Explain and Profile CodeQuery

query_code has three root execution modes. results is the default and returns the existing query-result object. explain lowers and selects a plan without accessing analyzer data during that phase; a host such as the one-shot CLI may still initialize and index its workspace before the request runs. profile executes the query and returns the ordinary result together with opt-in measurements. The mode is an output control: it is excluded from query-plan identity, cannot appear inside a set branch, and is not available to static-analysis policy selectors.

Request an explanation

Set execution_mode on a JSON query:

{
  "schema_version": "1",
  "execution_mode": "explain",
  "union": [
    {"match": {"kind": "class", "name": "Legacy"}},
    {"match": {"kind": "class", "name": "Replacement"}}
  ]
}

Or wrap an RQL query:

(profile
  (references-of :proof proven
    (enclosing-decl (method :name "handle"))))

Use (explain QUERY) for the planning-only form. explain and profile are mutually exclusive because both lower to the same root enum. A saved .json or .rql query may contain the mode directly, so --query-file and MCP query_file calls use the same contract.

Explain response

An explanation has format bifrost_code_query_explain/v1 and these fields:

• query_schema_version identifies the parsed CodeQuery language version. It is independent of the report contract version carried by format.
• parsed_query is canonical JSON for the resolved query schema version, including defaults.
• logical_plan is a dependency-first directed acyclic graph. Stable, plan-local node IDs and ordered dependency arrays expose shared work; a dependency ID may appear more than once when authored branches share the same logical seed.
• physical_plan maps each physical node to its logical node, selected operator, output kind, and ordered physical dependencies. A semantic step also carries semantic_request, which states with eight booleans which semantic facets the step would need. Explain publishes that demand without resolving a host registration or materializing semantic artifacts. Logical and physical IDs are separate contracts even where the current planner selects them one-to-one.
• scheduling records the scheduling policy, the selected strategy, and the maximum concurrency for the selected plan.

The explanation omits internal cache fingerprints, storage generations, representation versions, and executor-only suffix flags. Those values are implementation details, not a stable public planning contract. Explain fields lists every field.

Profile response

A profile has format bifrost_code_query_profile/v2. Its result field is the exact ordinary CodeQueryResult; result ordering, diagnostics, provenance, truncation, and completion semantics do not change. The remaining fields are observations from that same execution:

• explain contains the selected public plan described above.
• timings_ns separates planning, execution, result rendering, and the execution-core total elapsed wall time.
• request_timings_ns measures the whole tool request around the query: transport queueing, workspace readiness, analyzer admission, preparation, input decoding, query execution, and rendering. The MCP server fills every field. bifrost --tool query_code, bifrost --query-file, and the Rust SearchToolsService::query_code_result method also fill them, but report zero for the three transport wait fields; query_code_result also reports zero for rendering_serialization. Lower-level Rust execution functions such as execute_workspace leave every field at zero.
• work reports budget-accounted files, source bytes, fact nodes, pipeline rows, references, provenance steps, resolved import files and edges, and semantic work. Its nested semantic object reports materialization, semantic IR element counts, cumulative work and resident and per-root work peaks across the semantic lanes, retained bytes, traversal steps, and whether the separate semantic budget was exhausted. When typestate, value-flow, or class-set type-flow analysis runs, semantic.typestate, semantic.value_flow, or semantic.type_flow adds that solver’s counters. A pure witness projection adds no solve work.
• cache_layers reports deterministic, completeness-sensitive outcomes for request-local seed results, seed-path structural facts, reference and call relations, forward and reverse import relations, and the snapshot-local complete direct-import topology. The topology layer includes build files, edges, elapsed time, retained bytes, cancellation and unavailability, and request-local fallbacks. Only complete immutable topology values are reused; unsupported reverse domains and failed or bounded builds fall back to the request-local resolver.
• access_path separates compatibility-budget work from physical structural access. It reports the selected positive posting terms and their pre-intersection candidate counts, provider, scoped, and candidate facts, compatibility-admitted fact nodes, whether exact source-anchor verification remained necessary, cache readiness, materialization and source inspection, index lifecycle and build cost, fallback, and uniquely retained bytes. scoped_fact_nodes is exact when indexed metadata is available and zero for scan-only execution; admitted_fact_nodes is comparable on both paths. A narrow scope may reuse a ready snapshot index. Bifrost does not build a whole-provider index for a query on its own when the scoped files are fewer than eight or less than one quarter of that provider. For a viable scope, the first use scans and records reuse interest; a later use builds the snapshot index. This avoids whole-workspace index construction for one-shot queries.
• scheduling reports peak concurrency and, when bounded dispatch actually occurred, queue, coordinator, budget-wait, and dispatch observations.
• operators reports authored invocation identity, disposition and termination reasons, row counts, elapsed, wait, merge, and scheduling time, work and cache deltas, truncation, cancellation, and a lower-bound temporary container-capacity estimate.

Times are monotonic elapsed wall time in nanoseconds. timings_ns.total covers the executor request core through ordinary result construction. It excludes analyzer query-scope setup and cleanup, public-report projection, serialization, transport, and host workspace initialization, so it is not end-to-end tool latency; use request_timings_ns.total for that. Operator-inclusive intervals can overlap and must not be summed as CPU time. Temporary memory excludes allocator metadata and heap payloads owned by strings, paths, traces, and nested vectors, so it is a lower bound rather than RSS or retained heap.

Cancellation fields are observable through the public Rust execute_request_with_cancellation API, which returns a profiled cancellation-safe partial result.

Profiling is opt-in: an ordinary results query does not allocate the plan explanation, operator observation vector, or per-operator timers. A profiled request may warm analyzer caches, so compare cold and warm profiles using separate, deliberately controlled analyzer and storage lifecycles. Profile fields lists every field.

Current scheduling policy

The auto scheduling policy always selects the sequential_union operator. Bifrost keeps a bounded two-worker parallel_union implementation and a parity and benchmark harness, but it has no request option that forces parallel execution.

The final M4 release experiment used TypeScript workspaces of 257, 513, and 1,001 files on an Apple M4 with ten logical processors. Exact unions scanned one file per branch; broad unions scanned 128, 256, or 500 files per branch. Each warm cell used eight paired measurements, and the second round reversed stable candidate order. Representative absolute elapsed times were:

Workspace	Shape	Round 12 cold sequential / parallel	Round 13 cold sequential / parallel	Round 12 warm median sequential / parallel	Round 13 warm median sequential / parallel
257	exact	4.597 / 9.001 ms	10.558 / 10.252 ms	5.361 / 9.459 ms	5.952 / 8.559 ms
513	exact	16.612 / 16.946 ms	36.146 / 22.976 ms	15.684 / 25.430 ms	31.705 / 19.476 ms
1,001	exact	16.647 / 31.457 ms	142.964 / 38.171 ms	22.798 / 35.798 ms	26.796 / 50.630 ms
257	broad	47.622 / 44.388 ms	329.183 / 69.796 ms	34.340 / 37.236 ms	233.110 / 112.244 ms
513	broad	104.182 / 91.845 ms	513.850 / 297.272 ms	80.580 / 83.848 ms	386.303 / 151.409 ms
1,001	broad	841.686 / 428.313 ms	877.218 / 619.713 ms	222.465 / 402.767 ms	782.926 / 329.696 ms

Exact cold results changed sign at every scale after candidate order reversed. Broad warm medians favored sequential at every scale in round 12 and parallel at every scale in round 13. Absolute times varied by several multiples under system load even though results, work, cache contracts, and source fingerprints stayed equal. Bifrost therefore has no stable, observable pre-execution signal for a production threshold. This is a measured negative policy, not a claim that parallel unions can never win.

Three adjacent alternatives stopped at different evidence gates:

• bitmap-backed sets were ineligible because no stable dense identity domain was established;
• the normalized SQL edge-table negative control was discarded after it returned a non-equivalent consumer result and lacked an exact publishable snapshot identity;
• arbitrary recursive graph splitting remained a safety and ownership non-goal because current branch-local caches, resolver state, fair budgets, and derived dependencies do not yet have a safe general concurrent model.

A future selector must demonstrate a stable cold-and-warm crossover, an observable pre-execution cache/cardinality signal, exact result and budget parity, and behavior under concurrent request load before the auto policy can select parallel execution.

Explain fields

Every count is a non-negative integer. IDs are plan-local integers that are meaningful only inside one report.

Explain object

Field	Type	Meaning
format	string	Always bifrost_code_query_explain/v1.
query_schema_version	integer	CodeQuery schema version of the parsed query. The only version is 1.
parsed_query	object	Canonical JSON for the query, with defaults filled in.
logical_plan	object	The logical plan. See Logical plan.
physical_plan	object	The selected physical plan. See Physical plan.
scheduling	object	The scheduling decision. See Explain scheduling.

Logical plan

Field	Type	Meaning
logical_plan.root	integer	ID of the node whose output is the query result.
logical_plan.nodes	array	Logical nodes in dependency-first order.
nodes[].id	integer	Logical node ID.
nodes[].operation	object	What the node computes. operation.kind selects the shape below.
nodes[].output_kind	string	The row domain the node produces, such as structural_match, declaration, file, or procedure.
nodes[].dependencies	array of integers	Ordered input node IDs. An ID repeats when two authored branches share one logical node.

operation.kind values:

kind	Extra field	Meaning
seed	seed: object	A structural match seed, as canonical JSON.
occurrence_seed	seed: object	An occurrence seed, as canonical JSON.
scope_seed	seed: object	A lexical-scope seed, as canonical JSON.
binding_seed	seed: object	A binding seed, as canonical JSON.
generation_site_seed	seed: object	A generation-site seed, as canonical JSON.
export_seed	seed: object	An export seed, as canonical JSON.
path_seed	seed: object	A path seed, as canonical JSON.
dependency_evidence_seed	seed: object	Stored static dependency evidence filters, as canonical JSON.
configuration_facts_seed	seed: object	A configuration-facts seed, as canonical JSON.
step	step: object	One pipeline step, as canonical JSON, for example {"op": "procedure_of"}.
set	op: string	A set operation: union, intersect, or except.
limit	count: integer	The public result limit.

For the seeds and steps themselves, see Code Query Vocabulary.

Physical plan

Field	Type	Meaning
physical_plan.root	integer	ID of the physical node whose output is the query result.
physical_plan.nodes	array	Physical nodes in dependency-first order.
nodes[].id	integer	Physical node ID.
nodes[].logical_node	integer	ID of the logical node this physical node implements.
nodes[].operator	string	The selected operator. See the table below.
nodes[].output_kind	string	The row domain the node produces.
nodes[].dependencies	array of integers	Ordered input physical node IDs.
nodes[].semantic_request	object, optional	Present only on a step that needs semantic facts. See the table below.

operator values:

Operator	Meaning
seed_scan	Evaluates a structural match seed.
occurrence_scan	Evaluates an occurrence seed.
scope_scan	Evaluates a lexical-scope seed.
binding_scan	Evaluates a binding seed.
generation_site_scan	Evaluates a generation-site seed.
export_scan	Evaluates an export seed.
path_scan	Evaluates a path seed.
dependency_evidence_scan	Reads the selected durable static dependency publication.
configuration_facts_scan	Evaluates a configuration-facts seed.
pipeline_step	Runs one step over its input rows.
sequential_union	Computes a union, one branch at a time.
parallel_union	Computes a union on bounded parallel workers. The auto policy does not select it today.
sequential_intersection	Computes an intersection.
sequential_except	Computes a set difference.
limit	Applies the public result limit.

semantic_request has exactly these eight booleans. Each one is true when the step needs that facet:

Field	Meaning when true	Steps that set it
procedures	The step resolves or reads source-backed procedures.	Every step that has a semantic_request.
dispatch	The step resolves call dispatch targets.	For example call_bindings, dispatch_outcome, dispatch_targets, class_set, absent_member, nilness_operations, and the result-contract steps.
program_points	The step reads control-flow program points.	For example cfg_entry, cfg_exits, state_events_of, statement_reachability, guards_of.
control_edges	The step reads control-flow edges.	For example cfg_successor_edges, cfg_predecessor_edges, control_relations, loop_relations, switch_coverage.
switch_facts	The step reads switch facts.	switch_coverage.
typestate	The step runs a registered typestate protocol.	typestate.
value_flow	The step runs value-flow propagation.	value_flow, class_set, absent_member, keyed_read_value.
taint	The step projects a registered taint result.	taint.

Explain scheduling

Field	Type	Meaning
scheduling.policy	string	Always auto.
scheduling.selected	string	sequential or parallel. It is parallel only when the plan contains a parallel_union node.
scheduling.max_concurrency	integer	1 for a sequential plan. For a parallel plan, the smaller of the scheduler worker count and the parallel union’s branch count.

Profile fields

The profile object has these top-level fields:

Field	Type	Meaning
format	string	Always bifrost_code_query_profile/v2.
result	object	The ordinary query result, unchanged.
explain	object	The explain object for the executed plan.
timings_ns	object	Execution-core times. See Timings.
request_timings_ns	object	Whole-request times. See Request timings.
work	object	Budget-accounted work for the whole request. See Work.
cache_layers	array	Cache outcomes for the whole request. See Cache layers.
access_path	object	Structural index and scan use. See Access path.
scheduling	object	Observed concurrency. See Profile scheduling.
operators	array	One observation per operator invocation. See Operators.

Counters are unsigned 64-bit integers that saturate at their maximum instead of wrapping. Every time is in nanoseconds of monotonic wall time.

Timings

Field	Meaning
timings_ns.planning	Time to build the plan.
timings_ns.execution	Time to execute the physical plan.
timings_ns.rendering	Time to build the ordinary result rows.
timings_ns.total	Executor request-core time through ordinary result construction. It is not end-to-end tool latency.

Request timings

Field	Meaning
request_timings_ns.transport_queue_wait	Time from MCP request acceptance until analyzer execution starts.
request_timings_ns.workspace_readiness_wait	Time from MCP request acceptance until a deferred workspace is ready. Part of transport_queue_wait.
request_timings_ns.analyzer_admission_wait	Time spent waiting for an analyzer execution slot after the workspace is ready. Part of transport_queue_wait.
request_timings_ns.workspace_ready	Snapshot acquisition, including any wait for a deferred workspace build.
request_timings_ns.preparation	Remaining request setup before input decoding.
request_timings_ns.input_decode	Decoding the query JSON or query file.
request_timings_ns.query_execution	Running the query, including the execution core measured in timings_ns.
request_timings_ns.rendering_serialization	Rendering and serializing the tool response.
request_timings_ns.total	The whole request, including transport_queue_wait.

Work

work covers the whole request. Each operators[].work has the same fields and covers one operator invocation.

Field	Meaning
scanned_files	Source files scanned.
scanned_source_bytes	Source bytes scanned.
fact_nodes	Structural fact nodes charged to the execution budget.
pipeline_rows	Rows produced by pipeline steps.
examined_references	Reference candidates examined.
provenance_steps	Provenance steps recorded.
import_files_resolved	Files whose imports were resolved.
import_edges_resolved	Import edges resolved.
semantic	Semantic work. See below.

work.semantic fields:

Field	Meaning
materialization_attempts	Attempts to materialize a semantic artifact for a file.
unique_materialized_files	Distinct files materialized.
request_cache_hits	Semantic artifacts reused from the request-local cache.
source_bytes	Source bytes read for semantic materialization.
procedures	Procedures materialized.
blocks	Basic blocks materialized.
program_points	Program points materialized.
values	Values materialized.
allocations	Allocation sites materialized.
call_sites	Call sites materialized.
memory_locations	Memory locations materialized.
captures	Closure captures materialized.
source_mappings	Source mappings materialized.
evidence	Evidence records materialized.
gaps	Recorded gaps (facts the lowering could not model).
events	State events materialized.
control_edges	Control-flow edges materialized.
nested_entries	Nested entries materialized.
retained_bytes	Bytes retained by materialized semantic artifacts.
cumulative	Object. Cumulative semantic work per lane: retained admissions plus performed work. Holding and scanning the same bytes can count them twice, so this telemetry is not a bound.
resident_peaks	Object. Largest live resident census per lane.
root_work_peaks	Object. Largest work balance of any one root or prepass per lane.
resident_nested_entries_peak	Largest resident nested-entry count.
resident_retained_bytes_peak	Largest resident retained-artifact byte count.
traversal_steps	Semantic traversal steps taken.
budget_exhausted	Boolean. true when the separate semantic budget ran out.
typestate	Object, present only when typestate work happened.
value_flow	Object, present only when value-flow work happened.
type_flow	Object, present only when class-set type-flow work happened.

cumulative, resident_peaks, and root_work_peaks each carry the same sixteen lane keys: source_bytes, procedures, blocks, program_points, values, allocations, call_sites, memory_locations, captures, source_mappings, evidence, gaps, events, control_edges, nested_entries, and owned_text_bytes. In these objects, source_bytes counts admitted or scanned source bytes rather than the materialization reads in the top-level source_bytes field, and owned_text_bytes counts artifact-owned text bytes. resident_nested_entries_peak is the nested_entries lane of resident_peaks, and resident_retained_bytes_peak repeats the retained-artifact peak that retained_bytes reports.

work.semantic.typestate fields:

Field	Meaning
solves	Typestate solves run.
cache_hits	Solves answered from cache.
summary_hits, summary_misses	Procedure summary lookups that hit or missed.
summary_rejections	Summaries found but rejected.
summary_evictions	Summaries evicted from the bounded cache.
summary_recomputations	Summaries computed again.
reached_rows	Rows the solver reached.
findings, omitted_findings	Findings retained and findings omitted by the finding budget.
witnesses, omitted_witnesses	Witnesses retained and witnesses omitted.
witness_steps, witness_bytes	Size of the retained witnesses.
fixed_point_solves	Solves that reached a fixed point.
cancelled_solves	Solves stopped by cancellation.
budget_exhausted_solves	Solves stopped by the solver budget.
failed_solves	Solves that failed.
finding_budget_exhausted	Boolean. true when the finding budget ran out.

work.semantic.value_flow fields:

Field	Meaning
solves	Value-flow solves run.
cache_hits	Solves answered from cache.
reached_rows	Rows the solver reached.
meetings	Origin-to-observation meetings found.
sink_outcomes	Observation (sink) outcomes produced.
omitted_endpoints	Endpoints omitted by the endpoint cap.
witnesses, omitted_witnesses	Witnesses retained and witnesses omitted.
witness_expansions	Witness expansion steps taken.
witness_steps, witness_bytes	Size of the retained witnesses.
fixed_point_solves	Solves that reached a fixed point.
cancelled_solves	Solves stopped by cancellation.
budget_exhausted_solves	Solves stopped by the solver budget.
failed_solves	Solves that failed.
endpoint_truncated	Boolean. true when endpoints were omitted.
witness_truncated	Boolean. true when witnesses were omitted.

work.semantic.type_flow fields (the class_set and absent_member steps):

Field	Meaning
field_slot_builds	Fresh whole-workspace field-slot indexes built.
field_slot_memory_hits	Ready field-slot indexes reused from memory.
field_slot_persistence_hits	Complete stored field-slot indexes loaded and accepted.
field_slot_persistence_misses	Stored field-slot lookups that found no matching index.
field_slot_persistence_rejections	Stored field-slot rows found but rejected by validation.
field_slot_publications	Fresh field-slot indexes written to storage.
root_result_persistence_hits	Stored finding-free root results loaded and accepted.
root_result_persistence_misses	Stored root-result lookups with no exact key.
root_result_persistence_rejections	Stored root results found but rejected.
root_result_store_failures	Failures while reading or writing stored root results.
root_result_publications	Fresh finding-free root results written to storage.
solves	Distinct root transactions solved.
cache_hits	Solves answered from cache.
snapshot_cache_hits, snapshot_cache_misses	Snapshot cache lookups.
dispatch_cache_hits, dispatch_cache_misses	Dispatch cache lookups.
binding_cache_hits, binding_cache_misses	Binding cache lookups.
summary_cache_hits, summary_cache_misses	Summary cache lookups.
root_summary_cache_hits	Root relations restored from reusable summaries.
root_summary_observation_rejections	Restored root relations rejected because they lacked a meeting for a current sink.
published_summaries	Summaries published for reuse.
solver_attempts	Bounded solver runs. One root can need more than one.
summary_profile	Object, present only when nonzero. Counts reasons a reusable summary could not be prepared, looked up, or published: preparation_call_contract, preparation_dependency, preparation_source_limit, preparation_recursive_dependency, preparation_repository_failure, preparation_surface, lookup_procedure, lookup_source_behavior, lookup_source_partition, lookup_entry, lookup_relation, lookup_live_remap, publication_incomplete_result, publication_projection, publication_entry, publication_relation_shape, publication_dispatch_reads, publication_reused_conflict, publication_transfer, publication_dependency.
class_set_rows	class_set rows produced.
finding_rows	Finding rows produced.
witnesses, omitted_witnesses	Witnesses retained and witnesses omitted.
witness_expansions	Witness expansion steps taken.
witness_steps, witness_bytes	Size of the retained witnesses.
witness_truncated	Boolean. true when witnesses were omitted.
incomplete_roots	Roots whose analysis did not complete.
failed_solves	Solves that failed.

Cache layers

cache_layers always has nine entries, in this order. Each entry is {"layer": <name>, "metrics": {...}}. operators[].cache_layers has the same shape for one operator invocation.

layer	What it caches	metrics.kind
seed_result	Request-local seed results.	complete_value
seed_structural_facts	Structural facts read by seed scans.	structural_facts
inbound_reference	Inbound reference relations.	complete_value
outbound_reference	Outbound reference relations.	complete_value
incoming_call	Incoming call relations.	complete_value
outgoing_call	Outgoing call relations.	complete_value
import_forward	Forward import relations.	complete_value
import_reverse	Reverse import relations.	complete_value
direct_import_topology	The snapshot-local complete direct-import topology.	complete_value

complete_value metrics:

Field	Meaning
lookups	Cache lookups.
hits, misses	Lookups that found or did not find a value.
builds	Values built.
waits, wait_ns	Waits for another builder of the same key, and their total time.
complete_hits, incomplete_hits	Hits on complete and on incomplete values.
complete_builds, incomplete_builds	Builds that produced complete and incomplete values.
unknown_outcomes	Lookups whose completeness was unknown.
replayed_items	Cached items handed to the consumer before relation-specific filtering. This can exceed emitted rows.

direct_import_topology adds these fields:

Field	Meaning
cancelled	Builds stopped by cancellation.
unavailable	Lookups where the topology could not be built.
over_budget	Builds that exceeded their budget.
fallbacks	Requests that fell back to the request-local resolver.
build_files, build_edges	Files and edges in built topologies.
build_ns	Time spent building.
retained_bytes	Bytes retained by built topologies.

structural_facts metrics:

Field	Meaning
lookups	Structural-facts lookups by seed scans.
memory_hits	Lookups answered from memory.
persisted_hydrations	Lookups answered by loading stored facts.
extractions	Lookups that parsed and extracted facts.
unavailable	Lookups for which no facts were available.
unknown_outcomes	Lookups with an unknown outcome.
replayed_files	Files whose facts were replayed to the consumer.

Access path

Field	Meaning
selected	scan_only, posting:<terms> (for example posting:kind+name), or mixed when seeds used different paths. Posting terms are kind, name, role_name, and kwarg.
representation_version	Version of the structural index representation that was used. 0 for scan-only execution.
estimated_provider_files	Files in the structural provider, as estimated for index selection.
scoped_files	Files inside the query’s path and language scope.
scoped_fact_nodes	Total facts in the scoped files. Exact when an index supplies metadata; 0 for scan-only execution.
admitted_fact_nodes	Facts admitted by compatibility source filtering and charged to the execution budget. Comparable on both paths.
candidate_files, candidate_facts	Files and facts left after posting intersection.
selected_terms	Array of {"label", "candidate_facts"}: each posting term used and its candidate count before intersection.
source_verification_required	Boolean. true when candidates still needed exact source-anchor verification.
cache_ready_lookups	Lookups that found a ready index before they started.
materialized_files, materialized_fact_nodes	Files and fact nodes materialized.
inspected_source_bytes	Source bytes inspected.
examined_fact_nodes	Fact nodes examined by the matcher.
index_lookups, index_hits, index_misses	Structural index lookups and their outcomes.
index_builds	Structural index builds.
index_waits, index_wait_ns	Waits for an index build, and their total time.
index_cancelled	Index builds stopped by cancellation.
index_unavailable	Lookups where no index could be provided.
index_over_budget	Index builds that exceeded their budget.
scan_fallbacks	Seeds that fell back to scanning.
index_build_files, index_build_source_bytes, index_build_fact_nodes, index_build_facts_bytes	Size of the inputs to index builds.
index_build_ns	Time spent building indexes.
retained_bytes	Bytes uniquely retained by the access path.

Profile scheduling

Field	Meaning
scheduling.peak_concurrency	Highest number of operators that ran at once.
scheduling.bounded_dispatch	Object, present only when bounded parallel dispatch enqueued tasks.
bounded_dispatch.worker_limit	Maximum workers allowed.
bounded_dispatch.workers_spawned	Workers started.
bounded_dispatch.tasks_enqueued, tasks_started, tasks_completed	Task counts.
bounded_dispatch.tasks_observed_cancelled_before_start	Tasks that saw cancellation before they started.
bounded_dispatch.queue_wait_ns	Time tasks waited in the queue.
bounded_dispatch.budget_wait_ns	Time tasks waited for budget.
bounded_dispatch.coordinator_wait_ns	Time the coordinator waited.
bounded_dispatch.dispatch_overhead_ns	Time spent dispatching.
bounded_dispatch.peak_concurrency	Highest number of tasks that ran at once.

Operators

Each entry in operators describes one operator invocation. A shared plan node that runs for several branches has one entry per invocation.

Field	Type	Meaning
node	integer	Physical node ID.
branch	array of integers	Zero-based set-branch slots from the root to this invocation. Empty outside set operations.
operator	string	The physical operator. See Physical plan.
disposition	string	completed; skipped, when the operator did no work of its own because cancellation arrived first or a dependency was cancelled or halted; or cancelled, when the operator saw cancellation during its own work.
timings_ns.elapsed	integer	Operator-local time, excluding inline dependency execution.
timings_ns.total	integer	Inclusive time from entry until the operator returned.
timings_ns.dependency_execution	integer	Time spent running dependency subtrees inline.
timings_ns.dependency_wait	integer	Idle time waiting for an already-running scheduled dependency.
timings_ns.merge	integer	Time spent attaching branch provenance and diagnostics and combining sets.
timings_ns.scheduling_overhead	integer	Ready-queue, enqueue, and dequeue overhead.
input_rows	integer	Rows received from dependencies.
rows_visited	integer	Input rows the operator actually visited. Can be smaller than input_rows after cancellation or an early cap.
relation_expansions	integer	Relation expansions after relation-specific filtering, before de-duplication.
rows_discarded	integer or null	Exact discarded-row count for row-to-row operators. null for expansion operators.
output_rows	integer	Rows passed to the parent. A skipped operator can pass on a dependency’s partial rows.
temporary_capacity_bytes_lower_bound	integer	Lower bound of temporary container capacity. Excludes heap payloads owned by strings, paths, traces, and nested vectors.
work	object	Work charged to this invocation. See Work.
cache_layers	array	Cache outcomes for this invocation. See Cache layers.
terminations	array of strings	Reasons the operator did not consume all available work. See below.
operator_truncated	boolean	The operator clipped or did not fully produce its own output.
result_truncated	boolean	The result passed upward was incomplete.
result_cancelled	boolean	The result passed upward was cancelled.

terminations values:

Value	Meaning
cancellation_before_work	Cancellation arrived before the operator started its own work.
cancellation_during_work	Cancellation arrived during the operator’s own work.
dependency_cancelled	A dependency was cancelled.
dependency_pipeline_halted	A dependency halted the pipeline.
terminal_cap	A seed stopped at the row cap derived from the query’s terminal limit.
result_limit	The public result limit was reached.
execution_budget	The execution budget ran out.
pipeline_budget	The pipeline row budget ran out.
import_graph_budget	The import-graph budget ran out.
analysis_limit	An analysis cap or budget clipped the output, for example a reference, call-relation, semantic, typestate, value-flow, or taint limit.
unsupported_analysis	A language adapter or capability does not support the requested analysis.
analysis_incomplete	An analysis reported a partial or incomplete derivation.