Skip to content

config.md

Latest commit

 

History

History
104 lines (77 loc) · 3.21 KB

config.md

File metadata and controls

104 lines (77 loc) · 3.21 KB
Raw

Configuration

中文 | English

Default config file:

opencode-plusplus.config.yml

Local private config file:

opencode-plusplus.local.yml

Do not commit opencode-plusplus.local.yml.

Example

target: opencode
evidencePolicy: advisory
feedback:
  enabled: true
  telemetry: false
  network: false
  useLocalQualitySignals: false
tokenBudget: 60000

include:
  - "**/*"

exclude:
  - node_modules/**
  - dist/**
  - build/**
  - coverage/**
  - .next/**
  - .venv/**

tokenizer:
  mode: chars_approx
  # mode: cl100k_base
  # model: gpt-4.1

agents:
  mode: minimal
  maxTokens: 1200
  manualSources:
    - AGENTS.manual.md
  include:
    - commands
    - safety
    - entrypoints
    - contextLinks

llm:
  enabled: false
  provider: openai-compatible
  baseUrl: xx
  apiKey: xx
  model: xx
  temperature: 0.2
  maxTokens: 1200

rag:
  provider: lightrag
  chunkTokenLimit: 900

outputs:
  agents: true
  modules: true
  graph: true
  tasks: true
  readiness: true
  rag: true

Config files use tokenizer.mode: chars_approx; the CLI option --tokenizer chars-approx is accepted and normalized for command-line convenience.

Evidence Policy

evidencePolicy controls the minimum trusted evidence that can close test and contract requirements:

  • advisory preserves compatibility. Manual evidence can satisfy requirements, but reports mark it as a claim rather than system verification.
  • balanced accepts manual evidence for non-critical checks, but source or config changes require harness-captured command evidence or CI for tests. This is the recommended explicit mode for normal PR workflows.
  • strict requires current-working-tree command or CI evidence for both tests and contract validation. Manual evidence remains visible as supporting context but cannot close a blocking requirement.

The default remains advisory so existing configuration files keep their previous behavior. CLI commands that evaluate evidence accept --evidence-policy advisory|balanced|strict; MCP runtime tools accept the equivalent evidencePolicy field.

Context Feedback

feedback.enabled controls local Context quality feedback. It defaults to true; records contain only entry/source/version/revision/file identifiers and a fixed label. Task text, source content, absolute repository paths, and secrets are not stored.

feedback.telemetry and feedback.network both default to false. Network submission occurs only when both are explicitly enabled and feedback.endpoint is an HTTP(S) URL without embedded credentials. A network failure never removes the local record.

feedback.useLocalQualitySignals defaults to false. When enabled, local useful/not-useful statistics add a bounded, explainable ranking signal. The signal cannot override an exact Context ID match and never affects evidence, Guard, Policy, or finalize decisions.

Local annotations and maintainer feedback are separate: annotations are repository knowledge that may be explicitly injected as untrusted Context; feedback is quality metadata and is never injected into model context.

LLM Credentials

Committed examples should keep baseUrl, apiKey, and model as xx. Real credentials belong only in opencode-plusplus.local.yml.