v0.5.0

0.5.0 - 2026-09-18

• Reject duplicate --line-counts values in benchmark_latency.py,
  benchmark_semantic_index.py, and benchmark_routing.py before any
  measurement runs, since each size becomes a workload run id and a repeated
  size made --result fail with a duplicate-id error after the benchmark had
  finished. The workload result JSON Schema now pins canonical measurement
  names to their unit and rejects identical runs, matching the reference
  validator; run-id uniqueness beyond that stays with
  workload_results.validate_result, and the schema and documentation say so.
• Run on-demand semantic indexing on a pool bounded by
  EPHEMERAL_SEMANTIC_PREFETCH_WORKERS instead of one thread per capture, and
  cancel queued on-demand jobs when their capture is evicted or cleared, so
  hybrid searches that exceed the wait budget on captures that are then
  evicted no longer accumulate indexing threads beyond the bound. A hybrid
  search whose capture is evicted while it waits answers lexical-first within
  its budget instead of indexing the evicted capture inline, and
  get_buffer_stats reports queued on-demand jobs separately as
  semantic_index_on_demand_queued. Benchmarks shut their engines down so
  background indexing never delays process exit. A capture
  pulled out of the prefetch queue whose indexing then fails is marked failed
  and retries through the lazy path. Empty captures now report
  semantic_coverage (complete for semantic and hybrid searches,
  not-requested for BM25) like every other search response.
• Add a versioned, tool-agnostic workload result format
  (coding-agent-workload-result, format version 1) with a reference
  validator and JSON Schema in workload_results.py and
  workload_result.schema.json. Every benchmark, evaluation, and the Codex
  A/B runner accept --result PATH (- for stdout) to emit it alongside
  their existing reports, so latency, phase timings, output volume, token
  estimates, resource use, and success status can be compared across
  producers without reading producer-specific records.
• Add compare_workload_results.py, which compares two or more workload
  result documents (or single runs selected with PATH#RUN_ID) without
  rerunning them: it reports absolute and percentage deltas per run,
  measurement, phase, and statistic, classifies each as improved, regressed,
  changed, unchanged, missing, or incompatible, shows the parameter and
  environment differences that explain a delta, filters runs by label and
  metrics by name, and offers JSON output plus a --check mode that fails on
  regressions or non-success results for automated regression checks.
  Non-finite --tolerance values (nan, inf) are rejected as invalid
  input instead of failing with a traceback while writing the comparison.
• Add experiment groups and metadata to workload results: every benchmark,
  evaluation, and the Codex A/B runner accept --experiment GROUP,
  --metadata KEY=VALUE, and --redact KEY, recorded in an optional
  experiment block that consumers read alongside run summaries. Metadata
  keys that name credentials are always stored as [redacted]. Add
  list_workload_results.py to find result documents under directories and
  list or filter them by group, metadata, workload, and status (per document
  or per run, keeping failed and invalid documents visible), and teach
  compare_workload_results.py DIR@GROUP and DIR@GROUP,KEY=VALUE
  references plus an experiment differences line. run_agent_ab_experiment.sh
  now writes workload result documents tagged from AGENT_AB_EXPERIMENT and
  AGENT_AB_VARIANT.
• Add a semantic-index latency benchmark that reports ingestion, lazy embedding
  materialization, first and subsequent hybrid or semantic search timings,
  indexing throughput, and needle rank by capture size with the real model.
• Use the upstream fp32 ONNX export of bge-small-en-v1.5, registered with
  FastEmbed as BAAI/bge-small-en-v1.5-fp32, as the default embedding model.
  It produces identical vectors to FastEmbed's reduced-precision catalogue file
  but its kernels parallelize. A release-preparation comparison on a 16-vCPU
  Linux x86_64 host measured 69.1 semantic chunks per second and 1.85 seconds
  median indexing for the fp32 export versus 3.1 chunks per second and 41.15
  seconds for the catalogue file at 1,024 lines (about 22x); these are
  host-specific diagnostic measurements, not universal guarantees. The
  catalogue file remains selectable as BAAI/bge-small-en-v1.5. First use
  downloads about 130 MB instead of 67 MB.
• Add EPHEMERAL_EMBEDDING_THREADS to bound ONNX Runtime threads for embedding
  inference; the runtime default is fastest but uses every core.
• Index semantic embeddings over windows packed separately from the BM25
  sliding grid, bounded by EPHEMERAL_SEMANTIC_CHUNK_LINES (default 8),
  EPHEMERAL_SEMANTIC_CHUNK_BYTES (default 1024), and
  EPHEMERAL_SEMANTIC_CHUNK_OVERLAP (default 0), fuse hybrid results in line
  space so exact BM25 ranges are preserved, and report chunk_index per match.
  This stops embedding the two-line overlap twice and roughly halves first
  semantic or hybrid search latency on large captures.
• Replace best-effort semantic prefetch admission with a queue drained
  newest-first by the bounded worker pool, so ingestion bursts never silently
  leave captures unprefetched; searches index a still-queued capture inline,
  and diagnostics report queued and running counts.
• Enable semantic prefetch by default so the first semantic or hybrid search
  after a capture usually finds the index ready; set
  EPHEMERAL_SEMANTIC_PREFETCH=0 to restore lazy-only indexing.
• Bound first hybrid-search latency on very large captures with
  EPHEMERAL_SEMANTIC_WAIT_SECONDS (default 10): hybrid search waits at most
  the budget for a capture's semantic index, then returns BM25 results marked
  semantic_coverage: pending while a background job finishes indexing, so
  repeating the search is fully hybrid. Responses always report
  semantic_coverage (complete, pending, unavailable, or
  not-requested); semantic mode still waits for the index. Searches on a
  queued or unindexed capture now index it on a dedicated thread instead of
  inline, get_buffer_stats reports the budget and on-demand job count, and
  benchmark_semantic_index.py records the first search's coverage and needle
  rank separately from complete-index quality.
• Add durable phase-level executions with persisted outputs, metrics, status
  history, restart recovery, explicit retries, and unsafe-side-effect resume
  confirmation for long-running agent workflows, including recoverable compact
  responses when detailed execution metadata exceeds the tool response budget,
  bounded list-page fallbacks, owner checks, and crash-durable directory sync.

Verification

• Distributions and SHA256SUMS were produced by the tagged release workflow.
• Build workflow and provenance attestation

v0.4.0

0.4.0 - 2026-09-15

• Warm the embedding model asynchronously after server readiness by default,
  expose warm-up readiness and failures, and preserve lexical hybrid results
  when semantic initialization is unavailable.
• Add a warm-up benchmark covering startup readiness, first-query latency, and
  process RSS impact, with an opt-out for lexical-only deployments.
• Add explicit versioned length-prefix framing for socket requests and
  responses, including bounded and malformed-frame handling.
• Add opt-in runtime semantic-index budget adjustment with deterministic LRU
  eviction, diagnostics, concurrency coverage, and documentation.
• Add opt-in, session-scoped data-path byte counters to local diagnostics for
  capture input/retention, tool responses, and framed socket traffic.
• Extend the Codex A/B records to collect per-run data-path byte counters from
  content-free MCP metrics snapshots.
• Persist benchmark metrics snapshots after each MCP tool call so subprocess
  shutdown still leaves counters available for comparison runs.

Verification

• Distributions and SHA256SUMS were produced by the tagged release workflow.
• Build workflow and provenance attestation

v0.3.1

0.3.1 - 2026-09-13

• Package Codex and generic agent session launchers alongside ephbuf.
• Allow explicitly configured stdio-only operation when the host denies Unix
  socket creation while preserving dual-mode startup by default.

Verification

• Distributions and SHA256SUMS were produced by the tagged release workflow.
• Build workflow and provenance attestation

v0.3.0

0.3.0 - 2026-09-11

• Harden capture, search, consolidation, socket, timeout, shutdown, and
  process-group handling with stricter bounds and cleanup guarantees.
• Improve diff parsing, fallback tokenization, hybrid search context handling,
  semantic prefetch behavior, and diagnostic signal accuracy.
• Add repository-shaped agent A/B evaluation fixtures, repeatable Codex
  experiment tooling, provider-usage diagnostics, and privacy-safe baselines.
• Expand release, benchmark, lifecycle, and operational coverage and refresh
  the documented effectiveness measurements.

Verification

• Distributions and SHA256SUMS were produced by the tagged release workflow.
• Build workflow and provenance attestation

v0.2.0

0.2.0 - 2026-09-07

• Add bounded consolidate_captures support for building one searchable JSON
  view over multiple active captures while preserving source IDs and line
  numbers.
• Add opt-in, content-free local usage metrics and aggregate diagnostics for
  observing capture, search, retrieval, eviction, cleanup, and process events.
• Add paired and consolidated effectiveness benchmarks, including reproducible
  README results that quantify context-size reductions and document benchmark
  limitations.

Verification

• Distributions and SHA256SUMS were produced by the tagged release workflow.
• Build workflow and provenance attestation

v0.1.2

0.1.2 - 2026-09-06

• Add opt-in runtime diagnostics for content-free version, platform, socket,
  buffer, embedding, and memory reporting.
• Add privacy-safe structured operational logging for timeouts, process
  termination, embedding readiness and failures, limits, eviction, cleanup,
  and socket failures.
• Document field-observation procedures and add a privacy-conscious GitHub
  issue form for reporting real-world behavior.
• Expand project guidance and defensive-path coverage for ongoing maintenance.

Verification

• Distributions and SHA256SUMS were produced by the tagged release workflow.
• Build workflow and provenance attestation