Jacu Fast

A workflow for coding agents, with an optional local evidence runner. The skill helps the agent work; the runtime records what actually ran against which files. Neither is an independent proof that every business requirement was understood.

Choose a mode

The 0.3.0 runtime fixes the unsafe 0.2.0 cache identity. Do not reuse 0.2.0 receipts or the old binary. Source and binary versions are checked before execution.

Install in Claude Code

claude plugin marketplace add https://github.com/jacu-dev/jacu-fast.git
claude plugin install jacu-fast@jacu-fast

Start a new local Claude Code session, including the Code tab in Desktop. The plugin skill is:

/jacu-fast:jacu-fast implement the requested change and verify delivery

You may also ask in ordinary language: “Use Jacu Fast for this task.” Model invocation is enabled; the SessionStart hook explains the installed skill name. This makes discovery possible, not a guarantee that a probabilistic agent always selects it. The exact slash invocation is the unambiguous entry point.

/jacu-fast is the standalone skill name, not the namespaced plugin command. There is no separate /jacu-fast:run command. Installing this plugin does not install it in unrelated cloud chat sessions.

For updates, refresh the marketplace and plugin, then start a new session:

claude plugin marketplace update jacu-fast
claude plugin update jacu-fast@jacu-fast
claude plugin list

The package-level diagnostic is python3 PLUGIN_ROOT/scripts/host.py doctor. Replace PLUGIN_ROOT with the installation path reported by Claude. It distinguishes “files present” from “loaded in this host session”; it cannot inspect a Desktop session from a separate process.

Install only the skill

npx --yes skills add jacu-dev/jacu-fast -g -y --skill jacu-fast -a claude-code

The standalone invocation is /jacu-fast. The copy is useful without a binary. The agent must report skill-only mode, run the project's existing tools itself, and not invent a runtime completion result. Do not install both forms merely to work around a command-name mismatch.

Agent Skills-compatible hosts may reuse the procedure. The native event adapter in this repository is for Claude Code. A portable manifest does not mean Codex, OpenCode or another host's event integration was tested. There is no OpenCode npm plugin and no MCP server.

What happens during a task

The agent captures the original request and a compact outcome contract, inspects branches and linked worktrees before duplicating work, and uses existing project conventions. With the runtime available, it prepares a session against the actual delivery directory/branch, runs useful checks at iteration checkpoints, and obtains current delivery evidence. A Stop hook is activated only for that task and host session.

The hook evaluates saved evidence; it never starts the test suite or a model call. Missing evidence sends a bounded recovery instruction to the same agent. An unchanged repeated stop or real terminal blocker ends with an explicit incomplete result instead of an approval prompt or infinite retry. Inactive Stop hooks do no repository work.

The runtime discovers existing Cargo, Go, Swift and common Node scripts when .jacu-fast.json is absent. Discovery does not install dependencies or invent missing tests. Unknown test output remains unknown. Use an explicit policy for project-specific runners or ignored fixtures.

What the runtime verifies

• Content identity includes tracked and untracked file bytes, index state, file modes, symlink inputs, initialized submodules, and explicitly declared ignored inputs. It is not just the commit or Git status.
• A receipt is bound to the request, candidate, check definition, effective environment and executable/toolchain fingerprint. Failures are not success; a diagnosed transient failure can be retried explicitly.
• The delivery directory and branch are pinned. Required checks and already-bound outcome statements cannot be silently weakened or waived by a model-written amendment.
• Recognized zero-test runs, timeouts, truncated output and unsupported test counts do not pass. A semantic score needs its configured threshold and cannot replace behavioral evidence.

Limits: the agent still maps the request to outcomes and declares implementation state. Source quotes are checked when supplied, but coverage of an arbitrary natural-language request is not proven. Local receipts are mutable by the same OS user; this is not a tamper-proof attestation or security boundary against that user. Repositories and configured commands must already be trusted for execution. No general dependency-impact analysis, worktree-content search, automatic merge, assertion-semantic analysis or speedup benchmark is claimed.

Runtime and development

Prebuilt targets: macOS Apple Silicon and Linux x86-64 built on Ubuntu 24.04 (glibc 2.39 or newer). The launcher and hooks require Python 3.9+; the compiled CLI itself has no Python dependency. macOS Intel, Linux ARM and Windows use skill-only mode unless a compatible developer runtime is explicitly supplied. No runtime download or compilation occurs silently during an agent task.

bin/jacu capabilities
cargo test --locked
cargo build --release --locked
JACU_TEST_BIN="$PWD/target/release/jacu" python3 -m unittest discover -s tests -p 'test_*.py' -v
JACU_BIN="$PWD/target/release/jacu" python3 scripts/host.py doctor

JACU_BIN is an explicit developer override, not a PATH fallback. Packaged builds instead require runtimes/manifest.json, a matching source fingerprint, binary version and checksum. Developer overrides are identified as such by the diagnostic; they are not claimed to have packaged-build provenance.

See the runtime contract, the engineering scope, and the acceptance evidence map. CI tests the actual compiled CLI on Linux and macOS, the Python hook protocol, package consistency and an isolated Claude CLI installation. An authenticated end-to-end model session and your specific Desktop installation remain separate checks, not implied by unit-test success.

prepare, verify, report, clean, and capabilities are the native commands. clean only removes owned expired diagnostics for the selected repository; it does not delete branches, worktrees, build caches or receipts. Timing fields are measured operation timings, not claimed token or cost savings.

License

MIT. Keep the copyright and license notice when redistributing. Contributions and public examples must be English and synthetic; do not publish credentials, private projects or raw user sessions.