Exploration Strategy

Status: v0.1 baseline draft — derived from the case catalog and axis definitions; awaiting trace evidence from Fable 5.1 case runs.

When to apply

• You are dropped into a repository you have never seen and asked to change or fix something in it.
• A task description references symbols, modules, or behaviors you cannot yet locate in the code.
• You are about to edit code whose callers, config sources, or side effects you have not confirmed.
• A bug report describes a symptom but you do not yet know which layer produces it.
• Your context was compacted or the session resumed, and your prior map of the codebase may be stale.
• You catch yourself editing based on what a file "probably" contains rather than what you read.

Core loop

1. Restate the task as a location question. Before any file read, write one sentence: "To do X, I need to find where Y happens and what touches it." This defines the exploration target; every subsequent read is judged against it. If you cannot phrase the target, the task itself is underspecified — resolve that first.

2. Breadth-first structural scan (hard cap: ~10 tool calls). List the top-level directory tree. Read the manifest (package.json / Cargo.toml / go.mod / pyproject.toml / Makefile) and the README headings only. Output for yourself: language, build system, test command, and 3-6 candidate directories relevant to the task. Do not open individual source files yet.

3. Locate entry points. Find where execution starts for the behavior in question: main, HTTP route tables, CLI arg parsers, event handlers, exported public API, scheduled jobs. Grep for the task's key nouns first — error strings, endpoint paths, flag names, config keys. Literal strings lifted from the task description are the highest-signal search keys available: near-zero false positives.

4. Trace ONE representative path end-to-end. Pick the single most task-relevant flow and follow it from entry point to observable effect (response written, file persisted, state mutated), reading only files on that path. Depth-first on one path beats shallow reads of ten files: the traced path reveals the layering conventions, error-handling style, and data shapes that generalize to the rest of the codebase.

5. Maintain a known/assumed ledger. As you trace, keep two explicit lists:

   • KNOWN — facts confirmed by reading code or running commands, each with a file:line or command receipt.
   • ASSUMED — inferences from names, docs, comments, or convention. Every ASSUMED item is a liability. Promote it to KNOWN by reading the code, or carry it into the plan as a named risk. Never let an assumption become load-bearing silently.

6. Probe dynamically if cheap. If the project runs or tests in one command, run it once before editing. One observed runtime behavior — a log line, a passing suite, a reproduced error — outweighs ten inferred ones, and it validates the build/verify loop before you depend on it.

7. Apply the sufficiency test. Stop exploring when you can answer all four:

   • (a) Where exactly will my edit go?
   • (b) What calls into it, and what does it call?
   • (c) How will I verify the change worked?
   • (d) What is the most likely way my change breaks something else? Four yes answers → begin editing immediately. Any no → that question names the next thing to read; read only that.

8. Declare the model, then act. Write a 3-6 line summary: architecture in one line, the traced path, the edit site, the verification plan, and the surviving ASSUMED items. This is your checkpoint — if an assumption later proves wrong, you know exactly which belief to revise instead of restarting exploration.

9. Re-explore only on surprise. During implementation, return to exploration only when reality contradicts the ledger (a KNOWN item was wrong, or an ASSUMED item turned out to matter). Fix that one entry, re-trace the minimal affected path, resume. Do not rescan the repository.

Heuristics

• Read the manifest and directory tree before any source file; the build system reveals the architecture faster than the code does.
• Grep for literal strings from the task (error messages, route paths, config keys) before grepping for concept words.
• One end-to-end trace of a representative path teaches more than skimming N sibling files; siblings usually mirror the traced one's structure.
• Time-box the structural scan: if step 2 hits ~10 tool calls without candidate directories, switch to keyword grep instead of browsing further.
• Exploration budget scales with blast radius, not repo size: a one-line fix in a 1M-line repo needs only the traced path around that line.
• If two consecutive reads produce no new KNOWN entries, exploration has saturated — run the sufficiency test now.
• Tests are compressed documentation: a module's test file states intended behavior and canonical call patterns; read it before the implementation when both exist.
• Never trust a function by its name alone; if your change depends on its behavior, read its body — name/behavior mismatch is a top source of wrong edits.
• Prefer one cheap runtime probe (run tests, curl the endpoint, execute the script) over three more file reads when the project boots in under a minute.
• After context compaction or session resume, re-verify the top three KNOWN items you are about to build on; stale certainty is worse than admitted ignorance.

Anti-patterns

• Edit-first archaeology. Making the "obvious" change immediately, then discovering three call sites break. → Corrective: pass the four-question sufficiency test before the first edit; question (b) — callers and callees — is the one this failure skips.
• Exploration as procrastination. Reading a twentieth file "for context" when the edit site was clear ten files ago, because reading feels safer than committing to an edit. → Corrective: the no-new-KNOWN-in-two-reads rule; when it fires, write the model summary and start editing.
• Breadth without depth. Opening the first 30 lines of fifteen files and knowing the shape of nothing. → Corrective: pick one representative path and trace it fully to its side effect before opening any file off that path.
• Silent assumption stacking. Chaining "this probably validates input" → "so the bug must be downstream" without confirming the first link. → Corrective: every "probably" enters the ASSUMED list out loud; any assumption load-bearing for the fix gets promoted by reading the code before building on it.
• Doc-trust over code-trust. Building the mental model from a README or comments that no longer match the implementation. → Corrective: docs generate ASSUMED entries only; only code and command output generate KNOWN entries.
• Full restart on surprise. One wrong assumption during implementation triggers re-reading the whole subsystem from scratch. → Corrective: revise the single falsified ledger entry, re-trace only paths that depended on it, resume.

Worked example

Illustrative construction, not a recorded run.

Task: "POST /orders sometimes returns 500 when a discount code is applied. Fix it." Unfamiliar Node.js repo, ~40k lines.

• Agent restates the target: "Find where POST /orders handles discount codes, and what in that path can throw." (step 1)
• Agent lists the root tree and reads package.json → Express + Prisma, npm test exists. Candidate dirs: src/routes/, src/services/, src/db/. No source files opened yet. (step 2)
• Agent greps "orders" under src/routes/ → routes/orders.js maps POST to orderService.create. Greps discount → services/discount.js, services/order.js. Entry point located in 3 tool calls. (step 3)
• Agent traces one path: routes/orders.js → order.create() → discount.apply() → Prisma write → response, reading each file on the path fully. Observes discount.apply() calls code.toUpperCase() on the raw request field with no null guard. (step 4)
• Ledger. KNOWN: route wiring (routes/orders.js:24); apply() dereferences code unguarded (services/discount.js:11); error middleware converts any throw into a 500 (middleware/errors.js:8). ASSUMED: clients sometimes omit code; no upstream validation layer strips or defaults the field. (step 5)
• Agent promotes the assumption because the fix depends on it: greps for validation middleware on the orders route → none registered. Runs npm test — green, so the test loop works. Reproduces with curl -d '{"items":[...]}' → 500 with TypeError: Cannot read properties of undefined (reading 'toUpperCase'). The assumption becomes KNOWN via runtime evidence. (steps 5-6)
• Sufficiency test: edit site known (discount.js:11); callers/callees known (only order.create calls it; it only reads the codes table); verification known (rerun the curl repro, add a regression test); blast radius known (single call site). Four yes → exploration stops. (step 7)
• Agent declares the model in 4 lines, adds a guard returning 400 for missing/unknown codes, adds a test for the no-code case, reruns curl → 400, npm test → green. Total exploration: ~9 tool calls, one path traced end-to-end, zero unrelated files read. (steps 8-9)

Trace evidence

Case	Session	What it showed
awaiting case runs	—	—