Agent workbench
npx -y skills add MuNeNiCK/agent-workbench --skill agent-workbenchAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
One thing to look at
- 0 stars0 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
What its author says it does
Copied from the file, not written here
Use when a coding agent needs durable work, design, review, evidence, interruption, recovery, and completion state through the native Agent Workbench CLI.
The file declares its own license as MIT. That is the author’s claim about this one file, and it is not the same thing as the license GitHub reports for the repository, which is listed with the other numbers below.
SKILL.md
2.9 KB, 576 tokens by cl100k_base, as published. Nobody here has run it
Agent Workbench
Use the native agent-workbench executable directly. The Skill is guidance
only; it does not bootstrap, download, wrap, or route the product.
State
Choose an explicit state file. Private Workbench metadata belongs under a project state area by default, but the user may place it elsewhere:
agent-workbench --state .agent-workbench/state.sqlite3 init \
<owner> <outcome> <completion-boundary>
Never inspect or edit the SQLite file directly. Managed-project source and its toolchain are independent of the Workbench state location.
Recover before acting
At the beginning of work and after interruption, run:
agent-workbench --state <state-path> status
agent-workbench --state <state-path> next
status is read-only. next returns either one revision-bound executable
command with its constraints or one concrete blocker. Run the printed command;
do not reconstruct a different action.
A state initialized by an older or interrupted setup may return a
revision-bound start command:
agent-workbench --state <state-path> start \
<revision> <owner> <outcome> <completion-boundary>
An active work item returns continue. A resumable item returns resume.
Projection damage returns repair. Each command rechecks the current state and
rejects a stale action.
Mutations
Design, review, decomposition, evidence, and completion mutations use one typed JSON request per transaction:
agent-workbench --state <state-path> apply <request.json>
Every request contains a unique operation, the exact expectedRevision, a
supported command, and that command's fields. The native CLI parses the
request, while the authoritative Lean transition kernel decides acceptance.
The CLI does not duplicate policy.
Read request-format.md for the supported request shapes. Read native-workflow.md for the normal design-to-completion sequence.
Review authority
Review claims are advisory. The caller records a separate adjudication with a reason and remains responsible for adoption. Review observations do not become requirements merely because a reviewer raised them.
Validation
Record positive observations required by the current accepted design. Do not turn a rejected or removed approach into a permanent inverse requirement or an absence test. Evidence must retain its exact revision, producer, observation, repository snapshot, and artifact identity.
Errors
Unknown commands, malformed JSON, stale revisions, invalid transitions, and state corruption fail with a non-zero exit. Do not retry a changed request under the same operation identity. On an uncertain external effect, reconcile the remote observation before retrying.
What ships with it: 4 files
4.5 KB alongside SKILL.md, 2 of them executable
agents/
- openai.yaml258 B