Dobeu-tech-eco

dt-onboard

"Explain the dts2 harness (eight agents, twenty-eight skills, roster, team tiers, lessons, headless runner) and recommend where to start. Use when the user asks what dts2 does, how to use it, which skill fits a job, or is new to the harness."

Dobeu-tech-eco 0 Updated 2d ago
GitHub

Install

npx skillscat add dobeu-tech-eco/dts2-marketplace/dt-onboard

Install via the SkillsCat registry.

SKILL.md

Cursor procedure

Find the plugin root: the nearest ancestor of this file that contains .cursor-plugin/plugin.json. The examples below say $PLUGIN for that directory.

Run python3 "$PLUGIN/scripts/dt_gate.py" for every gate. Exit 2 from scaffold, preflight, or approve stops the command. Exit 2 from bridge that prints BLOCKED unsupported-runtime-script is a precise blocker for that one step: record it and do not imitate the script. The interview, approve, and scaffold still run.

--yes, --ship, and --scaffold-only are rejected. A JSON key with the same name as a kind is not authorization. approve records the sha256 of the file the user accepted, and later gates require that digest.

Phase files for this command are in $PLUGIN/skills/dt-onboard/.

Merges, fleet broadcast, hook install, the headless SDK, and paid provider calls are CURSOR_UNSUPPORTED even after a permission phrase is recorded.

Cursor edition: run scripts/dt_gate.py from this plugin for every gate. Do not read tokens or credential files. Do not merge, broadcast, install hooks, or spend a provider API. Carry out file work with Read, Edit, Grep, Glob, and git.

dts2 harness — orientation

This skill explains only, except the skill-parity step and the optional hooks step. Skill parity updates agent backends without asking. Hooks change files only after the user picks them. Do not invoke agents.

Agents (all prefixed dt:)

  • lead plans a DAG of tasks and gives each builder a separate file set.
  • builder implements one builder section; several builders can run at once on separate file sets.
  • reviewer returns severity-rated findings on one axis or all axes. Never edits.
  • verifier runs configured commands and checks for stubs. Never edits.
  • scout sizes a repo or workspace and recommends a team tier.
  • auditor inventories this machine's agent tooling against a blueprint profile.

Skills

Skill Use it to
machine-prep Scan the computer, prune plugins to the blueprint keep-set, set up the slim startup. Once per machine.
setup Interview to an approved end state, then scaffold .harness/ in one repo, many repos, or a multi-repo workspace.
run Do a task: scout sizes it, lead plans, N builders build, reviewers and verifier check. Two checkpoints: Plan checkpoint and Ship checkpoint. Open secrets and decisions go to .harness/action-items.md; close a secret with run add secret NAME.
learn Turn reviewer findings into lessons that every future run reads.
sdk Run the same pipeline headless with the Agent SDK, using your existing Claude login, or an API key or key command for CI.
evolve Change the harness itself in its source repo: edit, validate, test, bump both manifests, commit; push only on an explicit yes.
action-items List open secrets, decisions, handoffs, deploy steps and promotions with the command that resolves each; done, add, assign. Pass-off for any agent: .harness/handoff/latest.md, resumed with run resume.
gtm Go-to-market in one seamless pass: auto-installs marketing-skills if missing, runs a deep-dive, picks the applicable skills, and ends in a readiness.md scorecard.
ship /dt-ship opens a PR on your own initiative: pre-check, type/lint/test/build, secret and do-not-touch scan, verifier and reviewer, receipt, PR template, push and open; then sweeps PRs and branches so only main and dev remain. Never merges.
feature /dt-feature <idea> adds a feature to the roadmap: clarity score, a scoped deep interview when it is unclear, the lead writes it into the end state and every plan doc, then a PR.
autopilot /dt-autopilot <steps> runs steps unattended: three frontier providers must agree on the plan and the build (the latest frontier Claude Chairman breaks a split), one PR per step merged on green CI before the next. Progress shows in check-in.
dpp /dt-dpp [dry-run] runs up to five autopilot lanes in parallel as headless sons under one head: lane plan and feasibility from the check-in list, approval, supervision, serialised merges, one report and the task ledger.
dpp-full /dt-dpp-full is dpp with --full: delegation scan for extra lanes. Outside autopilot it asks before each new lane while dpp_full_ask_each_lane is true (the default).
fullauto /dt-fullauto is dpp-full with autopilot on, looping check-in between rounds until nothing is outstanding and every lane is done or blocked in the ledger.
model /dt-model picks which agent provider leads (the lead, advisory and sign-off seats); default Claude. Builders are unaffected.
find-skills /dt-find-skills is available any time: rescan the repo, answer one to three questions, get suggested skills, and wire the ones you pick into the project.
ask-expert /dt-ask-expert [repo] is available any time, before setup: tell it your goal, and an expert panel reads the code and writes a one-page memo: keep, fold into another repo, sell as a template, archive or delete.
repo-sweep /dt-repo-sweep [owner] does the same quick pass over all your repos, oldest first, then one cleanup screen. Archive is reversible; delete is one repo at a time with a backup and a typed name.
check-in /dt-check-in is available any time: a read-only checklist from where the project started through the approved end state, each item marked complete or outstanding, with an N of M complete percent.

Setup and run pull in companion skills from mattpocock/skills at the moment they are needed (grill-me, improve-codebase-architecture, writing-for-agents, pr, resolve-merge-conflicts (from warpdotdev/common-skills, since mattpocock retired its version), plus ask-matt and research for your questions, improve-codebase-architecture again before tests, diagnosing-bugs again at the hard 5-round fix cap, loop-me when a run starts, setup-pre-commit once per repo at the commit gate, and code-review on first contact with a new repo), recorded in keep.yaml. A pre-check before every run command installs any that are missing, proposes updates for you to approve (it never updates them in bulk), swaps one that is gone upstream, refreshes plugin marketplaces, and when a newer harness is published saves your spot and upgrades; the next run finishes the upgrade itself (there is no separate upgrade command). After every ship, open harness lessons are offered through evolve automatically.

Setup maps the whole project into sections; run keeps builders busy on ready sections and parks stuck ones instead of stopping. The run skill pivots to the workspace root when started inside a member repo, and a context guard stops it and re-runs setup on a context spike.

How work is organized

The frontier tier is for advice and sign-off only (the lead, the reviewers' final review, the verifier). Builders never run a frontier model: they run the cheapest capable model at a complexity-scaled effort (scripts/delegate.py), each in its own worktree. A stalled lead hands off and fails over to the next eligible provider, and every detected agent gets a seeded project memory file at scaffold. One run branch per task; you push and open the PR with the receipt. run review-pr <n> reviews an existing PR.

Every run checks which agents are usable and what budget is left (the budget check), then assigns work by cost tier, cheapest capable agent first.

Agent roster (profiles/agents/)

Agent Invocation Typical role
claude live lead, verifier, builder, reviewer
codex live builder, second-opinion reviewer
gemini live builder, reviewer, second-opinion reviewer
grok live builder, reviewer, second-opinion reviewer
opencode live builder
muse live builder, second-opinion reviewer
copilot live builder, security reviewer
cursor handoff builder
antigravity handoff builder

Only agents detected on this machine join the roster; handoff agents get a written brief, and the run pauses until you report it done.

Built-in API agents claude-api, codex-api and grok-api need only a key (docs/API-AGENTS.md). Another agent on a pay-as-you-go API key is added with /dt-model, section "Add or rename an agent"; keys live in your keyring, 1Password or Doppler and only the reference is written. A new API-key agent starts read-only (review, docs and research only, no deploy, merge, release, secrets or delete, $0.50 a run, $5 a day); the same section shows and edits these per-agent rules.

Skill parity

After the agent list above, run python3 "$PLUGIN/scripts/dt_gate.py" bridge roster.py detect --project . and pass every installed name to python3 "$PLUGIN/scripts/dt_gate.py" bridge lead_skills.py gaps --agents <names> --home "$HOME". For each entry in behind, say one line: <agent> is behind the orchestrator: <reason>. Then run lead_skills.py sync --agents <those names> --home "$HOME". The sync is silent: do not ask. A non-lead may show an older /dts2-harness until this step, or until /dt-model makes it the lead. Claude is the source and is skipped.

Where to start

  1. New machine or cluttered tooling → machine-prep.
  2. New repo or repos → setup.
  3. Any multi-file task → run.
  4. After a few runs → learn.
  5. CI, cron, or a second machine without the desktop app → sdk.
  6. Changing the harness itself → evolve. Launching a product → gtm.
  7. Optional Claude Code hooks (command guard, auto-format, tests on stop, notifications) → the Hooks step below.
  8. Many app plugins or MCP servers → composio (one connection, then trim overlap).
  9. Database tooling: /dt-find-skills suggests database MCP servers and skills for your stack.
  10. Not sure a repo is worth keeping → /dt-ask-expert for one repo, /dt-repo-sweep for many.

Offer ask-expert when the user doubts a repo

When the user's answers show doubt about keeping a repo (it is old, unused, abandoned, or they are thinking of deleting or archiving it), suggest /dt-ask-expert before setup, in one sentence. For many such repos, suggest /dt-repo-sweep. Only suggest; do not start either.

Tiers the run skill picks automatically

Tier Typical size Builders Reviewers
S under 20k LOC, single package 1 1 (all axes)
M under 150k LOC or a small monorepo up to 2 correctness, security
L under 600k LOC up to 4 + performance
XL larger up to 4 + style

Worktrees are used only in workspace mode or when use_worktrees: always is set; single mode shares the tree on separate file sets.

Builders never exceed the number of builder sections the lead planned, and never exceed max_builders in .claude/dts2-harness.local.md.

Composio (optional)

Ask once with AskUserQuestion: "Do you have a Composio account?" (yes or no). Composio is one connection to 1000+ apps, so agents need fewer separate plugins and MCP servers. On either answer, hand off to /dt-composio, which records the answer, installs and maps on yes, and offers a signup link on no. Skip the question if ~/.dts2/composio.json already holds a decision. Nothing is changed here; /dt-composio asks before it disables anything.

Hooks (optional last step)

The plugin's own hooks are already active: the policy gate, the merge guard, the stop checks and the context guard. Toggles are in .claude/dts2-harness.local.md. This step offers extra hooks for the user's project.

  1. Run python3 "$PLUGIN/scripts/dt_gate.py" bridge hooks-setup.py recommend.
  2. Show the recommended hooks as a short table: hook, when it runs, why, already present.
  3. Ask once with AskUserQuestion (multi-select) which to add, and the scope: project (shared, the default), local, or user.
  4. On picks, run hooks-setup.py apply --ids ... --scope ... --dry-run, show the result, then run apply for real.
  5. Report what was added, what was skipped and the backup path.

When a database is detected, db-guard (blocks destructive database commands) and migration-check (reminds you to review migrations) are recommended too. The harness never runs destructive database actions; it always asks.

Existing hooks are never overwritten or removed. Nothing is written without a pick. /hooks in Claude Code shows the result.

Keep the reply under 200 words unless the user asks for detail. Single-file fixes do not need the harness.