Map Codebase

Map Codebase (Brownfield Arrow Bootstrap)

This skill maps an existing codebase into the arrow of intent. It works bottom-up: read all the code in scope first, then propose lens-based clusterings for the user to choose among, then generate skeleton docs that describe what actually exists.

See brownfield-bootstrap.md for detailed guidance per phase.

Five Critical Rules

These govern every phase. Apply consistently.

1. Read actual code, don't guess. Every claim in generated artifacts traces to file/line evidence. Speculation is flagged explicitly rather than presented as fact.
2. Each STOP is mandatory. The workflow has multiple stop points. None are optional. Rushing past a stop is how brownfield mapping produces bad LLDs that poison subsequent work.
3. LLDs describe current reality, not aspirational design. Output is what the code is, not what a greenfield version would be. Inferred design decisions carry [inferred] markers; known technical debt and behavioral quirks go in Open Questions.
4. Thoroughness over speed. Token budget is real but not dominant; skimming produces mappings that miss behaviors and lock in the wrong segmentation.
5. Humble but guide. The agent is not the expert on the user's system; the user is. But don't silently defer — when the user's framing conflicts with the evidence, surface the tension with evidence rather than just going along.

At invocation

Ask one question first:

• Whole project, or specific parts?
  • Whole project → implies Full LID mode. Scope is the entire project.
  • Specific parts → implies Scoped LID mode. Ask the user to name the parts (directories, file lists, or component names). The declared parts are both the sweep scope and the LID scope going forward.

This question determines scope and mode simultaneously — the user is not asked a separate "Full or Scoped?" later at terminal verification. Default to Full (whole project) if the user is undecided.

Then ask:

• Subagent parallelism — offer as an option. Recommended for large codebases; single-agent works for smaller ones.

Token-intensity warning. Tell the user upfront this is token-intensive by design — reading every file, proposing multiple lenses, drafting skeletons for every segment, multi-step reconciliation. Not a lightweight operation. Users expecting a quick one-shot map should reconsider.

Undo. The workflow's STOPs between phases are the undo mechanism — aborting at any STOP leaves nothing written to disk. Agent harnesses also provide their own session-level rewind. LID does not ship a dedicated /unmap-codebase command; users roll back via the agent framework's rewind or by reverting a git commit.

State dispatch

Inspect the project before starting:

• Partial LID docs exist (HLD or some LLDs, but not complete). Ask the user: treat existing docs as authoritative (draft skeletons only for uncovered segments) or supersede them? Do not silently overwrite.
• Full LID docs exist but no docs/arrows/. Redirect the user to /arrow-maintenance — that command bootstraps the overlay from existing docs without the brownfield sweep. Do not proceed here.
• No LID docs, no overlay. Standard brownfield flow (below).

Phase 1 — Sweep (Reconnaissance)

Read every file in the declared scope. Not a sample. Sampling risks missing behaviors that only surface in edge-case files and locks in segmentation based on incomplete view.

For each file, record a structured summary:

• Purpose — what this file appears to do.
• Exports — functions, classes, types, endpoints exposed to other parts of the system.
• Dependencies — what the file imports or calls.
• Data shapes — structures it produces or consumes.
• Side effects — filesystem, network, database, logs.
• Role — how this file fits into the larger system (UI component, API handler, background job, pure utility, etc.).
• Observations — anything unusual, deprecated-looking, or flagged by comments.

Output: a flat list of observed behaviors with file/line references. No segmentation attempted here.

Capacity constraint handling. If the declared scope exceeds the invocation's capacity (single-agent context window, or the chosen subagent budget), surface the constraint with concrete sizing evidence, warn the user that a sampled sweep produces lower-quality mapping, and recommend narrowing scope or enabling subagent parallelism. The user may override and proceed with sampling anyway. Under override, preserve state across truncation points via per-subagent files (.lid/map-codebase/sweep-{N}.md) or by incrementally writing arrow-doc partial drafts during reconnaissance — never silently discard information the orchestrator cannot hold.

When subagents ran in parallel, each subagent writes its sweep to its own file; the orchestrator processes them in chunks during Phase 2.

See subagent-sweep-prompt.md for the prompt template given to sweep workers.

Phase 2 — Seam Identification: Lens Selection

Propose 3–5 fundamentally different clusterings, each using a distinct lens. Not variations on one theme — entirely different mental models.

Good lenses to propose:

• Data flow — what data originates where, how it moves between modules.
• User-facing capability — clusters organized around things a user can do (sign in, check out, export data).
• Domain concept — clusters matching domain language (order, inventory, keeper, entry).
• Behavioral boundary — where the system changes state in coordinated ways (authentication flow, payment pipeline).
• Creative / unconventional — a lens not already tried, presented as a counterweight.

Anti-pattern lenses to explicitly avoid:

• Frontend vs. backend split (deployment-location, not intent).
• Files that deploy together (infrastructure grouping, not intent).
• Team ownership (org chart, not intent).
• Utils / shared / common directory (tooling leftover, not a real concept).

For each proposed clustering, present: name, lens, the clusters it produces, pros, cons, and best-for (what kind of reasoning it supports well).

STOP. User picks a lens. Multiple lenses are the primary edge-detection mechanism — the user's choice of lens reveals latent intent in a way no single clustering can.

See reconciliation-template.md for the presentation format.

Phase 3 — Seam Identification: Slicing Granularity

Within the chosen lens, propose 2–3 slicing variations:

• Coarse — 3–4 large segments. Fewer LLDs to maintain, less precise tracking.
• Medium — 6–8 segments. Balanced.
• Fine — 10+ finer-grained segments. More precise tracking at the cost of more docs.

Coarse absorbs more code per LLD; fine gives precise segment-scoped tracking. Pick based on project maturity and the user's appetite for maintenance.

When a fine slicing produces more segments than sit comfortably at one level, propose grouping related leaves under sub-HLD (grouping) nodes — a shallow tree rather than a long flat list. Most projects map flat (depth-2: leaves directly under the root); nesting is offered only when the leaf count makes a grouping level genuinely clarifying. The tree placement chosen here drives the parent/children links and the mirrored artifact paths in Phase 5.

STOP. User picks a slicing.