LeanRigor

Less context. Full engineering rigor.

A local-first, cross-agent engineering harness. LeanRigor removes unnecessary context from coding-agent sessions while preserving the design, test, security and verification gates the task's risk level actually requires.

npx leanrigor init

The contract

This is the part to read before the numbers.

• Required engineering gates are never removed to save tokens. Verification is mandatory at every risk level; a critical task cannot skip its threat model, approval or rollback plan for any reason, including a token budget.
• Savings count only for work that passed. A reduction achieved by producing a wrong answer is not a reduction, and it is excluded from every total.
• Every number states its measurement mode and coverage. A local estimate is never called provider usage, and two measurement modes are never added together into an unlabelled total.
• Nothing is destroyed. Every projection carries a handle that restores the original bytes, or is explicitly marked summary-only.
• Telemetry is off by default, and leanrigor telemetry inspect prints the exact payload before you decide.
• Energy figures are versioned estimate ranges, never datacenter measurements. See docs/environmental-methodology.md.

Measured result

From the deterministic corpus in evals/, reproducible with one command:

npx leanrigor benchmark

Metric	Value	Cases	Measurement
Median context reduction	93.3%	8 passing	byte-only
Pass-rate delta vs baseline	0.0 points	8	deterministic verifier
Completion rate	100%	8	—

Per-case figures, the raw result and the release-gate verdict are in docs/benchmarks/. Every percentage there states the case count and the measurement mode behind it.

What this does not claim. These cases measure what LeanRigor's own transformations do to a payload — no model is involved.

Skill evaluation

The skills have now been evaluated against a real model (Codex CLI, gpt-5.5), baseline versus with-skill, with deterministic checks:

Skill	Baseline	With skill	Uplift	Reps
senior-system-design	2/6	6/6	+66.7 points	1
product-brainstorming	0/5	4/5	+80.0 points	1
verification	16/20	18/20	+10.0 points	4

All three trigger descriptions are bounded: the router selected none of them on any of the nine non-trigger prompts.

Do not quote those numbers without the caveat. Running the verification suite three times on an unchanged configuration produced +40, +20 and −20 points. A 60-point swing means anything below roughly twenty points at n=1 is noise, and the two large results above are single runs. Full write-up, including the five defects found in the evaluation harness itself and the two found in the skills, is in docs/benchmarks/skill-eval.md.

What it does

Piece	Job
MCP gateway	Exposes 4 tools to your host instead of 200. Tools are searched, not broadcast; large results are stored locally and returned as a compact, handle-backed projection.
TokenLeaf Engine	Measures what was actually saved, per measurement mode, and refuses to count savings from failed work.
Rigor Gates	Classifies task risk deterministically — no model call — and selects the smallest sufficient set of engineering gates.
Verified Skills	Three portable Agent Skills with licenses, provenance records, context budgets and evaluation suites.

Supported hosts

Host	Status
Claude Code	supported
Codex	supported
Gemini CLI	not yet; adapter planned

leanrigor init detects what is installed, previews every file change, backs up anything it modifies, and never writes before you confirm.

Commands

npx leanrigor init              # install, with a preview and confirmation
npx leanrigor init --dry-run    # show the plan, write nothing
npx leanrigor init --uninstall  # restore the original files
npx leanrigor doctor            # diagnose the installation
npx leanrigor mcp serve         # run the gateway (hosts launch this)
npx leanrigor benchmark         # run the reproducible benchmark
npx leanrigor report            # local session report
npx leanrigor report --share    # local SVG card, aggregate counts only
npx leanrigor skills list
npx leanrigor skills install verification
npx leanrigor telemetry status

Privacy defaults

No account. No network call of its own until you enable telemetry. Prompts, source code, file paths, repository names and tool payloads are never sent anywhere, at any setting — the ledger and telemetry schemas have no field that could carry them. See docs/privacy.md.

Uninstall

npx leanrigor init --uninstall

Files are restored to their original bytes. The only thing left behind is an append-only audit record under .leanrigor/.

Current limitations

Stated plainly, because a harness that overstates itself is worse than none:

• Two of the three skill-uplift results are n=1, and run-to-run variance on this suite has reached 60 points. Re-run with --repeat before relying on them.
• Four of the five verification cases do not discriminate — the baseline passes them every time, so they measure nothing. That suite needs harder cases.
• Ablation has not been run. No section of any skill has yet been shown to earn the context it costs.
• Every skill number comes from one CLI and one model.
• The benchmark's gateway+workflow and gateway+workflow+skill conditions are not implemented yet; only baseline and gateway run today.
• Risk classification is regex-and-path based. It is deliberately conservative and will over-classify before it under-classifies, but it is not clever.
• The Codex adapter rewrites config.toml, which drops TOML comments. The install plan warns about this and the original is backed up.
• No Gemini CLI adapter yet.
• The published package bundles the internal @leanrigor/* workspaces; those APIs are not stable and are not published separately.

Documentation

• Getting started
• Measurement boundaries — what is measured, estimated and unknowable
• Environmental methodology
• Privacy
• Evaluations
• Security policy
• Governance

Contributing

Bounded, ownable extension surfaces are documented in docs/extensions/: projectors, host adapters and skill packs. Each has an acceptance contract, so a contribution can be judged against a stated bar rather than a maintainer's mood.

License

Apache-2.0. Third-party notices, and the reuse ledger, are in THIRD_PARTY_NOTICES.md. No third-party source or prose has been copied into this repository.