Skip to content

configuration.md

Latest commit

 

History

History
74 lines (59 loc) · 2.83 KB

configuration.md

File metadata and controls

74 lines (59 loc) · 2.83 KB
Raw

Configuration

config/stages.yaml

Per-stage settings. Override the file with --config <path>.

defaults:
  max_turns: 25
  permission_mode: acceptEdits   # never bypassPermissions
  repair_attempts: 1             # re-emit chances on schema-validation failure

stages:
  recon:
    model: claude-opus-4-7
    concurrency: 1
    tools: [Read, Grep, Glob, Bash]
    max_turns: 60
    repair_attempts: 2
  hunt:
    model: claude-sonnet-4-6
    concurrency: 50
    tools: [Read, Grep, Glob, Bash]
  validate:
    model: claude-opus-4-7        # different from hunt — deliberate disagreement
    concurrency: 10
    tools: [Read, Grep, Glob]     # no Bash: pure analysis
  # … gapfill, dedupe, trace, feedback, report …

loops:
  gapfill_iterations: 2     # max Hunt → Validate → Gapfill cycles
  feedback_iterations: 1    # feedback passes after Trace

Per-stage fields

Field Meaning
model Model name (e.g. claude-opus-4-7). With a gateway, slash forms like anthropic/claude-opus-4-7 work
concurrency Max simultaneous agents for that stage (capped by --max-concurrency)
tools Allowed tool names for the stage's agents
max_turns Max agentic tool-use rounds per agent (falls back to defaults)
permission_mode SDK permission mode (acceptEdits by default; never bypassPermissions)
repair_attempts How many times the runner asks the model to fix schema-invalid output

Rules worth knowing

  • Hunt and Validate must use different models. This is the load-bearing "deliberate disagreement" rule and is enforced by a test.
  • Trace runs on the strongest model — it's the reachability gate.
  • Lowering concurrency (or --max-concurrency 1) is the simplest cost lever.

Loop counts

Key Effect
gapfill_iterations The Hunt → Validate → Gapfill loop runs up to this many extra times (so 2 ⇒ up to 3 Hunt passes). Set 0 to disable Gapfill.
feedback_iterations The Feedback → Hunt → Validate → Dedupe → Trace loop runs up to this many times. Set 0 to disable Feedback.

Both loops also break early when they stop producing new work.

Environment variables

Variable Purpose
CLAUDE_CODE_OAUTH_TOKEN Headless subscription token (claude setup-token)
ANTHROPIC_BASE_URL Gateway base URL (e.g. OpenRouter) — triggers gateway mode with a token
ANTHROPIC_AUTH_TOKEN Bearer token for a gateway
ANTHROPIC_API_KEY Metered Anthropic key (only honored with --allow-api-key)
ANTHROPIC_MODEL Forces every stage onto one model (useful with gateways)
AUDIT_ALLOW_API_KEY 1 to opt into API-key billing without the flag

.env is loaded automatically. See Authentication for how these interact and what gets scrubbed.