Which workflow fits?
Measured evidence for the routing case. Fired in 1 of 4 runs -- the prompt is underdetermined, which is a finding about the prompt, not a plugin failure.
| mode | model | run | outcome | skills fired | hook denials | mode denials | diff | cost | time |
|---|---|---|---|---|---|---|---|---|---|
plan | haiku | #1 | FAIL | -- | 0 | 2 | -- | $0.34 | 94s |
plan | haiku | #2 | FAIL | -- | 0 | 2 | -- | $0.32 | 88s |
plan | sonnet | #1 | FAIL | -- | 0 | 0 | -- | $0.12 | 18s |
plan | sonnet | #2 | PASS | arbeitsplan:arbeitsplan-start | 0 | 0 | -- | $0.40 | 271s |
Why the failing runs failed
- expected skill(s) never fired: arbeitsplan:arbeitsplan-start
What the permission mode refused -- 4 call(s), tool(s): Bash. One verbatim:
find in '/' was blocked. For security, Claude Code may only search files in the allowed working directories for this session: '/private/var/folders/pb/1z5lx35d7y99w8rsvn2t6z8r0000gp/T/arbeitsplan-cell-i9zfqjmu'.
What the run said (plan, haiku, run #1)
The working directory is empty and not yet a git repository — so before anything else, that needs
git init(+ an initial commit), since arbeitsplan's builder agents each work in their own git worktree. For "which arbeitsplan workflow fits":arbeitsplan-startis the entry point — run that first with a description of the feature. From there the pipeline is: 1.arbeitsplan-start— kicks off the run for your feature 2.arbeitsplan-patterns— picks the agentic shape (fan-out/build-referee
Recorded 2026-09-14 from wf-2026-09-14-acceptedits, wf-2026-09-14-plan, wf-2026-09-14-plan-sonnet with 2.1.236 (Claude Code). Every number above comes from a cell record; regenerate with test/workflows/curate.py.