Skip to content

forgegod/daidala

Repository files navigation

Daidala

Daidala

Daidala is a Hermes-native AI workshop that moves skill-backed work through interchangeable workflow packs and one explicit human approval gate—without introducing a second orchestration server.

Your daily driver for crafted, human-approved work.

Why Daidala

Daidala (pronounced DYE-dah-lah) is an Ancient Greek name for skillfully crafted or fashioned works. The word belongs to the tradition of Daedalus, the legendary maker, and the wondrous craft associated with Hephaestus.

The name fits a Hermes-native AI workshop built around disciplined craft rather than unconstrained automation. Daidala brings specialist agents and skills into an ordered process: a goal is defined, planned, approved by a human, implemented in isolation, verified, reviewed, and delivered with evidence. Skills provide the craft, workflow constraints shape the work, and Hermes supplies the agent runtime.

What Daidala adds to Hermes

Plain Hermes Kanban already owns cards, dependencies, profiles, retries, comments, and worker runs. Daidala adds pack-defined workflow policy around that runtime:

  • Workflow packs map exact skills onto define, plan, implement, verify, review, and deliver.
  • Provenance pins external skill sources, revisions, names, and complete directory digests.
  • Approval integrity binds human authorization to the SHA-256 digest of one complete plan revision.
  • Git safety rejects a dirty target and performs implementation in one Daidala-owned detached worktree.
  • Evidence retains definitions, plans, immutable diffs, changed paths, verification output, and review artifacts.
  • Conservative delivery reports a reviewed diff with committed: false and pushed: false.
  • Pack neutrality keeps pack-specific skill mappings in YAML rather than branching the Python engine.

How it integrates

Daidala loads in-process as a Hermes plugin. It creates linked cards on an existing Hermes Kanban board, assigns every executable stage to an explicit Hermes profile, and lets the gateway's existing Kanban dispatcher run ready cards. Daidala's SQLite data is only a policy and artifact ledger; Hermes Kanban remains lifecycle truth.

Daidala adds no MCP server, HTTP daemon, dashboard server, scheduler, model client, or nested hermes chat process. Its optional /daidala extension runs inside the existing Hermes dashboard; normal Kanban CLI, /kanban, and gateway operations remain available for progress and recovery.

flowchart LR
    S["explicit Daidala start"] --> D["define"]
    D --> P["plan"]
    P --> G["approval card<br>blocked"]
    G -->|"human approves exact plan digest"| I["implement"]
    I --> V["verify"]
    V --> R["review"]
    R --> DL["deliver<br>no commit or push"]

    H["Hermes Kanban<br>status, dispatch, retries"] --> D
    H --> P
    H --> I
    H --> V
    H --> R
    H --> DL
    W["Daidala policy ledger<br>digests and evidence"] -.-> G
Loading

Start a first workflow

Prerequisites:

  • Hermes Agent v0.18.2, the only verified host version;
  • Daidala installed and enabled in the profile that owns the workflow;
  • an existing named Kanban board;
  • the selected pack's exact skills installed in every assigned worker profile;
  • the Hermes gateway running so its Kanban dispatcher can claim ready cards;
  • a clean local Git target repository.
hermes plugins install forgegod/daidala --enable
hermes daidala doctor --pack aidlc
hermes kanban boards create project-board --name "Project board"
hermes gateway run

Run the gateway in a separate terminal on WSL. Then start explicitly with one profile for every stage:

hermes daidala start /absolute/path/to/repo "Implement the requested change" \
  --board project-board \
  --default-profile default \
  --pack aidlc \
  --workflow-id first-workflow

The command validates policy inputs and creates define → plan; it does not start another scheduler. Observe the board with hermes kanban --board project-board watch, the dashboard, or /kanban. After the plan card records a plan artifact, approve that exact digest:

hermes daidala approve first-workflow <64-character-plan-digest>

A generic hermes kanban unblock is not approval. Successful Daidala approval completes the gate and creates implement → verify → review → deliver in one persistent worktree. Use hermes daidala status first-workflow for combined policy facts and live card status; use normal Kanban comments, reassignment, and unblock for worker recovery.

See Getting started for the complete walkthrough, including pack setup, optional stage-specific profiles, recovery, and delivery.

Trigger and routing model

A workflow starts only through an explicit Daidala start action: the verified operator CLI above or an agent calling daidala_start. Cron is not required and is not part of Daidala's runtime. It may send a future prompt that asks an agent to perform the same explicit start, but Daidala owns no cron job, daemon, or polling loop.

The global Hermes kanban.orchestrator_profile limitation tracked in NousResearch/hermes-agent#34977 does not route Daidala stages. Daidala selects the board explicitly, assigns every executable card to an explicit profile, and creates the graph directly instead of asking Hermes goal decomposition to choose an orchestrator profile.

Support and limits

  • Supported host: Hermes Agent v0.18.2 on one local/single-host installation.
  • Supported entry points: native hermes daidala, standalone diagnostics, agent-facing plugin tools, and the optional Hermes dashboard extension.
  • Packs: Addyosmani agent-skills and the bundled AI-DLC v1.0.1 adapter.
  • Unattended runtime: the existing Hermes gateway Kanban dispatcher only.
  • Delivery never commits, pushes, deploys, or publishes without separate authorization.
  • Daidala does not copy secrets into artifacts and does not read or write the Hermes Kanban database.

Development and documentation

Start with the documentation index. Runtime claims and compatibility evidence are recorded in the Hermes integration guide; development commands and repository verification live in AGENTS.md. Release maintainers run both compatibility probes in scripts/; the release workflow enforces them for version tags and explicit manual dispatches.

python -m venv .venv
.venv/bin/pip install -e '.[dev]'
.venv/bin/lefthook install
.venv/bin/pytest
.venv/bin/ruff check .

License

MIT

About

A Hermes-native AI workshop for crafted, human-approved work through specialist agents and skills.

Topics

Resources

License

Stars

0 stars

Watchers

0 watching

Forks

Releases

No releases published

Packages

 
 
 

Contributors

Languages