Implement Self-Evolving Project Skills v2 run 24

[adriandemian]

Summary

• Adds learned-skill telemetry plumbing across the MCP tool surface, chat-service recording path, HTTP handlers, and storage repositories.
• Read tools now carry transport-owned conversation/run identity without exposing those identifiers in model-facing tool schemas.
• Learned-skill usage recording is expanded to cover launch, interactive stdin, and read-path attribution, with batch persistence support in the domain repository.

User Impact

• Learned-skill usage events can be attributed more precisely to conversations and agent runs.
• Read tooling can identify the active conversation/run when that identity is available while keeping orchestration identifiers out of the model-facing schemas.
• Skill injection and usage records are more complete across Claude and Codex harness flows.

Technical Context

• learned-skill-tools now distinguishes read and write tool transport headers and reuses runtime identity helpers for read calls.
• Chat-service composition propagates injected skill names from fresh launches and resume/recovery flows.
• learned_skill_substrate and the learned-skills HTTP handler resolve injected skill names, persist usage events, and suppress unresolved or cross-project records.
• Memory and SQLite learned-skill repositories, plus the domain repository interface, now support batch recording semantics.
• Harness adapters and repository tests were updated to align launch-time learned-skill injection and usage attribution.

Risks / Follow-Ups

• This change spans MCP tooling, chat-service orchestration, HTTP, and persistence; mismatches between transport identity and persisted metadata would affect attribution.
• Learned-skill suppression depends on resolving injected skill names against the current project scope, so unresolved or cross-project names remain excluded.

View full plan

Spec §PR C2 — Injection Telemetry

Implement C2 only on the current automation base: preserve C1’s exact injected project-skill names through the existing Claude and Codex launch structures, then record trustworthy, idempotent usage for compact-index injection, full loads, composer directives, and interactive stdin.

Current-base inspection shows useful substrate is already present: C1 emits structured learned:{project_skill_id} names; the four usage kinds and usage repository already exist; composer markers and partial composer/stdin recording are landed; and MCP runtime context already carries conversation/run identity. The remaining gaps are that Claude/Codex composition drops injected names, project-skill reads do not send or validate transport-owned attribution headers, usage writes are non-idempotent/non-atomic, and stdin telemetry is currently attempted before transport success.

The implementation will extend the existing provider-neutral composition, launch, HTTP/MCP, and ledger seams—without parallel composition types, payload identity, a new usage-kind schema migration, frontend work, or any C3 outcome-linking behavior. Usage telemetry remains best-effort and cannot break an otherwise authorized read or successful delivery; atomic persistence and typed diagnostics prevent partial or falsely attributed rows. Focused behavioral tests will prove both harnesses, exact-name propagation, linkage/scoring policies, invalid/stale identity suppression, project isolation, retry safety, failure isolation, and absence of telemetry on failed or suppressed paths. The linked Blueprint is intentionally unabridged: it contains the complete ten-step implementation sequence, a four-kind scoring/linkage matrix, a production-entry behavioral matrix, focused validation commands, and a proof-obligation completion gate.

---

Generated by RalphX

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}