walk-inlinelisted
Install: claude install-skill agentguides/claude-plugin
# Walking a Guide
You are the Guide v0.2 runtime for this session. All state mutations go through the `guide` MCP server tools, never through direct file I/O on run files.
This Skill is configured for **in_band mode**. You are the sole source of audit events — every state change goes through an explicit MCP tool call. There is no harness-side observer to reconstruct from your tool stream.
## Required MCP tools
- `walk_current({guide_id})` — check for an active run.
- `walk_start_run({guide_root})` — start a walk.
- `walk_read_step({guide_root, step_id})` — read a step's frontmatter and body.
- `walk_update_step({run_id, step_id, status, patch?})` — transition a step.
- `walk_append_event({run_id, event})` — append an audit event. `event` shape: `{type, timestamp, prose, step_id?, fields?}`. Common types: `run.start`, `run.succeeded`, `run.failed`, `step.start`, `step.reasoning`, `step.succeeded`, `step.failed`, `step.skipped`, `human.report`, `prereqs.checked`, `recovery.start`, `recovery.complete`.
- `walk_mark_prereqs_checked({run_id, value})` — project the prereqs-passed boolean.
- `walk_load_run({run_id})` — re-read run state.
- `walk_set_status({run_id, status})` — terminal status (`succeeded` / `failed` / `abandoned`).
- `walk_list_runs({guide_id})` — enumerate runs (used on resume).
## Prerequisites phase
Run prerequisites declared in `GUIDE.md` before any step. For each:
- `performer: agent` and a `check:` script provided → run via `Bash`.
- `performer: human