DRS Knowledge Map

Appearance

DRS configuration ​

DRS merges configuration from several sources, with later sources taking precedence. The loader is src/lib/config.ts.

Configuration sources ​

  1. Hard-coded defaults in src/lib/config.ts.
  2. Packaged workflow files from .pi/workflows/*.yaml.
  3. Project workflow files from .drs/workflows/*.yaml.
  4. .drs/drs.config.yaml.
  5. Legacy .gitlab-review.yml alias.
  6. Environment variables.
  7. CLI overrides.

The file .drs/drs.config.yaml is the canonical project configuration. Inline workflows: maps at the top level of the config file are rejected and must be moved to .drs/workflows/*.yaml.

Required settings ​

At least one review agent must be configured and a default model must be resolvable. The CLI will throw if these are missing:

yaml
agents:
  default:
    model: opencode-go/kimi-k2.7-code

review:
  agents:
    - review/unified-reviewer

The default model can also be set with DRS_DEFAULT_MODEL.

Pi runtime and providers ​

Custom providers are configured under pi.provider:

yaml
pi:
  provider:
    opencode-go:
      apiKey: OPENCODE_API_KEY
      api: openai-completions
      baseUrl: https://opencode.ai/zen/go/v1
      models:
        - id: kimi-k2.7-code
          name: Kimi K2.7 Code
          contextWindow: 262144
          maxTokens: 262144
          cost:
            input: 0.95
            output: 4

Model metadata is used for context compression and cost fallback. Provider-wide compat settings apply to every model unless overridden per model.

Runtime timeouts are under pi.runtime:

yaml
pi:
  runtime:
    operationTimeoutMs: 300000
    streamTimeoutMs: 900000
    streamPollIntervalMs: 2000

Environment overrides exist for each field: DRS_RUNTIME_OPERATION_TIMEOUT_MS, DRS_RUNTIME_STREAM_TIMEOUT_MS, DRS_RUNTIME_STREAM_POLL_INTERVAL_MS.

Provider retry is configured under pi.retry.provider:

yaml
pi:
  retry:
    provider:
      timeoutMs: 45000
      maxRetries: 2
      maxRetryDelayMs: 15000

Agent configuration ​

yaml
agents:
  default:
    model: opencode-go/glm-5.1
    skills: [cli-testing]
    thinkingLevel: medium
  namespaces:
    review:
      model: opencode-go/minimax-m3
    visual:
      model: opencode-go/kimi-k2.7-code
    task:
      model: opencode-go/kimi-k2.7-code
  overrides:
    task/docs-updater:
      run:
        prompt: "Summarize the latest API changes"
        output: .drs/docs-summary.json
        json: true

Resolution order for an agent model:

  1. CLI --model / per-agent env var.
  2. agents.overrides.<id>.model.
  3. agents.namespaces.<namespace>.model.
  4. agents.default.model.

The run override can supply a prompt, prompt file, output path, and JSON mode.

Review configuration ​

yaml
review:
  agents:
    - review/unified-reviewer
  ignorePatterns:
    - "*.test.ts"
    - "*.spec.ts"
    - "package-lock.json"
  describe:
    enabled: true
  cursorFixLinks:
    enabled: false

ignorePatterns supports * and ** globbing. describe.enabled runs the describe pass before review. cursorFixLinks adds deep links to the Cursor editor in summary comments.

Describe configuration ​

yaml
describe:
  model: opencode-go/kimi-k2.7-code
  includeProjectContext: true

Context compression ​

yaml
contextCompression:
  enabled: true
  thresholdPercent: 0.15
  maxTokens: 32000
  softBufferTokens: 1500
  hardBufferTokens: 1000
  tokenEstimateDivisor: 4
  summaryThresholdMultiplier: 3

When thresholdPercent > 0 and the model context window is known, the budget is contextWindow * thresholdPercent. Otherwise maxTokens is used. Diff entries are dropped when they exceed the soft/hard limits; above the hard limit multiplied by summaryThresholdMultiplier, all inline diffs are replaced with a summary.

Pricing overrides ​

yaml
pricing:
  models:
    opencode-go/glm-5.1:
      input: 0.0
      output: 0.0
      cacheRead: 0.0
      cacheWrite: 0.0

Used when the runtime reports zero cost. Values are USD per 1M tokens.

Fix checks ​

yaml
fix:
  checks:
    - name: typecheck
      command: npm run type-check
      matchPaths: ["src/**/*.ts"]
      timeoutMs: 60000

These checks become the drs_check tool available to the task/review-issue-fixer agent.

Temporal configuration ​

yaml
temporal:
  address: localhost:7233
  namespace: default
  taskQueue: drs-workflows
  workflowIdPrefix: drs
  workspace:
    mode: local
    root: /tmp/drs-temporal-workspaces

See Temporal execution for details.

Workflow default ​

yaml
workflow:
  default: local-changelog-review

Allows drs workflow run without a name.

Environment variables ​

  • GITLAB_TOKEN / GITHUB_TOKEN — platform tokens.
  • GITLAB_URL — GitLab URL (defaults to https://gitlab.com).
  • GITLAB_COMMIT_EMAIL_DOMAIN — optional private commit email domain for self-managed GitLab creator attribution; defaults to users.noreply.<instance-host>.
  • DRS_DEFAULT_MODEL / REVIEW_DEFAULT_MODEL — default model.
  • DRS_AGENT_<NAMESPACE>_<NAME>_MODEL / REVIEW_AGENT_* — per-agent model.
  • REVIEW_UNIFIED_MODEL — unified reviewer model.
  • REVIEW_THINKING_LEVEL — default reasoning effort.
  • REVIEW_AGENTS — comma-separated review agent ids (legacy alias).
  • DRS_RUNTIME_* — runtime timeout overrides.
  • DRS_TEMPORAL_* — Temporal overrides.

Legacy migrations ​

src/lib/config.ts rejects legacy DRS 3.x keys with a clear migration message:

  • review.default → agents.default.
  • review.defaultModel → agents.default.model.
  • review.paths → agents.paths.
  • review.postErrorComment and review.describe.postDescription — removed; posting is now explicit via workflows.
  • opencode → pi (deprecated alias, still merged with a warning).

See also ​