# Run > Run an FM-Agent full, incremental, or no-op correctness analysis for a Git project. Use when the user asks to analyze code, generate behavioral specifications, verify implementation against specifications, or assess changes for formal-methods bugs. - **Type:** Skill - **Install:** `agentstack add skill-fmagent-project-fm-agent-skill-run` - **Verified:** Yes — security-reviewed for prompt injection and unsafe behavior - **Seller:** [fmagent-project](https://agentstack.voostack.com/s/fmagent-project) - **Installs:** 0 - **Category:** [Agent Skills](https://agentstack.voostack.com/c/agent-skills) - **Latest version:** 0.1.0 - **License:** Apache-2.0 - **Upstream author:** [fmagent-project](https://github.com/fmagent-project) - **Source:** https://github.com/fmagent-project/FM-Agent-Skill/tree/main/plugins/fm-agent-skill/skills/run ## Install ```sh agentstack add skill-fmagent-project-fm-agent-skill-run ``` Requires the [AgentStack CLI](https://agentstack.voostack.com/docs/cli). Works with Claude Code, Cursor, and any MCP-compatible agent. ## About # FM-Agent analysis This is the sole public analysis entry point. The host coding agent is the plugin's equivalent of FM-Agent's original `main.py`: it coordinates local tools, reads the repository, writes specifications, and performs reasoning. Deterministic scripts own dispatch, locking, state transitions, and artifact validation. Never launch the original FM-Agent remote-LLM pipeline as a substitute for this workflow. Before invoking a script, read [runtime-path.md](../../references/runtime-path.md) and resolve `FM_AGENT_PLUGIN_ROOT`. This shared skill must work in both Codex and Claude Code; never use `CLAUDE_SKILL_DIR`. Read [agent-orchestration.md](../../references/agent-orchestration.md) before starting the selected pipeline. ## Public parameters Codex selects this skill from a natural-language request; this skill directory does not create a Codex slash command. When a client exposes a command entry, it may pass the following arguments to the same workflow: ```text [natural-language change note] [--submodule path ...] [--one-phase] [--extra-edge file-or-dir] [--knowledge file ...] [--isolate] ``` Treat all text not matching an option as the change note. Resolve paths relative to the target repository. Do not invent `$PROJECT`, `$RUN_ID`, or a scope: the caller supplies the project and `orchestrate.py dispatch` returns the run id, merged configuration, and mode as JSON. ## Dispatch and cleanup Parse the supplied argument list before dispatching. Do **not** pass `$ARGUMENTS` verbatim as `--note`: split these public options from the natural-language change note: - `--submodule PATH` (repeatable) - `--knowledge FILE` (repeatable) - `--extra-edge FILE_OR_DIR` - `--one-phase` - `--isolate` All remaining text is the change note. Preserve every option value and pass each option separately. First perform this read-only inspection. It never acquires a lock, writes state, or rebuilds CodeGraph: ```bash "$FM_AGENT_PLUGIN_ROOT/scripts/orchestrate.py" inspect \ --project "$PROJECT" \ [--submodule "$PATH"]... [--knowledge "$FILE"]... \ [--extra-edge "$FILE_OR_DIR"] [--one-phase] [--isolate] ``` If inspection returns `noop` and `refresh_observed_commit` is false, report its baseline commit and finish. Do not run `codegraph.py status`. If `refresh_observed_commit` is true, run the stateful `dispatch` command below **without** `--codegraph`; it writes the no-op record and refreshes only Git provenance, then finish. Only when inspection returns `full` or `incremental`, inspect CodeGraph: ```bash "$FM_AGENT_PLUGIN_ROOT/scripts/codegraph.py" status --project "$PROJECT" ``` If it is available, include the internal `--codegraph` option below and rebuild `$PROJECT/.codegraph/` automatically. Proceed directly. If it is not available, do not install software; dispatch without `--codegraph`, use `agent-static`, and record the fallback reason. A selected CodeGraph rebuild failure fails the run; do not silently change backend. After determining availability, run exactly one stateful dispatch command: ```bash "$FM_AGENT_PLUGIN_ROOT/scripts/orchestrate.py" dispatch \ --project "$PROJECT" \ --note "$CHANGE_NOTE" \ [--submodule "$PATH"]... \ [--knowledge "$FILE"]... \ [--extra-edge "$FILE_OR_DIR"] [--one-phase] [--isolate] [--codegraph] ``` The bracketed terms are placeholders, not literal shell text: omit an option when absent. For example, `review checkout changes --submodule backend --knowledge payments.md` must dispatch with note `review checkout changes`, `--submodule backend`, and `--knowledge payments.md` as distinct arguments. `dispatch` should return the inspected non-noop mode. Retain its `run_id` and execute only the selected pipeline. If `--codegraph` was selected, rebuild its generated index while the run lock is held before extraction or graph construction: ```bash "$FM_AGENT_PLUGIN_ROOT/scripts/codegraph.py" init --rebuild --project "$PROJECT" ``` During graph construction, read its normalized function and edge data with: ```bash "$FM_AGENT_PLUGIN_ROOT/scripts/codegraph.py" export --project "$PROJECT" ``` Before each phase call `pipeline.py phase-start`; after producing its artifacts call `pipeline.py phase-complete`. A failed gate means do not enter the next phase. On every exception, tool failure, or user-requested stop, run `pipeline.py fail`; it releases its owned lock while preserving artifacts and the final run record. `pipeline.py complete` likewise releases its owned lock. Report the last completed phase. When `pipeline.py phase-start` begins `phase_cleanup` in a `full` run, it automatically clears only old derived artifacts; it never removes business source or the current `fm_agent/phases.json`. Do not bypass this phase. ## Full and incremental execution Read [stage-gates.md](../../references/stage-gates.md) before starting. Then use [full-pipeline.md](../../references/full-pipeline.md) or [incremental-pipeline.md](../../references/incremental-pipeline.md) according to dispatch. Use [artifact-contract.md](../../references/artifact-contract.md) whenever writing an artifact. For specification work, read [specification-rules.md](../../references/specification-rules.md). For verification, read [hoare-reasoning.md](../../references/hoare-reasoning.md). For every `MISMATCH`, read [bug-validation.md](../../references/bug-validation.md). During `call_graph` or `rebuild_graph`, write one native top-down layer artifact for each phase in `fm_agent/phases.json`: `phase_01_topdown_layers.json`, then `phase_02_topdown_layers.json`, and so on. A layer entry must retain its original repository-relative `source_file`, so it can be checked against its phase. Do not merge phases unless `--one-phase` was selected. Then write the plugin-control precision record. For CodeGraph use: ```bash "$FM_AGENT_PLUGIN_ROOT/scripts/call_graph.py" record-precision \ --project "$PROJECT" --backend codegraph --precision exact \ --codegraph-index "$PROJECT/.codegraph/codegraph.db" ``` For agent-static analysis, use `--backend agent-static --precision best-effort` and state the fallback reason with `--reason`. Build the plugin control index after extraction: ```bash "$FM_AGENT_PLUGIN_ROOT/scripts/artifact_index.py" build --project "$PROJECT" ``` Write `MISMATCH` only for a function's own specification violation. If the function is affected solely by a mismatching callee, write `DEPENDENCY_RISK` with the affected callee IDs and do not send it to Bug Validator. For each direct CMake candidate, first run: ```bash "$FM_AGENT_PLUGIN_ROOT/scripts/probe_build.py" \ --project "$PROJECT" --run-id "$RUN_ID" --bug-id "$BUG_ID" ``` Use the resulting isolated build directory for the probe. Do not reuse a project `build/` directory or its `CMakeCache.txt`. Only after every phase gate succeeds may the agent call `pipeline.py complete` and release the lock as `idle`. Never modify business source with `[SPEC]` or `[INFO]` or expose raw full diffs in chat. Execute and describe only capabilities documented by this plugin's shared skills and references; do not infer features from the original FM-Agent project. ## Source & license This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code. - **Author:** [fmagent-project](https://github.com/fmagent-project) - **Source:** [fmagent-project/FM-Agent-Skill](https://github.com/fmagent-project/FM-Agent-Skill) - **License:** Apache-2.0 Install and usage instructions live in the source repository linked above. ## Pricing - **Free** — Free ## Security capabilities Automated source analysis of v0.1.0 — what this tool can access: - **Network access:** no - **Filesystem access:** no - **Shell / process execution:** no - **Environment & secrets:** no - **Dynamic code execution:** no *"Yes" means the capability is present in the source — more access means more to trust, not that it is unsafe.* ## Versions - **0.1.0** — security scan: passed — Imported from the upstream source. ## Links - Listing page: https://agentstack.voostack.com/l/skill-fmagent-project-fm-agent-skill-run - Seller: https://agentstack.voostack.com/s/fmagent-project - Browse the marketplace: https://agentstack.voostack.com/browse --- Listed on AgentStack — the marketplace for AI agent skills and MCP servers. Every listing is security-reviewed. Creators keep 70%.