Explore Codebase
You are building a mental map of an unfamiliar or large codebase region. The goal is structural comprehension — understanding what the code does, how it's organized, and what the key entry points and invariants are — without loading everything into your main context window.
This skill exists because codebases are too large to read in one prompt. The pattern (Recursive Language Model — see the RLM substrate section in CLAUDE.md) is: keep your root context small, dispatch sub-agents to read individual files or modules, and have each sub-agent return a structured summary. Synthesize the summaries into a coherent map.
This skill covers
any codebase the agent encounters — forks, dependencies, unfamiliar regions, user projects. For analyzing yoyo's own source to find bugs and gaps, use
instead.
When to use
Trigger this skill when ANY of these hold:
- A planned task touches >5 files you haven't recently worked on
- A community issue references a feature or module you don't have a mental map for
- You're investigating a bug whose surface spans multiple modules you haven't read recently
- A new dependency is being introduced and you want to know its public API + key invariants before integrating
- A user explicitly asks you to explore, understand, or map a codebase
- brought in a large project and you need structural context before acting
- You're working in a fork and need to understand what the fork changed relative to upstream
When NOT to use
- Small known regions. A single file ≤300 lines, or a function you wrote yesterday — just read it directly. Sub-agent overhead exceeds the savings.
- Precise edits across files. Refactoring needs mutual context (seeing all pieces at once), not summaries. Summaries lose the fidelity you need for surgical edits.
- Sequential workflows with strong mutual context. When each step depends on the full output of the previous step, fan-out doesn't help — you need serial reading.
- The region is already known. If you or the user can name the exact module and its API from memory, direct read is faster. Don't explore what you already understand.
- You're inside a sub-agent at depth 3. Stop. Return what you have. Do not dispatch further.
Procedure
1. Identify the region
Define the exploration scope — be specific:
- A directory:
- A file glob:
src/agent_builder.rs src/main.rs src/tools.rs
- A dependency:
~/.cargo/registry/src/*/yoagent-*/src/
- A commit range:
git diff main..feature-branch --name-only
If the scope is vague ("understand this project"), start with orientation (Step 2). If the scope is precise (a named set of files), skip to Step 3.
2. Orient — build a rough map
Before dispatching sub-agents, gather cheap structural signals directly:
bash
# Project structure
find <root> -type f -name '*.rs' | head -50
# or use /map if available in the REPL
# README / docs
cat <root>/README.md 2>/dev/null | head -100
# File sizes (to plan dispatch)
wc -l <root>/src/*.rs | sort -rn | head -20
# Recent activity
git log --oneline -20 -- <root>/src/
From this, build a file inventory with rough sizes. Files >300 lines are candidates for sub-agent exploration. Files ≤300 lines can be read directly if needed later.
3. Decide: direct read or sub-agent?
For each file in the region:
- ≤5KB (roughly ≤150 lines): Read directly with or . No sub-agent needed.
- >5KB: Dispatch a sub-agent (Step 4). Don't load large files into your main context.
If the total region is small enough to read directly (≤5 files, all ≤5KB), skip sub-agents entirely — just read and synthesize in your main context.
4. Dispatch per-file sub-agents
Store each file's content in shared state, then dispatch a sub-agent to summarize it. One file per sub-agent — files are the natural unit of structure in code.
4a. Store the artifact
shared_state set key="explore.<region>.<filename>" value="<file contents>"
Namespace convention:
explore.<region>.<filename>
(e.g.,
,
).
If a single file exceeds 30KB (~120,000 bytes), chunk it before storing:
- Split into chunks of ~80KB with 8KB overlap between consecutive chunks
- Store as
explore.<region>.<filename>.chunk-1
- Dispatch one sub-agent per chunk (same as analyze-trajectory's Section 3.5)
4b. Dispatch the sub-agent
sub_agent: You are exploring a source file to build a structural summary.
The file is stored in shared state under key "explore.<region>.<filename>".
Read it with: shared_state get key="explore.<region>.<filename>"
Describe this file's structure in a JSON response. Reply with ONLY a JSON object (no markdown fences, no prose):
{
"file": "<filename>",
"purpose": "1 sentence: what this file does",
"public_api": ["list of exported functions/structs/traits with 1-line descriptions"],
"key_invariants": ["non-obvious behaviors, constraints, or assumptions"],
"dependencies": ["other modules/crates this file depends on"],
"dependents": ["who calls into this file, if visible from imports/use statements"],
"complexity": "low|medium|high",
"deeper_question": "a follow-up question if something is unclear, or null"
}
Skills do not chain. Sub-agents don't load this skill or any other; include the full question and shared-state key reference directly in the sub-agent's prompt.
4c. Handle sub-agent responses
Parse each sub-agent's response as JSON:
- Valid JSON with all fields: Store the summary in shared state under
explore.<region>.<filename>.summary
- Malformed JSON but readable text: Extract what you can. Construct a partial summary:
{"file": "<filename>", "purpose": "<first 200 chars of response>", "public_api": [], "key_invariants": [], "dependencies": [], "dependents": [], "complexity": "unknown", "deeper_question": null}
- Empty or errored: Fall back to direct read of the file's first and last 50 lines. Produce a low-confidence summary manually.
5. Recurse on deeper questions
If a sub-agent returns a non-null
and
is
:
- Dispatch another sub-agent with the narrower question, referencing the same shared-state key.
- Merge the answer into the existing summary.
Hard cap: recursion depth = 3. That's: initial dispatch → 1st recursion → 2nd recursion. After depth 3, accept whatever you have. If you find yourself wanting depth 4, your initial scope was probably too broad — narrow the region and retry.
6. Synthesize into a mental map
After all per-file summaries are collected, dispatch a synthesis sub-agent (or do this in your main context if the total summary data is small enough, ≤5KB):
sub_agent: You are synthesizing per-file summaries into a structural map of a codebase region.
The following shared-state keys contain per-file summaries:
- explore.<region>.<file1>.summary
- explore.<region>.<file2>.summary
...
Read each summary, then produce a structural map as a JSON object:
{
"region": "<region name>",
"overview": "2-3 sentences: what this region does as a whole",
"module_graph": ["<file-A> -> <file-B>: <relationship>", ...],
"entry_points": ["the key functions/structs a caller would use"],
"invariants": ["cross-file constraints or assumptions"],
"risk_areas": ["files or interactions that look fragile or complex"],
"open_questions": ["things the summaries couldn't resolve"]
}
7. Use the map
The mental map is your working context for the rest of the session. Reference it when:
- Planning which files to modify for a task
- Estimating the blast radius of a change
- Deciding whether a refactor is safe
- Explaining code structure to a user or in a journal entry
Store the final map in shared state under
so sub-agents in later steps can reference it without re-exploring.
Pitfalls
- Don't explore what you already know. If you wrote the code recently or have it in active memory, skip this skill. It's for building new understanding, not confirming existing knowledge.
- Don't ask sub-agents to make decisions. They summarize structure; you decide what to do with it. Sub-agents that plan or recommend tend to drift.
- Don't dump multiple files to one sub-agent. One file per dispatch keeps the JSON output reliable and the summary focused. The synthesis step is where cross-file reasoning happens.
- Don't forget the recursion cap. 3 is the hard limit. If your region needs depth 4, the region is too broad — split it.
- Don't explore before acting on small tasks. If the task is "fix this one function," reading that function directly is faster than exploring the whole module. Match the tool to the task size.
- Don't re-explore within the same session. If you've already explored a region, the summaries are in shared state. Read them with instead of re-dispatching sub-agents.
- Per-file, not per-byte. Unlike analyze-trajectory (which chunks CI logs by byte offset), this skill fans out by file. Files are the natural structural unit in codebases. Only chunk within a file if it exceeds 30KB.
Verification
An exploration is "good enough" when ALL of:
- The map names concrete files and functions (not "some module that handles X")
- Each file in the region has a summary (even if low-confidence for some)
- The module graph shows how files relate (who calls whom, who depends on whom)
- Entry points are identified — a caller knows where to start
- The total exploration used ≤ N+2 sub-agent dispatches where N is the number of files explored (N per-file + 1 synthesis + 1 possible recursion)
- The work stayed within the depth-3 recursion cap
If the map fails any of these, narrow the region and re-explore the gap, or accept the partial result and document open questions.
What this skill deliberately does NOT do
- Does not modify code. Exploration produces understanding, not changes. The actual edits are a separate task.
- Does not find bugs. That's . This skill builds the map; self-assess uses the map to find problems.
- Does not auto-create documentation. If the map is worth preserving as docs, that's a separate decision outside this skill's scope.
- Does not write to the audit-log branch. The exploration results live in shared state for the current session only.