DailybotHQ

deepworkplan-addon-agentkit

Optional DeepWorkPlan addon that integrates coding-agents-kit (the `ak` command, DailybotHQ/coding-agents-kit pinned at v0.1.1, interface 1) as the headless delegation transport - a thin integrator that detects the kit through `ak doctor --json`, offers it (never imposes it) from onboard Phase 7b, documents the install from a clone of the pinned tag plus the kit's install.sh, records the acceptance in the .dwp/config.json addon registry, and maps a v7 plan's delegation operations (launch, observe, collect, cancel) onto one `ak run` per delegate in a dedicated git worktree. Never required, never a conformance gate; the pack never adds an autonomy or permission-bypass flag - autonomy stays the kit's explicit, per-run opt-in.

DailybotHQ 0 Updated 2w ago

Resources

3
GitHub

Install

npx skillscat add dailybothq/deepworkplan-vim/deepworkplan-addon-agentkit

Install via the SkillsCat registry.

SKILL.md

DeepWorkPlan — agentkit Addon (headless delegation transport)

Integrate coding-agents-kit — one command surface (ak) for every
terminal coding agent (claude, codex, cursor, opencode, pi, cline, grok and
their provider variants, multiple accounts through profiles) — as the
headless transport a v7 plan may use to delegate a parallel_safe
task. This is an opt-in addon, never required for a repo to be
AI-first and never a conformance gate: without it every task runs in the
current session exactly as before.

This addon is a thin integrator. The kit is its own product
(https://github.com/DailybotHQ/coding-agents-kit, MIT, works without
DWP); everything this addon relies on is the kit's frozen interface 1
(ak run, ak env, ak doctor --json, the exit codes) at the pinned tag.

Pin Value
Product DailybotHQ/coding-agents-kit
Tag v0.1.1 (ecosystem amendment A2 — supersedes v0.1.0)
Interface 1 (ak doctor --json → "interface": 1)
Registry key agentkit (.dwp/config.json → addons.agentkit)
Transport headless — provides subagents, cancel_children, model_routing; requires the agent_delegation grant

Read these first (all relative inside the skill)

  • `SPEC.md` — the normative contract: detection, offer,
    install, the transport mapping, the never-a-bypass-flag rule.
  • `templates/INTEGRATION.md` — reasoning
    guidance for one delegate end to end (worktree, prompt, run, collect).
  • ../../execute/delegation.md — when a plan may delegate at all.

When this runs

  • onboard Phase 7b offers it as an explicit opt-in when the developer
    wants a plan to hand bounded tasks to other coding agents.
  • execute uses it only through delegation.md, only on a v7 plan whose
    contract grants agent_delegation, only for a parallel_safe task (or a
    read-only delegate), and only when resources.py abilities shows
    addon:agentkit as a subagents source.
  • Direct invocation, to install or check the kit.

Trust boundary (write scope)

allowed-tools includes write-capable Edit, Write, and Bash. The
write scope is bounded:

  • Before consent: read-only. Detection is command -v ak and
    ak doctor --json (key names only, never values).
  • Writes (only after explicit acceptance): the kit install through its
    own pinned installer, the addons.agentkit registry entry, and — during a
    delegation — a dedicated git worktree for the delegate plus the collected
    result under the plan's analysis_results/delegations/<id>/.
  • It MUST NOT: add a permission-bypass or autonomy flag on its own
    (ak run without --auto is the default; --auto only on the
    developer's explicit per-plan opt-in, recorded, and only inside an
    isolated worktree or container); read, print or write any provider key
    value; let a delegate write outside its worktree; install a coding-agent
    CLI unasked (ak install is the developer's call); run a delegate the
    ledger refused; or treat a delegate's result as evidence.

The flow

Step 0 — Detect (read-only)

command -v ak → ak doctor --json → read interface (must be 1;
anything else is one warning and "not available"), the installed kinds and
whether each is logged in (file presence only), and the profile names. No
state installs anything.

Step 1 — Offer (never impose)

Explain what it adds — a plan may hand a bounded parallel_safe task to
another coding agent, in its own worktree, and verify the result with its
own gates — and what it costs: a machine-level install, provider accounts
you already have, and per-plan consent (agent_delegation). Declining is a
complete answer.

Step 2 — Install (pinned; point-don't-run by default)

git clone --branch v0.1.1 https://github.com/DailybotHQ/coding-agents-kit
./coding-agents-kit/install.sh

(Windows: the same tagged clone, then .\coding-agents-kit\install.ps1.)
The installer touches no network, installs no CLI and never overwrites the
env file. Run it yourself only interactively and with explicit acceptance;
re-detect afterwards.

Step 3 — Record

python3 ../../shared/config.py enable agentkit --version v0.1.1 --repo <repo>.
Delegation still needs each plan's contract to grant agent_delegation —
enabling the addon grants nothing.

Step 4 — Transport (during execute)

One delegate = one ak run in one dedicated worktree
(templates/INTEGRATION.md):

Operation Mapping
launch git worktree add <wt> -b dwp/<plan>/<delegation_id>; record ledger.py delegate launch (with prompt_digest); start ak run <kind> [@profile] --cwd <wt> --timeout <s> --output-format json -- "<prompt>" in the background, stdout to the result file
observe ledger.py delegate observe; the background process is alive or has exited
collect on exit read the one JSON object (exit 0 → completed, otherwise failed); ledger.py delegate collect with result_path; integrate the worktree's diff; run the task's gates here
cancel send SIGTERM to ak run (the kit kills the whole process tree, exit 5); ledger.py delegate cancel

Failure-mode guardrails

  • Never required, never blocking. Not installed, not logged in, unknown
    interface, refused by the ledger: run the task in this session.
  • The result is a claim. result_text is what the agent said; only the
    plan's own gate runner turns the work into observed evidence.
  • Secrets stay names. Provider keys are environment variable names; the
    kit never prints values and neither does this addon.

Validation checklist (component 4 — mirrored from SPEC §8)

  1. SKILL.md, SPEC.md, addon.json, templates/INTEGRATION.md exist;
    the descriptor pins DailybotHQ/coding-agents-kit v0.1.1, interface 1,
    transport headless.
  2. Detection is read-only and reports interface 1, or one warning.
  3. Install used the pinned tagged clone + install.sh; nothing piped.
  4. The registry entry exists only after acceptance.
  5. Every delegate ran in its own worktree, without --auto unless the
    developer's per-plan opt-in is recorded, and its result was gated here.