snagline 0.1.0

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog,
and this project adheres to Semantic Versioning.

[0.1.0] - 2026-08-27

This is the first tagged release. It comprises 87 merge commits on origin/master
from the first commit through d80a686 (feat --semantic baseline flag, PR #189).
Every user-visible change below was verified against git show <sha> --stat for
its merge commit, not against PR titles. CI at this commit is green with
reproduced counts: 651 passed, 3 skipped (pytest, py3.10 through 3.13),
88.04% line coverage, ruff check src tests and ruff format --check src tests
clean, mypy src clean. Overhead measured on this checkout:
benchmarks/overhead_benchmark.py reports median 5.31 us/step, p99 42.22 us/step
over 200,000 synthetic steps. Detection accuracy measured on this checkout:
benchmarks/detection_accuracy.py reports macro-F1 1.000 over 76 episodes
(40 labeled, 36 healthy controls), zero healthy-control false positives.

Added

Core and configuration

• Canonical schemas StepEvent / EpisodeMeta / FailureRisk and make_signature
  with SHA-256 normalization (PR #24 through #30, #15, #28).
• Monitor orchestrator with fail-open guarantee, per-episode lock sharding via
  StateBackend / MemoryStateBackend, and Metrics self-observability counters
  (PR #44, #47).
• Config dataclass with 12-factor env/file layering (SNAGLINE_* env vars,
  Config.resolve, optional JSON/TOML file), tunable thresholds for every
  detector, and CLI wiring (snagline --config) (PR #30, #36, #37).
• BaselineProfile fitting and persistence (fit_baseline_from_jsonl,
  save_baseline / load_baseline, BaselineStore versioned store per
  tenant/deployment) (PR #24, #45, #46).
• Monitor.snapshot / restore with versioned JSON, atomic tmp+replace,
  strict composition check, and per-detector dump_state / load_state
  (PR #168).

Detectors

• Loop detector with sliding window and repeat threshold, plus opt-in
  hardening modes: near-duplicate (volatile ID collapsing), cycle (periodic
  scan), stall (consecutive identical signatures) (PR #114, ce5e449).
• Error-cascade detector with windowed and consecutive modes (core), with
  config cascade_count_non_tool_errors (PR #16-era, #27 baseline).
• Latency-anomaly detector (Welford + CUSUM, stdlib only) with per-tool
  baselines, sigma floors, warm-up, and optional baseline re-fit
  (cusum_refit_every) (PR #25, #138, #167).
• Goal-drift detector (opt-in, compares live run to persisted
  BaselineProfile on error rate and latency z-score) (PR #25).
• ML ensemble MLOrchestrator (opt-in, noisy-OR over tier-1 detectors,
  model= hook) (PR #26).
• Token-runaway detector (opt-in, CUSUM over token volume plus
  episode_token_budget envelope, triggers token_runaway / budget_breach)
  (PR #107).
• Meltdown detector (opt-in, sliding-window Shannon entropy, low/high
  thresholds) (PR #107, #135).
• Silent-abort detector (opt-in, EpisodeFinalizer evaluated at
  end_episode, trigger silent_abort) (PR #107).
• Stagnation detector (opt-in, novelty-rate collapse) (PR #127).
• Side-effect guard (opt-in, duplicate non-idempotent action detection,
  side_effect field on StepEvent) (PR #140).
• Compaction tripwire (opt-in, governance-decay across context
  compactions, compaction / constraint_present contract) (PR #147).
• Horizon-scale time axis (opt-in, PR #167): max_episode_wall_seconds
  with wall-clock budget (wall_clock_budget), idle_warn_seconds
  (idle_gap), window auto-scaling (window_scale_steps, max_window),
  HeartbeatSink and snagline watch --follow --heartbeat liveness file.
• ML extra: ESN ensemble (ml/esn_ensemble.py, snagline[ml],
  one-class ESN + CUSUM + Mahalanobis baseline) (PR #135).
• Drift extra: semantic goal-drift (drift/goal_drift.py,
  snagline[drift], sentence-transformers, PR #81 / 09ae888).
• Auto-calibration from BaselineProfile (calibration="auto",
  CalibrationPlan, resolve_baseline_profile) (PR #138).
• Scheduled baseline retrain snagline baseline retrain contract with
  --windows-dir / --jsonl / --store-dir / --max-age staleness guard
  and docs/RETRAIN_CADENCE.md (PR #126, #102).
• Enforcement policy Monitor(policy="observe"|"callback"|"halt_webhook")
  with on_risk callback (fail-open), halt_url/halt_timeout_s
  (default 250 ms) / min_severity_for_halt (default 0.8),