Wizard development

Ask who owns the concern and what they should need to understand to change it. Product knowledge belongs behind typed configuration boundaries; runner and UI infrastructure should consume those boundaries.

Execution policy and model admission

• Pi is the default choice for new work. Harness choice and model provider are separate: Pi supports gateway-backed Anthropic and OpenAI transports.
• Prefer orchestration. A seed agent plans work, then task agents execute focused conversations. Start from metrics and its context-mill flow.
• Linear is for very simple tasks and legacy support. Composed program sub-runs are also structurally clamped to linear; orchestration cannot nest through that seam.
• Anthropic Agent SDK remains supported as a legacy fallback, deprecated as the default. Retain it for major Pi vulnerabilities or missing support for new Anthropic models.

DEFAULT_BINDING selects Pi + linear for a program whose config sets no binding; flags and development overrides sit on top. Set new bindings explicitly. Migrating an existing program requires checking its flow, tasks, and lifecycle hooks; changing the default constant alone is insufficient. Both harnesses implement run and runTask.

Adding a model or effort is a cross-repository change:

1. Define the Wizard model ID and capabilities in constants and models; select it through the switchboard or supported context-mill stage metadata.
2. Check the PostHog mint's WIZARD_MODEL_ALLOWLIST and allowed efforts. The token's policy is enforced by the gateway; a local constant or CLI override cannot authorize a model.
3. Check the gateway's provider/catalog support and required Wizard system-prompt policy, including the security-triage request shape. A model not supported by that infrastructure needs a coordinated gateway/mint change before use.
4. Verify the chosen harness transport and effective effort with the same admission policy used by the target environment.

Local commandments provide runtime and tool guidance. They do not replace the gateway's required safety prompt. Keep gateway policy owned there rather than copying it into skills. Do not infer per-model effort authorization merely from the local capabilities table.

Choose the extension surface

• For a framework, follow adding-framework-support. Detection is pure; use bounded filesystem helpers and preserve specific-before-generic ordering.
• For a capability, follow adding-skill-program. Prefer a context-mill command within an existing family when that is sufficient. A native command needs a command module in src/cli/commands/ and registration in runCli (src/cli/index.ts), as well as program registration.
• For a command that runs no agent, such as mcp add or doctor, follow Add a tool. A tool never goes through runProgram, and no program imports one.
• For screens or primitives, follow ink-tui. A program's flow in src/tui/programs/<id>/ drives its screen sequence. Programs never use the UI; they get a runner context. TUI code takes the store it reports through as an argument, and session mutations use store setters that emit changes.
• For headless exploration, follow exploring-the-wizard. Drive current legal actions and inspect error outros and pending questions throughout execution.

For a new concern, first look for an existing typed surface. Add an abstraction only when it gives a real owner a smaller, reusable boundary.

Lifecycle and security

runner/index.ts resolves the route from RunConfig.routing, bootstraps, and dispatches to a sequence. Sequences own their lifecycle; harnesses own SDK calls. ProgramRun.postRun, buildOutroData, customPrompt, and abortCases are consumed by the linear sequence, not by the orchestrator. Put orchestrated work in its flow/tasks and inspect its completion path when extending it.

Keep tool enforcement at the boundary: shared permissions, harness-specific adapters, and warlock scanning. Scanner failure must not silently allow an unsafe operation. Review each adapter's blocking and termination behavior rather than assuming every rejection terminates the run. New scanner rules belong in warlock; changes to how a match is handled belong in Wizard.

User-provided secrets should travel through secret-vault references. Tool implementations resolve values host-side where they are used; they must not return the raw value to the model. Inspect the wizard-tools implementations and the Pi adapter when extending this shared tool surface.

Verification and maintenance

Reuse existing tests for routing, registration, detection, state transitions, and security behavior. Add a focused regression only when it protects a meaningful failure mode that existing checks do not cover. Avoid tests of prose, duplicated shape checks, and assertions already enforced by TypeScript.

For docs, verify local links and claimed APIs, then format edited files. For code, run pnpm typecheck and the relevant existing Vitest files; build when the change affects bundling or runtime behavior. pnpm test already builds. Avoid repository-wide pnpm fix for a scoped edit. Keep new code comments to one line.

Keep desired design policy distinguishable from current runtime behavior.

Read references as needed:

• Architecture: routing, lifecycle, security and UI boundaries.
• Anti-patterns: evaluate whether an extension fits.
• Maintaining skills: update skills when architecture or APIs change.