beads/docs/METADATA.md
matt wilkie 3bf890cb21
docs(metadata): document portability of execution model/effort hints (#4358)
State the rule that makes execution_reasoning_effort and
execution_suggested_model portable across agent runtimes: stored effort
values are a canonical advisory scale that consumers map to their own
runtime, and suggested models are capability-tier hints that consumers
on another provider substitute rather than drop. Also update the
gh-3541 reference: that issue resolved to keep execution hints
metadata-only with no first-class CLI helper.

Agent-Signature: claude-fable-5-high on behalf of matt wilkie
2026-06-10 09:30:41 -07:00

3.6 KiB

Issue Metadata

The metadata field on issues accepts arbitrary JSON. Any valid JSON value is stored as-is.

Metadata is the preferred extension point for data that is specific to an integration, orchestrator, team workflow, or experimental automation. Before adding first-class fields, commands, or schema changes, check the Project Charter.

Example: Agent Execution Metadata

Agent execution hints are one example of using metadata to extend beads without adding new native database fields. Automation may store these hints so agents can make routing decisions without parsing prose. Agents enacting an issue should read metadata first, then use description and notes for scope and rationale:

bd show <id> --json | jq '.[0] | {id,title,metadata,description,notes}'

The current convention for execution hint keys is:

Key Meaning
execution_agent_type Suggested worker class, such as explorer, worker, or mixed.
execution_suggested_model Suggested model for the parent agent or spawned subagent.
execution_reasoning_effort Suggested reasoning effort, such as low, medium, high, or xhigh.
execution_mode Whether work should be local, delegated, or staged between delegated and local execution.
execution_parallel_group Grouping hint for work that can run alongside related tasks.

These keys are advisory metadata, not core issue fields. When a workflow uses them, they take precedence over free-form notes for execution routing. Notes remain useful for rationale, ownership, and exact prompts.

Model and effort values are portable hints, not runtime bindings. The execution_reasoning_effort values above are a canonical advisory scale: writers should store canonical values rather than runtime-local ones, and a consumer whose runtime uses a different scale should map the stored value to its nearest native level instead of dropping the hint. Likewise, execution_suggested_model is a capability-tier suggestion: a consumer on a different provider should substitute a model of the same tier rather than ignore the hint.

Parent/orchestrator agents must consume these keys before spawning subagents. Model and reasoning effort are normally fixed at launch, so reading metadata after delegation is too late.

Do not add a first-class helper such as bd show <id> --execution or bd plan <id> --json. Issue gh-3541 resolved to keep execution hints as metadata only; the JSON/JQ snippet remains the supported access path.

Example: Tracker Round-Trip Metadata

Tracker integrations map external issues into beads fields such as title, status, priority, type, labels, dependencies, and external_ref. When an integration needs to preserve tracker-specific fields that do not belong in the native beads schema, it can store those fields in issue metadata:

{
  "example_tracker": {
    "board_id": "ENG",
    "sprint_id": 42,
    "remote_type": "story"
  }
}

This keeps beads' core issue model stable while allowing the integration to round-trip fields it understands. Prefer namespaced keys and keep tracker-specific policy in the integration. If a value becomes broadly useful to beads itself, revisit whether it deserves a native field.

Reserved Key Prefixes

Prefix Reserved For
bd: Beads internal use
_ Internal/private keys

Avoid these prefixes in user-defined keys to prevent conflicts with future Beads features.