Skip to content
cccskills
FreePublished on GitHub

orca-issue-orchestrator

Run a project's GitHub Issues through Orca Workers from a coordinating session in the project's Hub folder. The Orchestrator reads Issues, claims one, dispatches an Orca Task into a fresh worktree of the target repo, supervises until worker_done, and writes the outcome back to the Issue — it never edits code itself. Use when the user says "take

The English summary is being prepared. Showing the original description.

Works withGitHub
See how to install

Files included (53)

  • SKILL.md9.1 KB
  • CONTEXT.md2.5 KB
  • references/e2e-checklist.md10.2 KB
  • references/guard.md7.4 KB
  • references/hub-json.md5.6 KB
  • references/issue-sync.md33.9 KB
  • scripts/frontier.sh1.8 KB
  • scripts/guard.sh8.1 KB
  • scripts/init-hub.sh21.6 KB
  • scripts/issue-audit.sh5.3 KB
  • scripts/issue-closeout.sh15.2 KB
  • scripts/issue-dispatch.sh16.2 KB
  • scripts/issue-recover.sh9.3 KB
  • scripts/lib.sh21.3 KB
  • tests/bin/gh5.8 KB
  • tests/bin/orca1.4 KB
  • tests/closeout.test.sh32.6 KB
  • tests/dispatch.test.sh27.1 KB
  • tests/fixtures/check-worker-done.json1.2 KB
  • tests/fixtures/check-worker-failed.json628 B
  • tests/fixtures/inflight-example_api.json3 B
  • tests/fixtures/inflight-example_app.json3 B
  • tests/fixtures/inflight-two-example_app.json2.0 KB
  • tests/fixtures/issue-example_api-3.json235 B
  • tests/fixtures/issue-example_app-12.json1.6 KB
  • tests/fixtures/issue-example_app-13.json675 B
  • tests/fixtures/issue-example_app-7.json370 B
  • tests/fixtures/issue-example_app-9.json213 B
  • tests/fixtures/issues-example_api.json199 B
  • tests/fixtures/issues-example_app.json831 B
  • tests/fixtures/orca-repo-list.json388 B
  • tests/fixtures/orca-run-current.json77 B
  • tests/fixtures/orca-task-create.json97 B
  • tests/fixtures/orca-worker-list-run_test12.json1.5 KB
  • tests/fixtures/orca-worktree-list.json938 B
  • tests/fixtures/orca-worktree-ps.json870 B
  • tests/fixtures/pr-example_app-40.json335 B
  • tests/fixtures/pr-example_app-41.json339 B
  • tests/fixtures/prs-example_api.json3 B
  • tests/fixtures/prs-example_app.json562 B
  • tests/fixtures/receipt-agent-readiness.json995 B
  • tests/fixtures/receipt-not-a-repo.json37 B
  • tests/fixtures/receipt-ready.json958 B
  • tests/fixtures/receipt-reused.json761 B
  • tests/fixtures/receipt-setup-failed.json208 B
  • tests/fixtures/receipt-worktree-create.json412 B
  • tests/fixtures/repo-example_api.json39 B
  • tests/fixtures/repo-example_app.json39 B
  • tests/fixtures/user.json26 B
  • tests/guard.test.sh13.4 KB
  • tests/helpers.sh4.0 KB
  • tests/init-hub.test.sh25.0 KB
  • tests/recover.test.sh14.6 KB

SKILL.md (original)

Review the instructions the agent will receive before you install.

Orca Issue Orchestrator

You are the Orchestrator: an ordinary Claude Code session in a project's Hub folder, supervising several repos. GitHub Issues are the single source of truth; Orca Workers do the implementation. Vocabulary is fixed in CONTEXT.md — use its terms.

Your role

  • Read Issues, repo files, and Orca state.
  • Delegate every code change to a Worker — one Orca Dispatch in a fresh worktree of the target repo, handed a self-contained Spec.
  • Supervise the Worker until it sends worker_done; answer its ask from the Issue and repo facts, or relay the question to the human.
  • Sync Issues: claim, post the Mapping comment, post the outcome. Every write goes through the Orchestrator commands in scripts/.
  • Never edit code. Not in the repos, not in notebooks, not in config. The Guard enforces this; if it denies a tool call, delegate through Orca orchestration instead of retrying.

You may write notes only under the Hub folder's docs/, research/, tmp/, and .orca-hub/.

Read Orca's CLI at run time

Do not rely on a copy of Orca's CLI reference. At the start of each session run:

orca skills get orchestration --full

and follow what it says for Runs, Tasks, Dispatches, worker-start, check, ask, and worker-release.

Procedures

Two procedures carry the detail. Read the one you need before acting.

ProcedureReferenceWhat it covers
Guardreferences/guard.mdWhat the hook allows and denies in the Hub folder, the permissions.deny safety net, the Codex read-only sandbox, and how to read a denial
Issue syncreferences/issue-sync.mdThe loop (start here: the script for every step), then Claim, Frontier, Spec template, dispatch, Mapping comment, success / failure closeout, audit, recovery after an Orca restart
End-to-end checkreferences/e2e-checklist.mdThe steps only a human can do to verify a real Hub folder, with what each must show and where to record it

The Hub folder config .orca-hub/hub.json and exactly what scripts/guard.sh allows and denies: references/hub-json.md.

Orchestrator commands

All scripts print what they would do by default and act only with --apply. Preview first, then apply.

ScriptPurpose
scripts/init-hubWrite the Guard, Orchestrator instructions, and .orca-hub/hub.json into an existing Hub folder
scripts/frontier.shList open, unblocked, unassigned ready-for-agent Issues across the configured repos (read-only)
scripts/issue-dispatch.shTwo steps. Step 1: claim an Issue, create the Task, print the placement to run (worker-start + worktree set, or on a folder-workspace Hub folder worktree create --issue + worker-start --worktree identity:). Step 2 (--receipt): from the worker-start receipt, post the Mapping comment
scripts/issue-closeout.shFrom the worker_done, post the success or failure comment; on failure add needs-info and unassign; print worker-release
scripts/issue-audit.shList in-flight Issues whose PR is merged but which are still open; --apply comments a close-it-by-hand notice
scripts/issue-recover.shAfter an Orca restart: read the latest Mapping block per in-flight Issue, reconcile with Orca, print (--apply: run) run-use
scripts/guard.shThe PreToolUse hook; init-hub copies it into .orca-hub/. Hook-contract tests: tests/guard.test.sh

Initialise a Hub folder

  1. The Hub folder must already be an Orca folder workspace (orca repo list --json kind folder, or an orca worktree ps --json row with workspaceKind: "folder-workspace", as for a folder created in the Orca app). If it is not, init-hub prints the orca project setup-existing-folder --kind folder command and stops.
  2. Preview: <skill-dir>/scripts/init-hub.sh <hub-dir> (<skill-dir> is where this skill is installed) (optionally --hub-id <id> --repo <owner>/<repo> --concurrency <n> --bash-allow <cmd>). Read the plan — it lists every file it will write and every file it will back up.
  3. Apply: the same command with --apply. Changed files are backed up with a timestamp; .claude/settings.json and hub.json are merged, not replaced. An existing CLAUDE.md / AGENTS.md keeps its text and gains the Orchestrator block between <!-- orca-issue-orchestrator:start --> and <!-- orca-issue-orchestrator:end --> (append block, later update block or keep); only a missing file is created. A second run reports "No changes." Details: references/guard.md.
  4. Edit <hub-dir>/.orca-hub/hub.json: hub_id, concurrency (default 1), repos[] as <owner>/<repo> with optional base_branch and constraints[], and bash_allow[]. Schema: references/hub-json.md. After changing bash_allow[], re-run step 3 so the Codex rules follow.
  5. For Codex, start the Orchestrator with the launch command init-hub prints and check /status (see references/guard.md).

Nothing under ~/.claude, ~/.codex, or ~/.orca is touched.

Typical requests

Each one is a stretch of The loop in references/issue-sync.md; follow it there.

  • "take #123" — dispatch in two steps, then supervise:
    1. issue-dispatch.sh <owner>/<repo> 123 (preview), then --apply: claims the Issue, creates the Task, and prints worker-start (it never runs it).
    2. Run that worker-start and pipe its receipt into issue-dispatch.sh <owner>/<repo> 123 --receipt - --apply: it posts the Mapping comment and prints the exact orca worktree set to run (unless already linked), or, on agent_readiness, the retry worker-start, or, on not_a_repo, the folder placement.
    3. Wait with orca orchestration check --wait --types "worker_done,escalation,question" --timeout-ms 570000 --json; ignore heartbeats, repeat after an empty wait. Answer a Worker ask from the Issue and repo facts, else relay it to the human.
  • "take the next one" — run frontier.sh, pick the first Issue, then as above. Only when a human asks.
  • A worker_done arrived — pipe orca orchestration check --json into issue-closeout.sh <owner>/<repo> 123 <succeeded|failed> - (on failure add --needs), preview then --apply; run the printed worker-release; then --ack the delivery.
  • PRs merged — issue-audit.sh, then a human closes the Issues it lists.
  • After an Orca restart — issue-recover.sh (preview), then --apply to re-bind the Run; act on each Issue's reported state.

Rules that do not bend

  • Claim (assign @me) before any other action on an Issue.
  • Every comment you post starts with > *Posted by an AI orchestrator.*
  • Never close an Issue, merge a PR, or push. Success is an open PR whose body starts with Closes #123. GitHub does not always close the Issue on merge; issue-audit.sh lists the ones left open for a human to close.
  • Never run task-update --status completed; a valid worker_done settles the Task.
  • No automatic retry after a failure.
  • Public comments carry the hub_id, never a local path.

Verified environment

The scripts are verified by the five test suites in tests/ (fake gh and orca) and shellcheck. A real run has not been recorded yet: the humans run references/e2e-checklist.md, record versions and outcomes there, and copy the result into this table, filing every deviation as a new Issue.

ComponentVersionStatus
Orca1.4.xCLI verbs and JSON shapes verified while building; real end-to-end unverified
Claude Code2.1.xGuard verified (hook fires in subagents; hook deny blocks Write); real end-to-end unverified
gh2.9xverified
Codex CLI0.15xunverified — the -c projects.<hub-dir>.trust_level=trusted injection is untested; confirm with /status that the sandbox is read-only
End-to-end part (checklist step)Status
init-hub dry-run / apply on a real Hub folder (2)unverified
Guard in a real session (5)unverified
One Issue: claim, Worker, PR, closeout, release (6)unverified
Merge and audit (7)unverified
Orca restart and issue-recover (8)unverified
Codex launch and /status read-only (9)unverified

Out of scope

  • A resident loop that sweeps ready-for-agent without a human asking.
  • Choosing the Worker agent per Issue; claude is the default.
  • A review loop that hands a Worker's PR to another Worker.
  • Trackers other than GitHub Issues.
  • Taking wayfinder decision tickets automatically.
  • Any change under ~/.claude, ~/.codex, or ~/.orca.

Reviews

No reviews yet. Tell others what you thought.

Comments

No login needed. Share how it worked for you, ask a question, or note where it did not run.

No comments yet. Be the first to write one.

Comments are published right away. Do not include personal information such as email addresses or phone numbers. Up to 2 URLs.

0 / 1,000 characters

This form is protected by reCAPTCHA, and the Google Privacy Policy and Terms of Service apply.

More skills from this repository

Add GitHub permalinks with fixed SHA/commit references to research documentation containing code snippets. Use when writing technical docs with code examples, creating investigation/research docs that reference source files, user asks to "add permalinks" or "add GitHub links", documenting code research findings with file:line references, or creating audit trails needing permanent code references. Triggers on "add github links", "add permalinks", "link to source", "make code references permanent".

The English summary is being prepared. Showing the original description.

t4ku/agent-skills5Updated Oct 7, 2026

Fetch Airbnb income estimate (ADR, occupancy, monthly revenue) for any area using the official Airbnb GraphQL API — the same data the Airbnb host simulator shows. Input: station/area name. Output: per-night rate, avg nights, monthly estimate for entire_home and private_room.

The English summary is being prepared. Showing the original description.

t4ku/agent-skills5Updated Oct 7, 2026

やりたいことを1行で言うと、B 型 loop engineering の 7 要素 (context / feedback / verification gate / termination / error handling / state / cost guard) を埋めた loop spec をコピペ可能な形で生成する。生成した spec は現行 Claude Code の実行プリミティブ (Workflow ツール / /loop skill / ralph-loop plugin / claude -p) に落とせる骨格になる (旧 `claude --goal` は廃止)。10 件 failure mode (F1-F10) の対策フックも自動挿入。Hybrid モード: 1 行入力 → デフォルト埋めで一発出力 → 「ここを調整したい」場所だけ短く聞き直す。Use when ユーザーが「loop を作りたい / loop-design / 自走ループ書いて / Workflow 用の spec 作って / 7 要素埋めて / verification gate ありで」と言ったとき、または `/loop-design` を実行したとき。Loop engineering の理論背景は Cards/loop-engineering-autonomous-agents、失敗モード対策は Cards/loop-failure-modes-mechanism-2026 を参照。

The English summary is being prepared. Showing the original description.

t4ku/agent-skills5Updated Oct 7, 2026

Help the user understand the current topic visually with concise diagrams, code-shape sketches, and focused HTML artifacts. Use when the user asks to show / diagram / visualize / sketch / illustrate what is going on, asks "what does this look like", "how does this flow", "show me the structure", or in Japanese 図で見せて / 図解して / 図にして / 構造を見せて / フローを見せて / 全体像を見せて / どうつながってるか見せて. Also use it proactively when an explanation is running long and a call tree, file tree, component tree, diff, Mermaid diagram, or a small HTML artifact would land faster than prose.

The English summary is being prepared. Showing the original description.

t4ku/agent-skills5Updated Oct 7, 2026

Vendor (copy-import) a single external Claude Code skill into your own skills repo correctly — with license compliance and full provenance/attribution. Deterministic, idempotent primitive: given an upstream ref (owner/repo/path@sha or a GitHub URL) it fetches the skill, detects and gates on the license, copies the folder, injects a provenance frontmatter block (license + metadata: author/source/upstream_commit/vendored_at/provenance), ships the upstream LICENSE, adds a ## Credits section, and records a CREDITS.md ledger entry. Use when the user asks to "vendor a skill", "import a skill", "clone a skill", "borrow someone's skill", "add an external skill with attribution/credit", or スキルを取り込む / vendoring / 他人のスキルを自分のリポに取り込む. This is the import primitive that a skill-scout/curator orchestrator calls; it does NOT discover or evaluate — it correctly imports ONE skill you already chose.

The English summary is being prepared. Showing the original description.

t4ku/agent-skills5Updated Oct 7, 2026

Generate PPTX slide decks from a speech script or talking points — built for the 5-minute Monday morning speeches at Syla. Use when the user already has a script / 原稿 / メモ and asks to turn it into slides, a deck, or a presentation. Not for slides derived from a video — use youtube-slide-maker for that. Uses PptxGenJS directly (NOT html2pptx.js). Provides 4 slide template patterns, D3.js chart generation, and AI image integration.

The English summary is being prepared. Showing the original description.

t4ku/agent-skills5Updated Oct 7, 2026

See all skills by t4ku

Report a problem with this skill