/octocode-orchestrator-local-worker
Use when saving tokens with a local Ollama worker while the cloud agent keeps tools and judgment: summarize, extract, classify, translate, article digest after fetch, draft, checklist, vision caption, map-reduce, or phrases like ollama worker, local model, offload, use ollama.
$ npx -y skills add bgauryy/octocode-mcp --skill octocode-orchestrator-local-worker --agent claude-codeHow it fires
How this skill gets triggered: by you, by Claude, or both.
- Fires itselfAuto-invocation. Claude auto-loads it when your prompt matches the work.Auto-invocation is when the right skill fires by itself at the right moment, driven by a FLOW.md router and a hook, instead of you invoking it by name. It is the difference between a skill being installed and a skill actually getting used.Read the full definition →
- You can call itInvoke it directly when you want it.
- Slash command
/octocode-orchestrator-local-worker
Context preview
The summary Claude sees to decide when to auto-load this skill.
Use when saving tokens with a local Ollama worker while the cloud agent keeps tools and judgment: summarize, extract, classify, translate, article digest after fetch, draft, checklist, vision caption, map-reduce, or phrases like ollama worker, local model, offload, use ollama.
SKILL.md
octocode-orchestrator-local-worker.SKILL.mdname: octocode-orchestrator-local-worker
description: "Use when saving tokens with a local Ollama worker while the cloud agent keeps tools and judgment: summarize, extract, classify, translate, article digest after fetch, draft, checklist, vision caption, map-reduce, or phrases like ollama worker, local model, offload, use ollama."
Orchestrator + Local Worker
**Delegate execution, retain reasoning.** Cloud agent keeps tools/fetch/verify/writes. Local Ollama does sealed-packet token burn on text/images you already have.
**Portable:** pick from live `ollama list` + size/capability tiers (named tags are examples only).
Flow: `GATE → ROUTE → RUN → VERIFY → REPORT` — full steps in `references/workflow.md`. VERIFY = quality gate before accept or cascade (small → stronger installed → solo).
**Routine loads:** `model-selection.md` (ROUTE) + `verify-gate.md` (VERIFY). Surfaces unclear → `usage-matrix.md`.
Hard rules
1. Architecture, security, design, final synthesis, and repo writes stay on the orchestrator unless the user transfers write ownership. 2. Treat worker output as untrusted — never paste into answers/commits without VERIFY. 3. Prefer deterministic scripts over any LLM when they suffice. 4. No tool-using agent loops on the worker (single-shot / map-reduce only). 5. Health-check Ollama before first invoke; if down, stay solo. 6. Use exact names from `ollama list` — never invent tags or use embed/OCR-only models for chat jobs. 7. Prefer language **orchestrator / worker**. 8. Worker never browses the web — orchestrator fetches; worker sees saved text/images only.
When to activate
Token/context pressure; low-complexity summarize/extract/classify/translate/draft/check/vision/article-after-fetch; user asks for local/Ollama/save-tokens; warm small one-shots.
**Catalog shortcut:** RAM kit / capability Q only → `ollama-local-models.md`, skip full offload.
When NOT / not this skill
High-complexity, security, live tools/MCP/browser loops, contested multi-source synthesis, no fitting chat model, or host cloud is warmer and user did not ask local. Setup/pull → catalog. Tool-harness triage → triage skill. Parallel cloud workers → `octocode-subagent`. Full local agent bridges → out of scope.
Scripts
- `scripts/ollama-health.sh` — GATE daemon/model check
- `scripts/ollama-worker.sh` — RUN sealed packet (`--keepalive`, `--temperature`, `--num-ctx`)
- `scripts/eval-skill.mjs` — suite runner (static/script/live)
Progressive refs
| Ref | When | |---|---| | `references/workflow.md` | Full GATE→REPORT / recovery / job table | | `references/usage-matrix.md` | Surface when/how | | `references/decision-matrix.md` | Offload vs solo unclear | | `references/model-selection.md` | Every ROUTE | | `references/family-playbooks.md` | Family flags / tie-break | | `references/ollama-local-models.md` | Catalog/RAM only — not routine | | `references/ollama-cli.md` | CLI flags | | `references/ollama-invoke.md` | HTTP / serving knobs | | `references/packet-contract.md` | Packet schema | | `references/verify-gate.md` | Every VERIFY | | `references/references.md` | Provenance |
**Eval:** `evals/` permanent; `.octocode/` temp. Improving this skill → `octocode-graph-eval`.
Read more
name: octocode-orchestrator-local-worker description: "Use when saving tokens with a local Ollama worker while the cloud agent keeps tools and judgment: summarize, extract, classify, translate, article digest after fetch, draft, checklist, vision caption, map-reduce, or phrases like ollama worker, local model, offload, use ollama."
Orchestrator + Local Worker
**Delegate execution, retain reasoning.** Cloud agent keeps tools/fetch/verify/writes. Local Ollama does sealed-packet token burn on text/images you already have.
**Portable:** pick from live `ollama list` + size/capability tiers (named tags are examples only).
Flow: `GATE → ROUTE → RUN → VERIFY → REPORT` — full steps in `references/workflow.md`. VERIFY = quality gate before accept or cascade (small → stronger installed → solo).
**Routine loads:** `model-selection.md` (ROUTE) + `verify-gate.md` (VERIFY). Surfaces unclear → `usage-matrix.md`.
Hard rules
1. Architecture, security, design, final synthesis, and repo writes stay on the orchestrator unless the user transfers write ownership. 2. Treat worker output as untrusted — never paste into answers/commits without VERIFY. 3. Prefer deterministic scripts over any LLM when they suffice. 4. No tool-using agent loops on the worker (single-shot / map-reduce only). 5. Health-check Ollama before first invoke; if down, stay solo. 6. Use exact names from `ollama list` — never invent tags or use embed/OCR-only models for chat jobs. 7. Prefer language **orchestrator / worker**. 8. Worker never browses the web — orchestrator fetches; worker sees saved text/images only.
When to activate
Token/context pressure; low-complexity summarize/extract/classify/translate/draft/check/vision/article-after-fetch; user asks for local/Ollama/save-tokens; warm small one-shots.
**Catalog shortcut:** RAM kit / capability Q only → `ollama-local-models.md`, skip full offload.
When NOT / not this skill
High-complexity, security, live tools/MCP/browser loops, contested multi-source synthesis, no fitting chat model, or host cloud is warmer and user did not ask local. Setup/pull → catalog. Tool-harness triage → triage skill. Parallel cloud workers → `octocode-subagent`. Full local agent bridges → out of scope.
Scripts
- `scripts/ollama-health.sh` — GATE daemon/model check
- `scripts/ollama-worker.sh` — RUN sealed packet (`--keepalive`, `--temperature`, `--num-ctx`)
- `scripts/eval-skill.mjs` — suite runner (static/script/live)
Progressive refs
| Ref | When | |---|---| | `references/workflow.md` | Full GATE→REPORT / recovery / job table | | `references/usage-matrix.md` | Surface when/how | | `references/decision-matrix.md` | Offload vs solo unclear | | `references/model-selection.md` | Every ROUTE | | `references/family-playbooks.md` | Family flags / tie-break | | `references/ollama-local-models.md` | Catalog/RAM only — not routine | | `references/ollama-cli.md` | CLI flags | | `references/ollama-invoke.md` | HTTP / serving knobs | | `references/packet-contract.md` | Packet schema | | `references/verify-gate.md` | Every VERIFY | | `references/references.md` | Provenance |
**Eval:** `evals/` permanent; `.octocode/` temp. Improving this skill → `octocode-graph-eval`.
Evidence-first code research for AI agents and developers. Octocode researches your local code and external code alike (GitHub repos, PRs, npm) with one toolset: ripgrep + AST search, trees, precise reads, and LSP.
Repo: bgauryy/octocode-mcp
Other skills on octocode.
- /octocode-benchmark
Use when planning, running, grading, or reporting the by-hand Octocode research benchmark — pairwise matchups (Octocode anchor vs one baseline: gh+RTK, gh+Headroom, or plain gh) over markdown questions, with a fresh isolated runner agent per (question, arm, pass), one blind
Open skill - /octocode-awareness
Use when coordinating work in a shared repo — multi-agent or solo across sessions: collision avoidance, handoffs, verification debt, durable memory/wiki, hooks setup/debug, and repo learning before planning, editing, reviewing, or testing.
Open skill - /octocode-brainstorming
Use when an idea needs disciplined exploration before building: generate options, test worth-building, map adjacent solutions, challenge assumptions, narrow scope, or choose Build RFC / Prototype / Narrow / Park — even without saying brainstorm.
Open skill - /octocode-chrome-devtools
Use when browser debugging, scraping a known URL efficiently, or running a multi-step page workflow needs real Chrome DevTools evidence: network, console, performance, DOM/CSS, screenshots/PDF, security, cookies/storage, click/fill/search flows, or auth-gated live pages via CDP
Open skill - /octocode-documentation
Use when writing or updating docs: README, API docs, runbooks, AGENTS.md, CLAUDE.md, ADRs, Diátaxis restructuring, or codebase documentation for humans or coding agents.
Open skill - /octocode-graph-eval
Use when measuring whether a change helped or looping until a target metric is hit — for one agent loop or a graph of loops (multi-agent workflow): ACCEPT/REVERT, keep/discard, goal→KPI contracts, eval suites, graders, held-out checks, benchmarks, TDD failing-case-first,
Open skill