Consolidate coding-agent tracing behavior

[Stephen Belanger (Qard)]

Summary

• centralize agent identity, capabilities, settings, managed-run packages, and hook event manifests in one AgentSpec catalog
• separate translator delivery checkpoints from terminal finalization, preserving Pi and OpenCode sessions across ordinary flushes while closing dangling work on shutdown
• source-qualify daemon state and span identities, reject unknown sources, fix Claude managed-run routing, and make Codex attachment/recovery durable
• consolidate the current enable/disable lifecycle across all four agents; disable removes only Braintrust-owned adapter state and preserves unrelated settings
• share JavaScript config/deduplication primitives and generated shell forwarders, including adapter version provenance
• add a genuinely trace-only OpenCode entrypoint so managed tracing cannot add or bundle Braintrust data tools
• make transcript attach incremental and identity-checked, including truncation/replacement recovery and ambiguous-subagent rejection
• consolidate npm build/publish/artifact validation and align shipped hook sets with managed-run coverage
• remove the Codex smoke hook-trust bypass and require reviewed hash-scoped trust configuration

Notable behavior fixes

• bt trace run claude now emits the canonical claude-code source instead of falling through to a debug translator
• session.flush is now a checkpoint and no longer closes Pi/OpenCode roots
• Codex honors an attached external root_span_id
• Pi preserves route.additional_metadata and OpenCode accepts the shared boolean spellings
• managed Pi/OpenCode instances claim capture once per run; OpenCode injection uses @braintrust/trace-opencode/tracing
• managed child exit status is preserved through a typed error contract
• enable, disable, managed run, shipped manifests, provenance, and lifecycle output now consume shared agent definitions
• persistent JS routing continues to come only from braintrust.json; invocation-local routing is owned by bt trace run

Validation

• make test
• cargo test --manifest-path bt-daemon/Cargo.toml --all-features
• cargo clippy --manifest-path bt-daemon/Cargo.toml --all-targets --all-features -- -D warnings
• cargo fmt --manifest-path bt-daemon/Cargo.toml -- --check
• package format, lint, typecheck, exact-artifact, forwarder, hook-event, shell, Python, and diff checks

All deterministic Rust and JavaScript suites pass. The four real-agent Rust tests remain intentionally ignored locally because they require installed agent CLIs; CI runs them with its packaged-agent harness. Codex post-deploy smoke now skips unless reviewed hook hashes are configured. Claude still has no separate post-deploy marketplace smoke.

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.