffs/ (reserved for handoffs) or anywhere else.