Track generic tool runtime metadata in the conversation substrate

[chubes4]

Summary

Data Machine and Data Machine Code now use a generic runtime metadata convention on tool declarations/results to avoid hardcoding product-specific tool names in the agent loop. Track whether that contract should move into Agents API as substrate-owned tool runtime semantics.

Related shipped PRs:

• Data Machine: Use tool runtime metadata in agent loop Extra-Chill/data-machine#1951
• Data Machine Code: Declare agent tool runtime metadata Extra-Chill/data-machine-code#360

Current state

Data Machine currently reads tool metadata in its caller-managed turn runner:

• runtime.duplicate_policy = repeatable lets safe inspection tools be called repeatedly with identical parameters.
• runtime.completion_signal = progress lets state-changing tools trigger immediate missing-completion assertion nudges.

This solved a product-boundary bug: Data Machine core should not know about Data Machine Code extension tool names like workspace_read, workspace_git_status, or create_github_pull_request.

Why this may belong in Agents API

Agents API already owns neutral pieces of the conversation substrate:

• tool declarations
• optional tool-call mediation
• tool execution result shape
• completion policy callbacks
• iteration budgets
• transcript persistence

If duplicate-call policy and runtime result metadata are generic behavior, the contract should be documented and/or implemented in Agents API rather than remaining a Data Machine convention.

Proposed direction

• Define a generic tool runtime metadata shape for declarations/results, likely under tool_declarations[*].runtime and tool_result.runtime.
• Preserve product-agnostic keys such as:
  • duplicate_policy: repeatable
  • completion_signal: progress
• Have Agents API mediated tool execution propagate normalized runtime metadata into tool_execution_results.
• Consider whether duplicate-call prevention should become an optional Agents API loop policy instead of caller-managed Data Machine behavior.
• Keep product-specific completion assertion wording/policy in Data Machine unless/until a generic substrate abstraction emerges.

Acceptance criteria

• Agents API has documented semantics for runtime metadata on tool declarations/results.
• Generic mediated tool execution preserves that metadata in tool execution results.
• Data Machine can eventually delete local helper glue or treat it as compatibility only.
• No product-specific tool names or Data Machine Code assumptions are introduced into Agents API.