Eval capture

Status: stable from v0.21.0. Schema versioning via schema_version on every row; additive changes increment the minor version; removals are breaking-schema-v2.

Audience: downstream consumers (primarily the sibling gbrain-evals repo) that replay captured real-world queries as a BrainBench-Real fixture.

The pipeline

MCP / CLI / subagent tool-bridge caller
     │
     ▼
src/core/operations.ts — query + search op handlers
     │
     │ (hybridSearch or searchKeyword)
     │
     ▼
{results, meta: HybridSearchMeta}                 ┌── captureEvalCandidate
     │                                             │    (fire-and-forget)
     ▼                                             │
return to caller                                   ▼
                                            scrubPii(query) ←── src/core/eval-capture-scrub.ts
                                                   │
                                                   ▼
                                           buildEvalCandidateInput
                                                   │
                                                   ▼
                                           engine.logEvalCandidate
                                                   │
                                    ┌──────────────┴──────────────┐
                                    │ success                     │ fail
                                    ▼                             ▼
                                INSERT into eval_candidates    engine.logEvalCaptureFailure
                                                                 (reason: db_down | rls_reject |
                                                                  check_violation |
                                                                  scrubber_exception | other)

gbrain eval export — the consumer contract

gbrain eval export [--since DUR] [--limit N] [--tool query|search]

Emits NDJSON to stdout. One JSON object per \n-terminated line. stderr receives progress heartbeats. Every line starts with "schema_version": 1 so a forward-compat parser can fail loudly on schema v2 instead of silently misparsing.

Typical usage from gbrain-evals:

# Snapshot the last week of real traffic for replay
gbrain eval export --since 7d > brainbench-real.ndjson

# Stream through jq for ad-hoc analysis
gbrain eval export --tool query | jq -c 'select(.latency_ms > 500)'

Row schema (v1)

Every exported row has this shape. Field order in JSON output is not guaranteed; consumers MUST key by name, not position.

Field	Type	Notes
schema_version	number	Always 1 on v1 rows. Forward-compat gate.
id	number	Autoincrement primary key. Stable across exports.
tool_name	"query" | "search"	Which MCP operation captured this row.
query	string	Already PII-scrubbed by scrubPii unless eval.scrub_pii: false. Emails / phones / SSN / Luhn-verified credit cards / JWTs / bearer tokens replaced with [REDACTED]. Max length 50KB (CHECK-enforced).
retrieved_slugs	string[]	Deduplicated slugs that came back in SearchResult[].
retrieved_chunk_ids	number[]	Every chunk id in result order (duplicates preserved — one per hit).
source_ids	string[]	Distinct sources.id values across the result set (v0.18 multi-source). Empty for pre-v0.18 rows that lacked the column.
expand_enabled	boolean | null	Whether the caller requested Haiku expansion. null for search (no expansion concept).
detail	"low" | "medium" | "high" | null	Detail level the caller requested. null when omitted.
detail_resolved	"low" | "medium" | "high" | null	What hybridSearch actually used after auto-detect. null when neither caller nor heuristic classified.
vector_enabled	boolean	True iff vector search actually ran. false when OPENAI_API_KEY was missing or the embed call failed. Replay MUST respect this — rows with false only exercised the keyword path.
expansion_applied	boolean	True iff Haiku expansion actually produced variants (not just "was requested").
latency_ms	number	Wall-clock duration of the op handler (includes capture itself — negligible since it's fire-and-forget).
remote	boolean	true for MCP callers (untrusted), false for local CLI. Partitions "real agent traffic" from "operator probing."
job_id	number | null	OperationContext.jobId when the caller was a subagent tool-bridge. Null for MCP + CLI.
subagent_id	number | null	OperationContext.subagentId for subagent-owned runs.
created_at	string (ISO 8601)	UTC timestamp of insert.

Ordering + determinism

listEvalCandidates orders by created_at DESC, id DESC. Same- millisecond inserts tie on created_at; id DESC is the stable tiebreaker. Replay tools can consume rows in order and assume:

• no duplicate rows across calls with non-overlapping --since windows
• no missed rows across calls that chain --since windows (window end of run 1 is the strict upper bound, not a soft cursor)

Schema versioning promise

• v1 (shipped v0.21.0) — this document. All fields listed above.
• Additive changes increment gbrain minor version (v0.25.0, v0.23.0 …) and ship with new optional fields. Consumers keyed on known fields ignore unknown keys and keep working.
• Breaking changes (rename, type change, removal) increment schema_version to 2. Consumers MUST branch on schema_version to stay compatible.

eval_capture_failures — companion audit table

Not exported by gbrain eval export. Surfaced via gbrain doctor:

gbrain doctor   # warns when failures in last 24h > 0

Reason enum (stable): db_down | rls_reject | check_violation | scrubber_exception | other. Cross-process visibility is the whole point — gbrain doctor runs in its own process and reads the table directly, so in-process counters wouldn't work.

Config + CONTRIBUTOR_MODE

Capture is off by default as of v0.25.0 (was on for everyone in earlier drafts). Two paths to turn it on:

Path A — env var (contributor opt-in, the common case):

export GBRAIN_CONTRIBUTOR_MODE=1     # in ~/.zshrc or ~/.bashrc

Path B — explicit config (~/.gbrain/config.json, file-plane only):

{
  "engine": "postgres",
  "database_url": "...",
  "eval": {
    "capture": true,
    "scrub_pii": true
  }
}

Resolution order (most explicit wins):

1. eval.capture: true in config → on
2. eval.capture: false in config → off (overrides CONTRIBUTOR_MODE=1)
3. GBRAIN_CONTRIBUTOR_MODE === '1' → on
4. otherwise → off

scrub_pii defaults to true independent of capture. Set eval.scrub_pii: false to preserve raw query text (only if you control the brain's distribution).

gbrain config set eval.capture false does not work — that command writes the DB-plane config, and the MCP server reads the file-plane. Edit the JSON directly or use the env var.