A deterministic workspace gatekeeper that decomposes codebase changes into isolated, test-verified, atomic Git commits. Enforces architectural cohesion and SOLID invariants through a lean execution protocol.
"Swarming your messy diffs before they reach production."
Load SKILL.md directly into the LLM context window. The model follows the
rules manually. No MCP server required.
The MCP server at src/index.ts (built to dist/index.js) exposes eleven tools programmatically:
init-- Map project conventions and verify environmentaudit-- Scan workspace diffs and generate commit planswarm-- Execute the full discovery-cluster-test-commit looptooling-- Inventory Makefile targets,scripts/helpers, and package scriptslint-- Run lint (Makefile target first, manifest fallback)format-- Run formattest-- Run testbuild-- Run buildverify-- Run the full quality gate (format, lint, test, build)compliance-- Audit deliverable tiersversion-- Report the current LCCST version
These tools map 1:1 to the slash commands documented in SKILL.md. The skill
runs standalone; the server is an optional programmatic front end.
Every path-taking tool accepts a path argument (default .). Absolute paths
preserved; ~ expands to $HOME; set LCCST_WORKSPACE to override the
workspace root. See src/index.ts for implementation details.
# Init detects the manifest and runs the native test command:
/init -> Detects go.mod -> swarm runs `go test ./...`
/init -> Detects Cargo.toml -> swarm runs `cargo test`
/init -> Detects pyproject.toml -> swarm runs `uv run pytest`
/init -> Detects package.json -> swarm runs `pnpm test`
/init -> Detects CMakeLists.txt -> swarm runs `cmake --build .`- UNIX philosophy over framework: One skill file and one server file. No scaffolding the domain does not justify. Over-engineering is a correctness defect.
- User conventions first: Existing patterns, manifest commands, and explicit preferences take priority. Atomic hunk isolation, the Tooling Ladder, and strict test-pass verification are non-negotiable.
- Quality over velocity: Structural integrity and complete test verification beat raw speed. Token discipline applies to output, not internal reasoning.
- Granularity over convenience: Locality Clustering groups diffs by domain so each atomic commit rolls back cleanly. The extra commits buy a clear, reversible history.
- Proportional defence: Validation, rate limiting, and caching apply only where module exposure justifies them. Omit fabricated attack or load scenarios.
- Ecosystem-native discovery: LSP, Tree-sitter, and native test runners trace side effects; the Tooling Ladder prefers project scripts over bare binaries.
- Token investment: LCCST puns on low-cost asset management while clustering by locality. Tokens are strategic capital -- spent on tests, typing, and the Tooling Ladder, not on boilerplate.
Download the latest release from the
releases page. Each release
bundles dist/index.js, SKILL.md, dist/index.d.ts, LICENSE, and
README.md. Set the path to dist/index.js in your agent config. No install
or build needed.
For instruction-driven workflows that need no background processes.
- Claude Code CLI:
claude "Review the active git diff using the parameters in ./SKILL.md" - GitHub Copilot & OpenCode: Attach
#SKILL.mdor@SKILL.mdin chat - Codex & harnesses:
cat SKILL.md | your-agent-runner "Apply this system execution skill" - Project-level binding: Symlink
SKILL.mdas.cursorrules,.clinerules, or.github/copilot-instructions.md - Global profiles: Paste into Cursor Rules, Windsurf Memories, VS Code
globalRules.json, or JetBrains Custom Prompts
npx skills add bladeacer/lccst
/lccstgit clone --depth 1 https://github.com/bladeacer/lccst
cd lccst && pnpm install && pnpm run buildAdd to your harness config:
{
"mcpServers": {
"lccst": {
"command": "node",
"args": ["/absolute/path/to/lccst/dist/index.js"]
}
}
}Replace the path with your actual
dist/index.js. The server is disabled by default (enabled: false).Note: A harness is a programme you use to interface with and run AI models
OpenCode: Add the above to opencode.jsonc under mcp.lccst. The
SKILL.md is auto-discovered as an Agent Skill -- use /lccst or @SKILL.md
to invoke it.
Read AGENTS.md for build/test commands, deliverable tiers, and structural invariants.
| Tool | Version | Purpose |
|---|---|---|
| pnpm | >= 9 | Package manager |
| TypeScript | >= 5.4 | Compiling engine source |
pnpm run build # Bundle deps + source -> dist/index.js
pnpm run test # Run all tests
pnpm run test:swarm # Swarm library unit tests only
pnpm run test:telemetry # Telemetry MCP unit tests only
pnpm run test:picker # Benchmark picker unit tests only
pnpm run test:e2e # Telemetry end-to-end tests on real harnesses
pnpm run test:mcp # MCP server integration tests only
pnpm run bench:list # Models that the benchmark picker offers
pnpm run bump 1.0.0 # Bump version across all filesThe end-to-end test starts a real harness with a real model, so it needs a
network connection. Set E2E_HARNESS or LCCST_E2E_MODEL to narrow the run.
Benchmarking has its own dependencies -- see playground/README.md.
Measures token impact of skill-guided vs plain code generation across three reference projects (Python HTTP server, React timer, Go login CRUD).
make benchmark-free # Ask for a harness and a model, then run the benchmark
make bench-list # Show the models that the picker offersThe picker lists the models of every harness that the path holds, and it asks
for one in a fuzzy search. Type part of a model name and press Enter. Without
fzf, the picker asks for a filter and a number instead. A model is offered
only when the registry states that it calls tools, reads text, and writes text,
and when the registry states no charge for it, or when its name carries a free
marker. Routers, image models, and families that an author rejects for agentic
work are left out. The run then happens in a clean room outside the repository.
The target seeds SKILL.md, README.md, guide.md, traps.md, and
agent-prompt.md into the clean room, and it copies agent-prompt.md to
AGENTS.md, which the harness reads as the project instructions. The target
then starts the harness in the foreground so you can steer it. It settles the
token counts, writes the report, removes the clean room, and refreshes the table
below. Set BENCH_PICK=0 to skip the picker and pass HARNESS, PROVIDER,
and MODEL_NAME yourself.
The picker needs a terminal, because the run that follows costs model tokens.
Set BENCH_ALLOW_PIPE=1 only when a script must choose a model.
The run measures the model it names. The run checks that the harness offers the model before it builds the clean room, and it gives the harness its own configuration directory and a private server, so the plugins, skills, agents, and servers of a user cannot change the run. The run also gives the harness its own state directory, because a harness restores the model that a user last chose for each agent, and that restored model wins over the model of the run. Kilo also receives the model on the command line. The report compares the model identifier that the run pinned with the model that the store of the host recorded. A report that names a different model is marked as wrong.
The run measures the prompt it names. The instructions of a run carry a run token, and the model must state the token in its first reply. The settle step reads the text of the model turns and looks for the token. A run whose model never stated the token is a run that measured a different prompt, so the report marks the prompt as unverified.
Both columns are measured the same way. The scanner applies one file filter and one test command to the
plainand theskill-guidedvariant, and it scores both with the same rubric. A test failure is a measurement. A test run that cannot start, that times out, or that holds no test is an error, and an error never scores as a pass. Theplainvariant readsREADME.mdandguide.md. Theskill-guidedvariant reads those files plusSKILL.mdandtraps.md.
A run enters the table only when every check passes. The table needs the version of the harness, a measured count for every phase, a model that the host store confirms, a stated run token, and a passing skill-guided variant at 100 percent for all three subprojects. A run that fails a check stays on disk, and the table leaves it out. The report states the reason.
Every report names the harness version. A harness version changes the prompt, the tool list, and the token accounting, so a report that names no version cannot be reproduced. The scanner asks the harness for its version and records the version with the path of the command.
How runtime tokens are measured. A model cannot read its own token usage, so the
lccst-telemetryMCP server measures it. The model marks the end of each phase, and the settle step then reads the token counts of every model turn of that phase from the store of the harness that ran it. The reader names no harness, so the same server measuresopencodeandkilo.The server checks the workspace before it records a phase. A call that arrives before the phase holds its directory and its manifest records nothing, and a second call for the same phase corrects the end of that phase. A call that came too early therefore costs one more call, and it never costs a phase.
Every run before
v3.7.0used counts that the model under test supplied, so those Agent Runtime Tokens (ART) figures were estimates. The old findings are removed. The table below fills again after the next benchmark run.The scanner changed after the first
v3.7.0run, so the file token counts and the robustness scores of a new report are not comparable with the removed reports. A report that predates the change counted the test files of one variant and not the other.
| Provider | Harness | Harness Version | Model | Skill Layer | Context Tools (MCP) | Subproject | Plain Score | Skill-Guided | Test Status | FCT (Plain) | FCT (Guided) | ART (Plain) | ART (Guided) |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
opencode |
opencode | v2.0.20 |
space-bunny-free |
v3.8.0 |
lccst-telemetry |
python-http-server | 83/100 | 100/100 | PASSED | 2,345 | 8,116 | 179,173 | 502,581 |
opencode |
opencode | v2.0.20 |
space-bunny-free |
v3.8.0 |
lccst-telemetry |
react-timer | 47/100 | 100/100 | PASSED | 2,544 | 4,166 | 899,775 | 565,717 |
opencode |
opencode | v2.0.20 |
space-bunny-free |
v3.8.0 |
lccst-telemetry |
go-login-crud | 100/100 | 100/100 | PASSED | 4,457 | 13,452 | 1,729,503 | 2,541,943 |
| Summary | Workspace Totals / Avg | 77/100 | 100/100 | 3/3 Passed | 9,346 | 25,734 | 2,808,451 | 3,610,241 |
Highest ART subproject:
go-login-crudconsumed the most guided runtime tokens. Highest FCT subproject:go-login-crudconsumed the most guided FCT tokens. Skill-guided implementation used +175% more FCT and +29% more ART compared to plain implementation across the workspace suite.
| Provider | Harness | Harness Version | Model | Skill Layer | Context Tools (MCP) | Subproject | Plain Score | Skill-Guided | Test Status | FCT (Plain) | FCT (Guided) | ART (Plain) | ART (Guided) |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
kilo |
kilo | v7.7.9 |
stealth-space-bunny-alpha |
v3.8.0 |
lccst-telemetry |
python-http-server | 83/100 | 100/100 | PASSED | 2,397 | 8,309 | 230,291 | 781,186 |
kilo |
kilo | v7.7.9 |
stealth-space-bunny-alpha |
v3.8.0 |
lccst-telemetry |
react-timer | 47/100 | 100/100 | PASSED | 1,533 | 3,725 | 444,659 | 1,145,016 |
kilo |
kilo | v7.7.9 |
stealth-space-bunny-alpha |
v3.8.0 |
lccst-telemetry |
go-login-crud | 100/100 | 100/100 | PASSED | 6,260 | 16,756 | 1,073,902 | 3,057,677 |
| Summary | Workspace Totals / Avg | 77/100 | 100/100 | 3/3 Passed | 10,190 | 28,790 | 1,748,852 | 4,983,879 |
Highest ART subproject:
go-login-crudconsumed the most guided runtime tokens. Highest FCT subproject:go-login-crudconsumed the most guided FCT tokens. Skill-guided implementation used +183% more FCT and +185% more ART compared to plain implementation across the workspace suite.
| Metric | opencode-opencode-space-bunny-free |
kilo-kilo-stealth-space-bunny-alpha |
|---|---|---|
| Harness | opencode | kilo |
| Harness version | v2.0.20 | v7.7.9 |
| Plain score | 77/100 | 77/100 |
| Guided score | 100/100 | 100/100 |
| Plain FCT | 9,346 | 10,190 |
| Guided FCT | 25,734 | 28,790 |
| FCT overhead | +175% | +183% |
| Plain ART | 2,808,451 | 1,748,852 |
| Guided ART | 3,610,241 | 4,983,879 |
| ART overhead | +29% | +185% |
| Tests passed | 3/3 | 3/3 |
All evaluated models (opencode-opencode-space-bunny-free and
kilo-kilo-stealth-space-bunny-alpha) achieved a perfect guided score of
100/100 under the protocol. However, their resource efficiency varied
significantly:
-
opencode-opencode-space-bunny-freeentered with the strongest plain baseline (77/100) and reached perfection with +175% FCT and +29% ART overhead -- representing a genuine quality investment rather than recovery from failure. -
kilo-kilo-stealth-space-bunny-alphaalso delivered a perfect guided score, with +183% FCT and +185% ART overhead.
Across all runners, go-login-crud remained the most resource-intensive
subproject.
opencode-opencode-space-bunny-free consumed the fewest tokens overall
(6,453,772): 9,346 plain FCT, 25,734 guided FCT, 2,808,451 plain ART, and
3,610,241 guided ART.
| Rank | Agent-Model | Plain Score | Guided Score | FCT Overhead | ART Overhead | Verdict |
|---|---|---|---|---|---|---|
| 1 | opencode-opencode-space-bunny-free |
77/100 | 100/100 | +175% | +29% | Best overall |
| 2 | kilo-kilo-stealth-space-bunny-alpha |
77/100 | 100/100 | +183% | +185% | Strong contender |
See model-ranking.md for the full ranking of all benchmark runs.
Only benchmark runs which perform well enough are included