SDK provenance context primitives across CLI, MCP, and event streams

Merged
UN
3mo ago
+2672-172

Summary

  • add bounded automatic model/version resolvers with privacy-safe opt-outs and extensible SDK descriptors
  • add versioned legacy-author interpretation plus patch-free provenance projections, filters, and bounded summaries across history, activity, and mutation events
  • keep SDK, CLI, MCP, completions, settings, generated contracts, and documentation aligned
  • add focused and end-to-end coverage while preserving exact repository-wide 100/100/100/100

PM lineage

Verification

  • pnpm typecheck and pnpm lint
  • 422 files and 6,747 tests at exact 100% statements/branches/functions/lines
  • pm-changelog 2026.8.2 check unchanged
  • isolated packed npm/SDK and Bun/bunx acceptance
  • packed-consumer npm audit: 0 vulnerabilities across 64 packages
  • GitHub Dependabot, CodeQL, and secret-scanning open alerts: 0

Summary by Sourcery

Introduce shared SDK/CLI/MCP primitives for immutable, patch-free provenance context across history, activity, and mutation event streams, backed by bounded automatic probes, versioned legacy author interpretation, and extended indexing and filtering.

New Features:

  • Add patch-free provenance projections, filters, and bounded summaries for history, activity, and mutation event streams across SDK, CLI, and MCP.
  • Introduce automatic, bounded provenance probes (including harness session file resolvers and version extraction) with privacy-safe opt-outs and workspace identity vocabulary for legacy authors.

Bug Fixes:

  • Treat provenance, compact, and full projections as mutually exclusive in history, activity, and events, and enforce CLI constraints such as disallowing provenance summaries while following event streams.

Enhancements:

  • Extend the history event index schema and queries to support harness, agent instance, and provenance-based filtering without scanning raw history streams.
  • Align CLI flags, completion scripts, tool schemas, and parameter metadata with new provenance options for history, activity, and events.
  • Refine agent identity and settings handling to support probe configuration, identity vocabulary, and ordered serialization of agent_identity settings sections.

Tests:

  • Add focused unit tests for provenance probes, history/activity/events provenance projections and filters, CLI option normalization and mapping, completion flag coverage, and updated settings and provenance coverage analysis.

Summary by cubic

Adds patch-free provenance reads and filtering across history, activity, and events, powered by bounded probes and a versioned legacy-author vocabulary to make identity readable and safe. CLI, SDK, and MCP share the same provenance options, with privacy opt-outs and constant-size summaries.

  • New Features

    • Patch-free provenance projections for history/activity/events, plus optional summaries; filter by harness, agent_instance, and provenance dimensions; resolve legacy authors via a versioned vocabulary.
    • Bounded provenance probes (model, version) with opt-out via settings; new agent_identity.probes_enabled and agent_identity.identity_vocabulary { version, aliases }.
    • Event index v2 adds harness, alias author, agent_instance, and provenance-dimension filtering without scanning raw streams; CLI/MCP/SDK schemas and completion scripts updated with repeatable --provenance, --provenance-summary, --harness, --agent-instance, and --provenance-filter.

  • Bug Fixes

    • Treat compact, full, and provenance projections as mutually exclusive and enforce CLI constraints (e.g., no provenance summaries while streaming events).
    • Update docs to link provenance items at canonical paths and record exact-head review evidence.
    • Add explicit provenance errors (invalid_provenance_predicate, unknown_provenance_dimension) and refresh public surface and token baselines.