Run-scoped artifact identity for subagent writes

[Whamp]

Parent PRD

#1

What to build

Implement run-scoped artifact identity for subagent outputs. Subagents should continue writing logical names such as context.md, while the framework stores those artifacts under a unique namespace keyed by the subagent run id. Repeated writes to the same logical artifact name within a single subagent run should update that run-scoped artifact rather than creating ambiguous sibling files. This slice should cover the end-to-end artifact identity path needed by the parent PRD's solution and implementation decisions.

Acceptance criteria

• A subagent can write a logical artifact name like context.md and the stored artifact path is namespaced by the unique subagent run id.
• Repeated writes to the same logical artifact name within one subagent run update the same run-scoped artifact identity rather than creating versioned siblings.
• read_artifact(...) can resolve the stored namespaced artifact correctly using the surfaced artifact name.
• Pure tests cover artifact namespace resolution and duplicate-write behavior.

Implementation approach

• Pass PI_ARTIFACT_RUN_ID and PI_ARTIFACT_SESSION_DIR env vars to subagent processes
• write_artifact uses these to store under <parentSessionDir>/artifacts/<runId>/
• read_artifact searches run-scoped directories in addition to session-scoped ones
• Pure tests for namespace resolution and duplicate-write idempotency