/local-agency

name: local-agency
description: >-
  Use when the adopter asks to delegate to local OSS agents (OpenCode) on a
  READY local runtime.
license: MIT

Local agency (OpenCode)

Optional **local coding agency** path. Not a workflow owner; select the canonical workflow in [execution workflows](../../references/execution-workflows.md) first.

**Local agency vs session agents**

| Path | When | Owner | | --- | --- | --- | | Session subagents / Task | Default orchestrator labor on the host | Host adapter + [delegation](../../references/delegation.md) | | Local agency (this skill) | Adopter asked for OpenCode / local OSS agency against local weights | This skill + READY runtime |

Hard rails (never regress)

• Prefer a **READY** local runtime: FreeToken (`:1919`), then Ollama / LM Studio /

llamacpp via [local-openai-compat](../local-runtimes/references/local-openai-compat.md), then Colibri (`:8000`) via [colibri](../local-runtimes/references/colibri.md) (frontier MoE; prefer FreeToken for coding agency).

• Configure OpenCode with **ephemeral** `OPENCODE_CONFIG` /

`OPENCODE_CONFIG_CONTENT` that sets `enabled_providers` to the READY local provider. OpenCode merges global config; the allowlist keeps other providers out of that process. `--pure` only disables external plugins.

• **Never** write into durable `~/.config/opencode` (helper refuses those paths),

**never** `ft launch` / `ft serve`, **never** start local servers from CE.

• Do **not** silently fall back to cloud OmniRoute. Missing local runtime:

tell the operator; continue with session agents / `SOLO` / OmniRoute only when the adopter explicitly asked for that path.

• Do **not** install OpenCode from this skill. Operator install stays on vendor

docs (README third-party table).

• Installer / doctor / status must **not** fail because OpenCode or a local

runtime is missing.

• Never bind or probe a non-loopback URL from ChaosEngine.
• **ROG FreeToken bind:** Task/box writers must **not** claim FreeToken.

`dispatch.py resolve --prefer freetoken` and [`require_rog_freetoken.py`](scripts/require_rog_freetoken.py) fail closed unless the hostname looks like ROG, cwd is the operator ROG checkout (see local-agency guide (repo-only `chaos-engine/guides/local-agency.md`)), or `CE_ALLOW_BOX_LOCAL_AGENCY=1`. Process-owner Shell with `machineId` on ROG is **mandatory** for ROG FreeToken/OpenCode writers. Do **not** dispatch those writers via Task until Grok Bot exposes `machineId` to Task/executor Shell. Diagnostic: [`assert_parent_rog_shell.py`](scripts/assert_parent_rog_shell.py). Clear error when FreeToken is not READY on this host.

Probe → resolve → ephemeral OpenCode

Helper: [`chaos-engine/skills/local-agency/scripts/dispatch.py`](scripts/dispatch.py). ROG gate: [`require_rog_freetoken.py`](scripts/require_rog_freetoken.py).

python3 .chaos-engine/skills/local-agency/scripts/dispatch.py resolve
python3 .chaos-engine/skills/local-agency/scripts/dispatch.py --prefer freetoken config
python3 .chaos-engine/skills/local-agency/scripts/dispatch.py argv --prompt '…' --workdir '<worktree>'

`resolve` ranks FreeToken, then OpenAI-compat peers, then Colibri. Use `--prefer colibri` when the adopter asked for Colibri. On `READY`, `config` / `argv` emit ephemeral OpenCode material (`OPENCODE_CONFIG` path). Run OpenCode yourself with that env; do not persist the config into the durable user file. `argv` defaults include `--pure` and `--variant` (`medium`; use `low` or `medium` for tool loops). `--no-pure` exists for callers that already isolate plugins. `--pure` still only disables plugins.

Design-turn brief and OpenCode preflight

On demand: [brief and preflight](references/brief-and-preflight.md).

Dispatch modes

`--mode mechanical` (**default** for small local coding models / 7B tool loops): treat the prompt as one bounded apply command; do not auto-attach CE brief (still allowed via `--with-ce-brief`).

`--mode design`: attach the locator-only CE brief (same as `--with-ce-brief`) for richer design/spec turns. Prefer this only when the host coach wants a design draft, not for mechanical apply.sh loops.

Design/spec turns must follow [design-turn-contract.md](references/design-turn-contract.md) (verbatim `CE_BRIEF_LOCATORS:` closing line; host reject gates).

Parent parser uses `allow_abbrev=False` so `--mode` cannot abbreviate to `--model` . Keep that guard when adding short overlapping flags.

python3 .chaos-engine/skills/local-agency/scripts/dispatch.py --mode mechanical argv --prompt 'bash <abs-dir>/apply.sh'
python3 .chaos-engine/skills/local-agency/scripts/dispatch.py --mode design argv --prompt '…' --project .

Mechanical dispatch

A READY local coder is a mechanical runner. Procedure: [mechanical-dispatch.md](references/mechanical-dispatch.md). The OpenCode bash timeout is **120s**: after a timeout kill of an apply script, re-run the same idempotent script from the orchestrator so product files are not left half-applied.

Session token usage

When `dispatch.py chat` returns OpenAI-compat `usage` and `--session-id` is set, dispatch records `session_token_usage.py` with `--channel local` and a coarse `--runtime-class` (`freetoken` / `openai-compat` / `colibri`). Never write model or provider ids into the ledger. `brief` / `config` / `argv` do not call the model, so they do not record usage.

When to use local

Implementers write code directly; this skill is optional. Use local only when the handoff cost is well below generation cost (bulk mechanical edits, spec or ticket drafting, log summarization, offline or private work). See [when-to-use-local.md](references/when-to-use-local.md).

Coach loop (when local is chosen)

A chosen local writer needs an active coach: verify every artifact, give grounded feedback, and never let it finish unsupervised. Playbook: [coach-loop.md](references/coach-loop.md).

Folded: local-coding-delegate

[local-coding-delegate](..