Pure, source-bound queries over validated private inspection records.
The compiler performs the joins once, before a snapshot publishes any
capability. Capability inputs and outputs are paired by capability_id, and
a validated input without an output is retained as an explicitly incomplete
interrupted attempt. MCP request and response bodies remain paired by
{capability_id, request_id}. Optional stdio mcp-stderr records join that
same identity and project onto the exchange when present. Captured stdio
exchanges are serialized per session so that join is one request's capture
window, not a mix of concurrent calls. Callers therefore never join private records
by timestamp or depend on file order beyond the artifact's validated
sequence.
Every collection uses the same bounded page shape as canonical trace
queries. Cursors are opaque, bind the immutable source identity, operation,
filters, ordering, and offset, and cannot be reused against another query or
capture. The requested item limit is an upper bound: a page returns the
largest prefix whose final encoded and retained sizes both fit the result
ceiling. Run-scoped private collections accept
"order": "asc" | "desc"; ascending sequence order is the default.
Inspection artifacts retain their versioned bare-hex source hashes. Query
results expose those hashes with the sha256: algorithm prefix required by
trusted component-override descriptors, so an effective-prelude result can
be copied into base_source_hash without reinterpretation.
Execution-phase diagnostics (execution-prints, execution-error,
explicit-failure-value) are
exposed as run-scoped collections alongside their counts in each list_runs
row. Their items retain the correlated evaluation_id, sequence, timestamp,
environment, and exact bounded diagnostic payload.
Projections include mission_name on every mission-owned result so
repeated component and capability names remain unambiguous.
V8 adds the dedicated explicit-failure-value collection. V7 added
correlated private callback exception diagnostics to capability
attempts. V6 added one singular, non-paginated terminal result projection per run and
joins static prelude-call analysis to generated source by evaluation_id.
Generated source and reconstructed turns copy the canonical
parent_evaluation_id edge rather than inferring one from snapshot-local
sequence numbers. Execution errors and generated sources add typed,
state-bearing navigation relationships only from validated identities,
direct boundary-value provenance, and exact source/prelude associations.