agent-skills Agent Guide

Agent-Skills is a transitional migration repository. The approved lifecycle is Foundry holding -> SDK checking and preparation -> private Tessl registry -> Codex runtime installation, with verified OpenAI-provider exemptions. Complete that lifecycle and repository independence before refining either product.

Jamie approved a bounded transitional exception on 2026-09-27: the selected existing custom skills may run from ~/.agents/skills without prior SDK/Tessl processing until SDK supports the required workflows. They remain unchecked, not SDK-cleared. This does not cover plugins or new external intake. Follow the selected-set and preservation rules.

Root Essentials

• On a fresh checkout, run bash scripts/bootstrap-ask.sh --json before requiring ./bin/ask; if the wrapper is unavailable, verify the fallback with python3 bin/ask repo status --json.
• Use ./bin/ask for repo operations; it forwards to Infrastructure/bin/ask.
• The repository root has no package manager install step. Use repo wrappers at the root, and use package commands only inside verified package roots.
• Inspect YAML through the managed wrapper lane, not ad hoc system Python: ./bin/ask repo yaml-inspect <repo-relative-yaml> --json --robot or mise run yaml-inspect -- <repo-relative-yaml>. Plain python3 is not a PyYAML contract for this repo.
• Run Infrastructure Python tests through the locked Infrastructure environment, not the system Python from the repo root. Use bash Infrastructure/scripts/run-infrastructure-python.sh -m unittest tests.<module> or bash Infrastructure/scripts/run-infrastructure-python.sh -m pytest <path> so dependencies from Infrastructure/uv.lock are available.
• Before changing skills, sync policy, runtime projections, or agent-facing docs, read UBIQUITOUS_LANGUAGE.md.
• Before changing Skills SDK plans, specs, atlas visuals, capability claims, or product-direction docs, use the product-boundary terms in UBIQUITOUS_LANGUAGE.md: Foundry holds candidates awaiting SDK processing; SDK owns the agent-facing workflow, reusable tooling and required registry/runtime adapters. Tessl holds checked versions privately; only Jamie decides public release. Codex home runtimes consume those versions. OpenAI plugins and system skills are exempt by verified origin. Follow the six rules in Path Ownership Boundaries. An explicit owner decision—not holding, checking or installation—decides editable source authority.
• For a package selecting skills-sdk.gold-standard.v1, read Skills SDK Authoring Contract, update references/contract.yaml: authoring_contract with the skill, and run ./bin/ask skills package verify <skill-path> --json --robot. That is a structural admission gate only: run the declared behavioral proof and the selected runtime or external lane before claiming the skill is ready.
• Edit canonical sources, not runtime projections. See Path Ownership Boundaries.
• Finish the named outcome through the smallest local change and focused proof. Feedback is diagnostic evidence, not automatic authority to stop delivery or add process. Select the High-Signal Steering Feedback only for its named conditions. Routine delivery does not require a steering ledger entry or new process artifact. When that route is selected, opening and reading it in the current turn is required; record the result in .harness/quality/steering-uptake.md and run python3 Infrastructure/scripts/validation-and-linting/validate_steering_uptake.py --json.
• Keep runtime handles operation-bound: wait, poll, or resume only the latest valid handle returned for that operation, following the tool's lifecycle rules. Unrelated tool calls do not invalidate a live handle. Perform these operations serially for that handle. See High-Signal Steering Feedback for recovery rules.
• Separation coordination: task 01a025cc-c335-71b2-89ba-e32d7b0bf848 is the master controller for OC - Skills SDK and OC - Foundry. It assigns bounded migration units and accepts their evidence; owners do not start product refinement or redefine programme completion. Use completion messages, not polling or periodic status loops. Follow PM Thread Coordination.
• Keep local, runtime, cloud, Tessl, and release proof separate. Use Tessl Live Skill Eval Workflow and Skills SDK Runtime Lane Contract for the applicable selected actions. Registry and runtime proof remain required programme outcomes; a local task does not authorise live side effects.

Project-specific correction boundary

Routine local skill corrections use existing local checks and do not invoke cloud execution, external Tessl distribution, live runtime mutation, publication, or release machinery unless promotion is explicitly selected. Pre-merge local review remains required for PR-bound changes. This source-only path does not waive the skill-management contract: runtime projection proof is required before claiming a skill is usable or using it as a runtime skill; that proof belongs to the separately selected runtime or promotion lane.

Further Guidance

• Repo command behavior and --robot mode: Agent Operating Contract
• Tooling and package command policy: Tooling and Command Policy
• Validation order and evidence rules: Validation and Checks
• Workflow, git, refactoring, docs, and safety: Workflow and Safety Guidance
• Skill install, audit, fold, and line-budget rules: Skill Management
• Browser and local preview fallback: Browser and Local Preview
• Instruction cleanup notes and stale guidance to remove: Contradictions and Cleanup

Start at Docs/agents/README.md when you need the full instruction map.