workflow-debuggerlisted
Install: claude install-skill home-dev-lab/workflow-toolbox
# Workflow debugger — diagnose a Workflow run from its journal
When Claude Code's Workflow tool runs a script, it writes a **run journal** to disk:
```text
$CLAUDE_CONFIG_DIR/projects/<project-slug>/<sessionId>/workflows/wf_<runId>.json
(default config dir: ~/.claude)
```
That journal — not the launch message, not the per-agent transcripts — is the single
structured record of what happened: the final `status`, the `error` (with stack) if it
threw, every `agent()` call's state and token cost, the phases, and the totals. This
skill reads it, says **what went wrong**, and decides **whether `resumeFromRunId` helps**.
> **Journal-first.** Claude Code's own Workflow-tool docs name the per-agent
> `agent-<id>.jsonl` transcripts as the FALLBACK "when no journal is available". So this
> skill diagnoses from `wf_<runId>.json` and only drills into a transcript when you need
> one agent's full reasoning.
## Run it
```bash
# primary (plugin install — the bundled, zero-dependency CLI; nothing to install):
node "${CLAUDE_PLUGIN_ROOT}/bin/wt-debug.mjs" [runId|latest|<journal-path>] [--json] [--project <slug>]
# npm alternative (any project with @workflow-toolbox/build installed) — and the
# only path for the audit REPORT, which the bundled bin does not do:
npx workflow-toolbox debug [runId|latest|<journal-path>] [--json] [--project <slug>]
npx workflow-toolbox report [runId|latest|<journal-path>] [--project <slug>] [--out <dir>] [--quiet]
```
(`pnpm exec workflow-toolbox …` is the e