redsen-lean-harness

An imperative, token-lean, self-documenting agent harness for GitHub Copilot CLI and VS Code Copilot.

It imposes a spec-driven pipeline, runs the implement→verify stage as a Ralph loop, drives parallel dynamic workflows on its own using git worktree lanes, and writes down everything it did. Mostly markdown. One runtime dependency.

---

Why it exists

Coding agents fail in predictable ways: they guess at requirements, load far too much context, lose what they learned between sessions, work sequentially when the work is parallel, and leave no trace of why anything was done.

This harness fixes each of those with a specific mechanism, not with prompt-engineering hope.

Failure	Mechanism
Guessing at requirements	interrogator asks questions with recommended answers; the pipeline blocks until answered
Context bloat	Only memory/INDEX.md is ever auto-loaded; everything else is pulled on demand
Amnesia between runs	Categorised memory shards, committed to the repo
Sequential work	Dynamic DAG + git worktree lanes with file-scope leases
No audit trail	ADRs, living specs, and a per-run journal — written by a cheap agent, continuously
Unverifiable "done"	Five explicit Ralph exit criteria, all machine-checkable

Design

Three layers, one hard boundary — we never reimplement an agent runtime (ADR 0001).

Layer	Owns	Where
Markdown	Behaviour — agents, skills, instructions	.github/
Node CLI (lh)	Determinism — index, gates, lanes, telemetry, memory	src/
Host	Execution — subagents, fleet, tools, models	Copilot CLI / VS Code

lh never calls a model. Agents never do git plumbing, parsing, or token arithmetic by hand.

Install

GitHub Copilot CLI

# direct from the repo
copilot plugin install redsentech/lean-harness

# or via the marketplace
copilot plugin marketplace add redsentech/lean-harness
copilot plugin install redsen-lean-harness@redsen

VS Code Copilot

Clone the repo and point VS Code's Copilot customization settings at it, or copy the .github/ tree into your project. .github/agents/, .github/instructions/ and .github/mcp.json are read by both hosts.

.github/skills/ is Copilot CLI only. VS Code gets equivalent behaviour through .github/agents/ and .github/prompts/.

The lh CLI

npm install -g @redsen/lean-harness   # or: npx @redsen/lean-harness <command>
lh doctor                             # verify the environment

Context7

Mandatory before any external-library work. The key is read from the environment only — the harness never writes secrets to disk.

export CONTEXT7_API_KEY="<your key>"   # add to ~/.bashrc or ~/.zshrc

Quick start

cd your-repo
lh init            # first-run wizard; writes .agents/harness.config.json
lh doctor          # confirm everything is wired

Then, in Copilot CLI or VS Code Copilot:

@conductor build a rate limiter for the public API

For a small brownfield change, skip the ceremony:

/fast-track fix the off-by-one in pagination

To learn an unfamiliar codebase first:

/onboard

Onboarding a new project

Before the plugin is published to a marketplace (or if you want to try it against a local repo first), scripts/onboard.mjs does the whole bootstrap in one shot: copies the behaviour layer (.github/agents, skills, instructions, prompts, mcp.json, copilot-instructions.md, AGENTS.md) into a target repo, npm links the lh CLI so it's on PATH there, runs lh init and lh doctor, and prints next steps.

# from inside this repo
node scripts/onboard.mjs /path/to/target-repo --dry-run   # preview, writes nothing
node scripts/onboard.mjs /path/to/target-repo --yes        # do it

Flag	Effect
--yes	Non-interactive: auto git init if needed, run lh init without asking
--force	Overwrite target files that already exist and differ (default: skip + warn)
--dry-run	Print the plan, write and run nothing
--no-npm-link	Skip npm link; print the absolute node .../src/cli.mjs invocation instead

Safe to re-run: files that are already identical are left alone, and re-running lh init on an already-initialized repo warns instead of failing (rerun with lh init --force to reset). npm install -g @redsen/lean-harness (once published) replaces steps 3–5 of the script with a normal global install.

Pipeline

design    → interrogator asks, user answers          → decisions.md      [USER GATE]
plan      → architect writes spec + acceptance criteria
            splitter emits plan.dag.json (lanes + file scopes)
build     → lh host picks the strategy
              read lanes  → shared checkout, N-wide scout fan-out
              write lanes → lh lane create → git worktree + branch
                            builder ⇄ verifier   (RALPH loop)
            checkpoint  → conductor RE-PLANS, may spawn/kill/re-scope lanes
            failure     → isolate → bounded retry → re-plan around it
integrate → integrator merges lanes SEQUENTIALLY, then one full verify
document  → scribe writes journal, ADRs, spec, conventions   (throughout)

Ralph exit criteria

A lane is done only when all of these hold. Otherwise it iterates; at max iterations it escalates to you.

1. Every declared verify command exits 0
2. Every acceptance criterion is individually checked off
3. lh graph structural gate passes
4. reviewer approves
5. No files were changed outside the lane's declared scope

Agents

Agent	Tier	Invokable	Responsibility
conductor	strong	yes	Entry point. Owns the pipeline and the dynamic DAG.
interrogator	strong	yes	Design-phase Q&A with recommended answers.
scout	cheap	–	Read-only recon, fanned out N-wide.
architect	strong	yes	Spec, acceptance criteria, architecture doc, ADRs.
splitter	mid	–	Decomposes the spec into lanes with file-scope globs.
builder	mid	–	Implements one lane inside its worktree.
verifier	cheap	–	Runs verify commands + the structural gate.
reviewer	strong	yes	Acceptance-criteria and scope gate.
integrator	strong	–	Sequential merge, conflict escalation, full verify.
scribe	cheap	–	Journal, ADRs, living spec, conventions.

Tiers map to models in harness.config.json and are fully overridable. The default map is cheap → claude-haiku-4.5, mid → claude-sonnet-5, strong → claude-opus-5. Expensive exploration is deliberately pushed onto the cheap tier; only summaries return to the main context.

lh commands

Command	Purpose
lh init	First-run wizard → .agents/harness.config.json
lh index [--budget N] [--focus g] [--fetch]	Token-budgeted tree-sitter repo map
lh graph	Structural gate, terse output by default. Exit 1 on violations
lh lane create|list|status|merge|drop	Worktree lanes + file-scope leases
lh run start|event|end	NDJSON telemetry + live board
lh memory get|put|compact|scan|list	Memory shards, compaction, secret scanning
lh host	Detect host capabilities, print the orchestration strategy
lh report [runId] [--journal]	Markdown telemetry report / run journal
lh doctor	Environment checks

Full flag reference

Every flag each subcommand reads (click to expand)

• lh init [--yes] [--force] [--json] — --yes skips prompts with defaults; --force re-initializes an existing .agents/harness.config.json.
• lh index [--budget N] [--focus <glob>] [--fetch] [--force] [--stats] [--json] — --fetch lazily downloads/caches missing tree-sitter grammars; --force rebuilds the cache instead of reusing it; --stats prints token counts instead of the map.
• lh graph [--severity <level>] [--fix-manifest] [--json] — terse brief is the default output; --fix-manifest rewrites codegraph.manifest.json-style baselines.
• lh lane create --id <id> [--title t] [--kind read\|write] [--scope glob...] [--depends-on id] [--base ref] [--json]
• lh lane list [--status s] [--run-id id] [--json]
• lh lane status --id <id> [--json]
• lh lane merge --id <id> [--abort] [--strategy s] [--json]
• lh lane drop --id <id> [--force] [--json]
• lh run start --objective "..." [--spec slug] [--host h] [--strategy s] [--attrs '{"k":"v"}'] [--json]
• lh run event --run <id> --type <t> [--name n] [--status ok\|fail] [--agent a] [--lane id] [--model m] [--operation op] [--input-tokens N] [--output-tokens N] [--duration-ms N] [--exit-code N] [--attrs '{"k":"v"}'] [--json]
• lh run end --run <id> [--status ok\|fail] [--summary "..."] [--json]
• lh run list [--json] / lh run show [runId] [--json]
• lh memory list [--shard s] [--json]
• lh memory get --shard s [--query q] [--limit N] [--json]
• lh memory put --shard s --title t [--body "..." | piped via stdin] [--tags a,b] [--allow-secrets] [--json] — writes are refused if a secret pattern matches unless --allow-secrets is set.
• lh memory compact [--force] [--json]
• lh memory scan — secret scan only, no write
• lh host [--strategy s] [--json]
• lh report [runId] [--run-id id] [--json] [--board] [--journal] [--spec slug] [--write]
• lh doctor [--json]

Repository footprint

The harness writes into one configurable directory:

.agents/
  harness.config.json
  architecture.md          # living, ADR log inside       [committed]
  conventions.md           # living                       [committed]
  memory/INDEX.md          # the ONLY always-loaded file  [committed]
  memory/{seed,failures,corrections,insights,conventions,quirks}.md
  specs/<slug>/{questionnaire,decisions,spec}.md · plan.dag.json
  runs/<id>/journal.md                                    [committed]
  runs/<id>/{board.md,events.ndjson}                      [gitignored]
  .cache/                  # repomap, symbols, wasm, lanes [gitignored]

The committed/gitignored split is chosen during lh init.

Observability

lh run event appends OTEL-GenAI-shaped NDJSON. No collector, no server, works offline.

• board.md regenerates on every event: lanes in flight, ralph iterations, elapsed, tokens, and live burn rate.
• lh report gives the post-run breakdown by phase, agent, model and lane.
• journal.md is the committed, human-readable record of what happened and why.

There is no hard budget cap — a deliberate choice. The control is visibility: live burn rate while running, full cost breakdown after. See ADR 0006.

Token discipline

1. Only memory/INDEX.md is always loaded; shards are pulled on demand.
2. lh index --budget caps the repo map and degrades signatures → names → counts.
3. Cheap-tier subagents absorb exploration; only their summaries re-enter the main context.
4. lh graph and lh report emit briefs, never dumps.
5. Skill bodies stay short; procedures live in templates/, loaded only when used.
6. lh memory compact merges the oldest entries when the budget is approached.

Architecture decisions

ADR	Decision
0001	Markdown owns behaviour, Node owns determinism
0002	web-tree-sitter WASM index with an offline regex fallback
0003	git worktree per write-lane; pluggable isolation backend
0004	No vendoring of Elastic-licensed code
0005	File-only memory with a single always-loaded index
0006	Local NDJSON telemetry, no server, no budget cap

Development

npm install
npm run validate     # manifests + every markdown frontmatter block
npm test
node src/cli.mjs --help
node scripts/onboard.mjs <target-dir> --dry-run   # preview bootstrapping another repo

npm run validate is the distribution gate: it checks plugin.json, marketplace.json, agent/skill/instruction frontmatter, delegation targets, and scans .github/mcp.json for committed secrets. npm test covers config, memory, lane scope safety, glob matching, secret scanning, telemetry, and the onboarding script end to end (tests/onboard.test.mjs).

Prior art

Ideas taken (no code, no dependencies) from pi-hermes-memory (categorised memory shards, secret scanning, consolidation) and pi-context-mode (token budgeting, compaction, checkpoint anchors — Elastic Licensed, deliberately not vendored, see ADR 0004). The structural gate is modelled on codegraph; the Ralph loop and run-state telemetry on rapid-prototyping-agent; the packaging on redsen-copilot-agents.

Roadmap

v1 — spec pipeline, Ralph loop, memory, index + structural gate, worktree lanes, integrator, telemetry.

v2 — devcontainer isolation backend with a generator/manager/updater ecosystem (lh lane is already backend-pluggable), semantic memory retrieval behind the existing getMemory() interface, and OTEL export.

License

MIT © Redsen