One candidate corpus
Every arm sees the same scanned files, so a weaker baseline is never manufactured by changing the search space.
The product
One problem.
A map that stays useful.
FixMap narrows the first step, explains its reasoning, and checks the work that followed. It is a map you can inspect—not a promise that the map is always right.
Measure it locally
Backtest the map on your own history.
fixmap benchmark --repo . --last 50 compares BM25-over-code, ordinary FixMap context, and FixMap with Impact Graph against historical parent snapshots.
Pre-change cutoff
Each case is evaluated on its parent snapshot. The target change and later Git history cannot leak into its evidence.
Raw cases included
All, mentioned, and unmentioned cohorts plus Wilson intervals make misses and small samples visible.
No repository code runs
The benchmark reads Git and source text in temporary worktrees without installing dependencies, running hooks, or executing tests.
While the agent edits
Watch the implementation drift—or stay aligned.
fixmap watch --report plan.json --repo . emits only when the working tree changes, then re-runs Verify and recalculates the Impact Graph around the real diff.
Changed states only
A lightweight fingerprint avoids repeating full scans when nothing moved.
Drift made visible
Unmapped edits, untouched leading context, and new impact relationships appear as evidence, not verdicts.
Agent-ready stream
Markdown stays readable; JSON Lines gives automation one complete record per update.
Still local-only
Watch reads Git and source text without running repository code, installing dependencies, or calling a model.
1
You describe the issuein plain English
TOKEN_TTL_MINUTES is ignored and reset links expire immediately.
2FixMap finds the few places most likely to matter
Files to inspect
1src/auth/reset-password.ts · high, score 51
2src/email/templates/reset.ts · medium, score 23
3src/auth/token-store.ts · medium, score 21
Likely impact
1test/auth/reset-password.test.ts · co-change
2src/http/routes.ts · co-change
3src/config.ts · imports
Risks to review
1low authentication · ranked files touch authentication; review this area before editing, but no diff evidence is available yet
01
Before the edit
Plan: find the few places that matter.
FixMap reads the task and repository together. It ranks primary context, then builds a separate Impact Graph from imports, reverse dependents, routed tests, and repeated Git co-change evidence.
- Ranked source files with reasons
- Impact files to inspect, not assumed edits
- Workspace-aware tests, risks, and diagnostics
02
Before the agent reads
Context: send the source, not just its address.
Context packages deterministic line ranges from primary and impact files inside an estimated source-token budget. Every snippet keeps its role, reason, confidence, range, and truncation state.
- Markdown or structured JSON
- Visible UTF-8 byte estimate
- Explicit omissions and scanner bounds
03
When relationships matter
Graph: make the evidence portable.
Graph exports the Impact Graph as Mermaid for review documents or versioned JSON for tools, preserving the direction and reason for every import, dependent, test, and co-change edge.
- Mermaid or versioned JSON
- Directional relationships
- No invented dependencies
04
When the map surprises you
Explain: ask why a file is missing.
A ranked list explains what it chose. Explain handles the harder question: whether a path ranked lower, fell below the cutoff, was excluded intentionally, or never entered the scan.
- Inspect one path directly
- Distinguish ranking from exclusion
- Surface scan-limit uncertainty
05
When you refine the task
Compare: check whether a better task moved the answer.
Naming a symbol, an error string, or a path usually changes the ranking. Compare puts two plans side by side and reports what entered, left, moved, or changed confidence — so refining the task is measurable rather than a matter of impression.
- Two plans, one delta
- Entered, left, moved, changed confidence
- Same report from CLI and MCP
06
After the edit
Verify: compare the plan with the real change.
FixMap checks the saved plan against a git diff. It points out unplanned files, untouched leading context, missing tests, risky areas, and recalculated impact around the files that actually changed.
- Plan versus diff
- Recalculated impact
- Advisory findings by default
What makes the map useful
Grounded signals, not a black box.
FixMap is deterministic. Every ranking point comes from repository evidence that can be inspected and challenged.
Repository connections
Paths, imports, symbols, workspace boundaries, and related tests establish how code fits together.
Calibrated confidence
Confidence is capped when evidence is thin. Vague tasks can return an empty map instead of a plausible guess.
Explicit uncertainty
Scan limits, missing refs, generated files, and remote-fetch behavior appear as diagnostics rather than disappearing.
Safe by construction
FixMap reads and ranks. It does not install dependencies, run scripts, execute tests, or upload local source.
The boundary matters
A starting map, not proof.
FixMap gives an agent a focused, evidence-backed starting set for more deliberate review. It cannot prove a change is correct, complete, or safe.
That is why every score has reasons, Verify is mostly advisory, and the benchmark page publishes failures alongside hits.
See the evidence