local-agency
Use when the adopter asks to delegate to local OSS agents (OpenCode) on a READY local runtime.
- License 许可
- MIT
- 最近更新
- 2026-10-20
- Stars
- ★ 0
- Forks
- 0
01 前言 Front Matter
- name
- local-agency
- license
- MIT
- description
- Use when the adopter asks to delegate to local OSS agents (OpenCode) on a READY local runtime.
02 资源文件 13
- 参考文档references/brief-and-preflight.md1.1 KB
- 参考文档references/coach-loop.md6.5 KB
- 参考文档references/design-turn-contract.md1.8 KB
- 参考文档references/mechanical-dispatch.md3.0 KB
- 参考文档references/parent-rog-shell.md1.6 KB
- 参考文档references/when-to-use-local.md2.0 KB
- 脚本scripts/assert_parent_rog_shell.py2.7 KB
- 脚本scripts/design_turn_gate.py3.8 KB
- 脚本scripts/dispatch.py25.8 KB
- 脚本scripts/executor_brief.py4.4 KB
- 脚本scripts/require_rog_freetoken.py7.7 KB
- 脚本scripts/tests/test_assert_parent_rog_shell.py1.2 KB
- 脚本scripts/tip_preflight.py9.7 KB
03 SKILL.md 正文
Local agency (OpenCode)
Optional local coding agency path. Not a workflow owner; select the canonical workflow in execution workflows first.
Local agency vs session agents
| Path | When | Owner |
|---|---|---|
| Session subagents / Task | Default orchestrator labor on the host | Host adapter + delegation |
| 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, then Colibri (:8000) via colibri (frontier MoE; prefer FreeToken for coding agency). - Configure OpenCode with ephemeral
OPENCODE_CONFIG/OPENCODE_CONFIG_CONTENTthat setsenabled_providersto the READY local provider. OpenCode merges global config; the allowlist keeps other providers out of that process.--pureonly disables external plugins. - Never write into durable
~/.config/opencode(helper refuses those paths), neverft 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 freetokenandrequire_rog_freetoken.pyfail closed unless the hostname looks like ROG, cwd is the operator ROG checkout (see local-agency guide (repo-onlychaos-engine/guides/local-agency.md)), orCE_ALLOW_BOX_LOCAL_AGENCY=1. Process-owner Shell withmachineIdon ROG is mandatory for ROG FreeToken/OpenCode writers. Do not dispatch those writers via Task until Grok Bot exposesmachineIdto Task/executor Shell. Diagnostic: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. ROG gate: 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.
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
(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. 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.
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.
Folded: local-coding-delegate
local-coding-delegate is a compat shim: hardware probe stays there; local OpenCode / agency routing lives here. Prefer this skill when the adopter names OpenCode or “local agents”.
Related
- Guide: local-agency.md (repo-only
chaos-engine/guides/local-agency.md) - FreeToken: freetoken
- Local OpenAI-compat: local-openai-compat
- OmniRoute (explicit cloud only): omniroute
- Identity push-back: identity-push-back.md
- Coach loop: coach-loop.md
- When to use local: when-to-use-local.md
- Design-turn contract: design-turn-contract.md
- Design-turn gate:
design_turn_gate.py - Parent ROG Shell playbook: parent-rog-shell.md
- Parent ROG Shell assert smoke: test_assert_parent_rog_shell.py
04 同仓库更多 Skills
ponytail
Lazy senior dev mode: the smallest change that fully solves the task, and a reply a busy human understands in one read. Use on any coding task (writing, fixing, refactoring, reviewing, choosing dependencies) and when the user says "ponytail", "be lazy", "simplest solution", "yagni", or complains about over-engineering or bloat. Levels: lite, full (default), ultra.
work-item
Use when opening or rewriting a work item on any git-based SCM. Source-control agnostic; GitHub, GitLab, and Azure Boards are adapters only.
self-improve
Use when running ChaosEngine Learning Session self-improve. Harness lessons are GitHub issues only. Product lessons may queue after delivery or on request.
omniroute
Use when local-runtimes selected the OmniRoute process for bounded work: runner, delegate continuity, capability enforcement and proof of dispatch.
local-runtimes
Use when choosing an optional local inference runtime for a narrow mechanical or offline job.
local-openai-compat
Loopback OpenAI-compatible peers (Ollama, LM Studio, llamacpp on 127.0.0.1). Use when a narrow task should go to an already READY loopback server.