gaze: Skill for OpenCode

Determinism First

Agent context is a scarce shared resource (Constitution Principle VI, "Deterministic Execution"). Deterministic procedures MUST execute in tools rather than be re-reasoned from prose on every run. This skill is the authoring-time check that keeps mechanical work out of instruction text.

Classify Before Writing

Before writing or modifying a command, agent, skill, or any instruction text, classify every step of the work:

• Deterministic — fixed procedure, same output every time. It belongs in a tool call (script, CLI command, or plugin), not improvised output.
• Judgment-based — a decision, an evaluation, or human interaction. It belongs in prose.

The test: if a competent agent follows the step twice, does it produce the same result both times? If yes, the step is deterministic and belongs in a tool. When classification is uncertain, look at the concrete examples below before defaulting to prose. The same test applies one level up: a fixed sequence of tool calls with no judgment between steps is itself deterministic.

Deterministic Steps Belong in Tools

Mechanical procedures that MUST be tool calls, not instructed prose:

• Parsing structured input (YAML, JSON, diffs, frontmatter)
• Existence and state checks (file present, branch current, tool available)
• Git choreography (worktrees, staging, merge-base diffs)
• Prescribed terminal output rendered from templates
• Enumeration loops over directories, files, or checks
• Orchestration of tool chains (ordering, batching, retry, failure-stop)
• Branching on structured results via lookup tables (verdict mapping, severity sorting)

When a deterministic step is needed:

1. Find an existing tool that performs it and call it.
2. If none exists, extract the procedure to an executable script or command with tests — do not restate the procedure in prose for the agent to re-derive.
3. Prescribed output MUST be tool-rendered, not reconstructed by the agent from a template embedded in instructions.
4. If the tool now sits in a judgment-free chain of calls, collapse the chain into a single intent-level tool; pass judgment products (composed text, gate decisions) as parameters.