Reporecall

Local-first context + memory for coding agents
Automatic injection + an explicitTrust Contractso agents (and you) always know when the context is fresh.

Reporecall helps Claude Code, Codex, Cline, Aider and other agents work more effectively onlarge, high-churn or unfamiliar codebases.

Both measured on a real 1,306-file / 5,591-chunk codebase over 30 pre-registered queries, withno model calls— so you can reproduce them exactly. The token figure is context-assembly cost, not end-to-end agent tokens; we don't publish an end-to-end number because we haven't earned one yet. SeeBenchmarking & Token Evidence.

User prompt │ ▼ Claude Code hook ──► Reporecall (local) │ │ │ [check freshness] │ │ [route intent] │ │ [select + compress]│ ▼ ▼ Injected context + banner (optional MCP tools for gaps)

Plays nicely with:Claude Code (hooks), Codex (MCP/CLI), Cline, Aider, and any MCP-compatible coding agent.

📖 Full docs + honest competitive analysis(current position, target position, and threat matrix):https://proofofwork-agency.github.io/reporecall/

npm install -g @proofofwork-agency/reporecall reporecall init && reporecall serve

reporecall lens --serve --open # one-file architecture dashboard reporecall explain "..." # per-question diagnostics + evidence reporecall stats # Trust Contract + freshness at a glance

See the full, honest comparison + threat analysis indocs/competitive-positioning-2026.md.

Full documentation is hosted onGitHub Pagesathttps://proofofwork-agency.github.io/reporecall/:

- Introduction
- Installation & Quick Start
- CLI Reference
- MCP Tools
- Configuration
- Architecture

flowchart TB Q["User or agent question"] Entry["Hook, CLI, MCP, or JSON command"] Intent["Intent classifier"] Search["Code retrieval"] Wiki["Wiki evidence"] Product["Product area evidence"] Memory["Project memory"] Resolver["Capability evidence resolver"] Selected["Selected context"] Agent["Agent reads selected files first"] Explain["explain --json"] Lens["lens --json / Lens HTML"] BusinessTools["MCP business tools"] Q --> Entry --> Intent Intent --> Search Intent --> Wiki Wiki --> Product Intent --> Memory Search --> Resolver Wiki --> Resolver Product --> Selected Resolver --> Selected Memory --> Selected Selected --> Agent Selected --> Explain Product --> Lens Wiki --> Lens Product --> BusinessTools

The important rule is file coverage over chunk volume. For trace and architecture questions, Reporecall tries to cover the relevant layers: entry/UI, state or service, controller or edge function, and shared helpers when those layers exist.

flowchart LR Query["Prompt"] Mode{"Mode"} Lookup["Exact lookup"] Trace["Flow reconstruction"] Bug["Symptom evidence"] Arch["Layer coverage"] Change["Affected surfaces"] Query --> Mode Mode --> Lookup Mode --> Trace Mode --> Bug Mode --> Arch Mode --> Change Trace --> Resolver["Capability evidence"] Arch --> Resolver Change --> Resolver Resolver --> Files["Selected files with provenance"] Resolver --> Areas["Product areas used"] Resolver --> Pages["Business pages used"]

Capability evidence is generic. It does not encode customer/project names or repository-specific file lists.

For trace, architecture, and change prompts, Reporecall can:

Reporecall exposes product-language context in three places:

These surfaces are additive product-language views over code evidence for external tools.

The schema is documented indocs/business-context-schema.md.

Business context is not fed back into core search as hard-coded rules. Hooks may append a small budgeted product-area evidence section for trace, architecture, and change prompts, but lookup prompts stay small and code retrieval remains grounded in source/wiki/graph evidence. Consumers should treat the business layer as a read-only product map with supporting evidence.

Most context tools are silent when they're wrong.
Reporecall is not.

This is why we collapsed the MCP surface and made freshness signals unavoidable.

{ "trust": { "banner": "⚠ ... STALE ...", "indexedCommit": "abc1234", "currentCommit": "def5678", "dirtyFiles": 14, "level": "stale" }, ... }

Context-assembly cost.Getting the right evidence in front of the model costs a median of 75.4% fewer tokens than reading the relevant files whole.

Measured, not estimated — and scoped precisely:

Fixture: benchmark/project-context-queries.json (30 queries, real 1,306-file repo) Baseline: whole-file tokens for the files that actually contain the answer, i.e. what grep-then-read costs. Counts ONLY known-relevant files, never the wrong files a real search would also open — so the measured saving is a floor, not a best case. Candidate: tokens RepoRecall injects for the same query. Guard: a query counts only if RepoRecall delivered every mustInclude file, so omitting evidence can never register as a saving. 30/30 passed. Model calls: none. Deterministic and reproducible. median baseline 5,215 tokens -> median injected 1,439 tokens median reduction 75.4% aggregate 86.4% by route R0 54.5% R1 70.4% R2 90.8%

npm run benchmark:context-cost -- --project /path/to/repo --output ./context-cost.json

For redacted aggregate injection + freshness evidence:

PRs with reproducible numbers on real repos are very welcome.

reporecall lens --serve --open reporecall lens --json > lens.json

- overview stats;
- Louvain communities;
- high-degree hub nodes;
- surprising cross-module edges;
- generated wiki pages;
- product areas that group related business capability pages;
- business capability pages with product-facing summaries and supporting files.