agentsclimarketplace

Roboports

Skill dylanmccavitt/loom/skills/roboports

The implement coordinator. Runs one tracked issue end-to-end as code — one issue to one branch/worktree to one PR — with localized subagent fanout and a minimal diff, and covers behavior-preserving refactors and measured performance work through lenses. Use when the user asks to start, continue, or ship one tracked issue, refactor without changing behavior, or optimize a proven bottleneck.From its SKILL.md

Install
npx -y skills add dylanmccavitt/loom --skill roboports

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

  • 1 stars1 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.

SKILL.md

5.0 KB, ~1.0k tokens by cl100k_base, as published. Nobody here has run it

Roboports

Blueprint's issue-decomposition lens stamps the planned work; roboports coordinates the bounded build network that turns one ready tracked issue into landed code: one issue → one branch/worktree → one PR, no more. The main agent is the roboport hub — it owns intake, integration, and handing the PR to rocket-launch; subagents do bounded, disjoint, localized work.

Lenses

The input packet's lens field selects which variant guidance loads: a named lens loads references/lens-<name>.md; when lens is absent, load the default references/lens-issue-delivery.md. Unnamed lens references stay unloaded. Lenses select guidance only; they never widen packet scope, change the implement-mode boundary, or grant extra delegation authority.

  • issue-delivery (default) — one ready issue through branch, implementation, proof, review, and PR readiness.
  • refactor — behavior-preserving refactor: upgrade in place or delete/salvage dead and duplicated code while tests stay green.
  • performance — optimize a proven bottleneck with measured before/after results, stopping at diminishing returns.

Side-effect boundary: resolve the packet's context (validation | live) per the shared contract before any tracker, PR, or live-HOME action; under validation, report intended side effects instead of performing them.

The bridge

Planning lives in the tracker; code lands as a PR; the tracker's PR integration stitches the two. Preserve one issue to one branch/worktree to one PR unless the repo envelope says otherwise. The branch name carries the tracked issue id — the PR auto-links and, on merge, auto-closes the issue. Never craft a branch that drops the id.

Required reading

Before editing, read the repo envelope assembler generated (tracker team/project/label map, domain glossary, branch/PR conventions, build/test commands — never hardcode commands or a tracker), the active tracked issue (description, comments, acceptance criteria), and any architecture/ADR/domain docs the issue names.

Localized roboport discipline

Do not wire every subagent into one giant global network. Fan out in localized small batches: each subagent gets a bounded, disjoint write scope and a single lens. Default to the main agent implementing the first pass; spawn bots only when they cut risk or context load; give reviewers distinct lenses. Do not spawn separate "read issue", "implement", "review findings", and "fix findings" bots — that makes coordination the work. Localized small-batch fanout, never a universal backbone.

Doctrine and specialist routing

  • Apply the minimal-diff doctrine (the biters minimal-diff lens reviews against it): reuse before you write, ship the minimum that works, never cut validation/security/error-handling/accessibility.
  • Use tdd for test-first / red-green-refactor work.
  • For a bug, failing check, or regression: route to the operator-local debug-tools diagnose loop when installed (docs/skills/operator-local-manifest.md); otherwise run reproduce→minimise→hypothesise→instrument→fix→regression-test inline.
  • Route triage of incoming work to blueprint's triage lens; roboports only builds an already-tracked, ready issue.
  • Use the biters drift lens when repo/tracker/proof drift could change the route; it checks only.
  • Use lab for targeted proof of the implemented behavior (the lab ui-proof lens for user-visible flows) before launch gates rely on it.
  • Hand the finished PR to rocket-launch; roboports does not own closeout.

Flow

  1. Confirm scope, blockers, and the proof plan from the issue's acceptance criteria; create the one branch/worktree (id in the name) if absent.
  2. Implement only the acceptance criteria — nothing the issue did not ask for.
  3. Run the envelope's targeted checks that prove the changed behavior.
  4. Fan out localized reviewer bots when the change is non-trivial; fix real findings; rerun the relevant checks once across the union of changes.
  5. Prepare a review packet (the tracked issue id and acceptance criteria, changed files/diff, checks run and any unrun with reasons, exact questions per reviewer lens, blockers) and open or update the PR.

Invariants

  • The branch name carries the tracked issue id for the bridge.
  • Implements only the acceptance criteria; reads commands from the repo envelope.
  • Prepares the PR review-ready but does not own closeout (that is rocket-launch), and never silently closes the issue.

What ships with it: 11 files

16.1 KB alongside SKILL.md

evals/

exemplars/

Keep looking

Skills are one crate of 325,949. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.