Skip to content

CHANGELOG.md

Latest commit

 

History

History
282 lines (187 loc) · 12.1 KB

CHANGELOG.md

File metadata and controls

282 lines (187 loc) · 12.1 KB

Changelog

All notable changes to SOMA. Format follows Keep a Changelog.

Versioning

SOMA uses CalVer in YYYY.M.D form, where each segment is the date of release (no zero-padding — PEP 440 canonicalizes leading zeros away anyway). Multiple releases on the same date append .N: 2026.4.30, 2026.4.30.1, 2026.4.30.2, ….

Versions before 2026.4.30 used a YYYY.M.PATCH counter where the middle segment was a sequential minor, not a month — confusing, and abandoned in favor of the date scheme. Historic CHANGELOG entries (2026.6.x, 2026.5.x, 2026.4.x pre-30) reflect that older numbering and are left as-is.

[2026.4.30] — 2026-04-30

First release under the new date-based CalVer scheme. Aggregates everything done since the deleted 2026.6.2 artifacts.

Added

  • tools/hook_bench.py — hook-latency benchmark scaffold.
  • dashboard: SOMA_DASHBOARD_TOKEN env-gated bearer-token auth on every API endpoint. Off by default for backwards compatibility; banner on stderr at startup announces the state.
  • calibration: schema_version migration framework. Older profiles run through registered migrators; newer-than-build profiles refuse to overwrite and log under SOMA_DEBUG=1.
  • persistence: update_engine_state one-shot helper around the existing engine_state_transaction (load → mutate → save under flock).

Changed

  • versioning: switch to date-based CalVer (this entry).

Fixed

  • analytics: SQLite trigger rejects NULL firing_id inserts into ab_outcomes (defense in depth for the firing-id contract).
  • analytics: hot-path indexes on guidance_outcomes(pattern_key, ts) and partial ab_outcomes(pattern, arm).
  • ab: purge orphaned ab_counters.json.tmp.* files on reset_counters (left behind by hook subprocesses that crashed between fsync and os.replace).
  • cli: friendly error on invalid mode argument; restore wheel build path.
  • errors: log_silent_failure helper wired into ~14 swallow sites in data-integrity / write paths. Gated on SOMA_DEBUG=1, silent by default.
  • config: moved config_loader from cli/ to top-level soma.config (no public API change).
  • audit: prune rotated audit logs past retain_rotated.
  • guidance: 1-hour TTL on the healing-transition cache, on time.monotonic() so NTP corrections can't freeze it.
  • dashboard: token comparison via hmac.compare_digest (closes timing oracle).
  • persistence: prune stale agents from engine_state.json on load.
  • tunables: behavioral thresholds centralized in soma.tunables.

Removed

  • Deprecated packages/soma-ai/ TypeScript stub.
  • All in-repo markdown docs (re-authored from scratch).

Notes

  • All historic GitHub Releases and PyPI artifacts (2026.4.x–2026.6.2) were withdrawn before this release. Tags and remote releases are gone; the codebase carries forward.

[2026.6.2] — 2026-04-29

Reliability + performance hotfix on top of 6.1.

Performance

  • hooks: in-process compile() instead of py_compile subprocess; tighten lint timeouts to 500 ms. Cuts hook-tool RTT in the common path.

Fixed

  • analytics: set busy_timeout=5000 + synchronous=NORMAL on SQLite connections. WAL mode already on.
  • ab: atomic write for ab_counters.json (tmp + fsync + os.replace under flock). Prior non-atomic write could corrupt counters under concurrent hook subprocesses.
  • guidance: short-circuit retired patterns in check_followthrough (no-op for _stats, drift, entropy_drop, context).
  • errors: SomaBlocked and SomaBudgetExhausted now correctly inherit from SOMAError.
  • tools: phantom_smoke.py resolves paths relative to script + reads env-overridable database location.

[2026.6.1] — 2026-04-29

Stop-ship hotfix: four security + correctness fixes immediately after 6.0.

Security

  • hooks: reject flag-shaped file_path (rejects --evil masquerading as a file path) and add -- separator before the path argument in the hook subprocess invocation.
  • cli: function-level allowlist for the --definition SQL identifier in validate-patterns (no SQL identifier injection through CLI args).

Fixed

  • release-gate: exclude NULL firing_id rows from _arm_counts so the count matches the t-test filter — gates were previously letting through under-powered comparisons.
  • ab: drop h=2 INSERT when pressure_after_h1 is NULL (closes the B1-class bias at horizon 1).

[2026.6.0] — 2026-04-27

The "proof pipeline" release: multi-horizon A/B recording, multi-definition guidance outcomes, dropguard, and pattern retirement after audit.

Added

  • ab: multi-horizon recording — INSERT at h=2 (or h=1), UPDATE-by-firing_id at h=5 and h=10. Outcomes can be analysed at any horizon without rerunning sessions.
  • analytics: firing_id column added to ab_outcomes + per-horizon pressure columns (pressure_after_h1, pressure_after_h5, pressure_after_h10).
  • guidance: compute_multi_helped — three orthogonal "helped" definitions persisted side-by-side (helped_pressure_drop, helped_tool_switch, helped_error_resolved).
  • analytics: guidance_outcomes columns for the three helped definitions.
  • cli: validate-patterns gains --horizon and --definition flags.
  • dashboard: pattern cards surface multi-helped breakdown.
  • feat: P2.1 — bash_error_streak predictor in OBSERVE mode.
  • feat: P2.2 — A/B coverage release gate blocks under-sized ships.
  • analytics: migration archives pre-firing_id legacy ab_outcomes rows into a quarantine table; validate-patterns excludes them.
  • tools: phantom_smoke runner — end-to-end pipeline verification without an LLM.

Changed

  • wrap + hooks: double-wrap guard + cross-IDE agent_id family routing.
  • setup: refuse to overwrite a corrupt settings.json; atomic write with .bak.

Fixed

  • ab: idempotency token on should_inject — kills A/B counter rebias on retried hook invocations.
  • budget: removed silent spend clamp that was suppressing the cost_spiral signal.
  • state: atomic writes for predictor / quality / fingerprint / task_tracker persistence.
  • hooks: flock around circuit_<aid>.json read-modify-write; wrap followthrough RMW block in circuit_transaction.
  • hooks: surface PostToolUse exceptions on stderr instead of silent pass.
  • baseline: from_dict defaults match the constructor.
  • Multiple corrections from independent code-review (pre-firing slice, firing_id collision window, RMW rollback, timeout pressure).

Performance

  • statusline: mtime-keyed cache with 1-second TTL; cuts repeated reads when the user holds the prompt open.

Removed

  • Patterns retired after audit:
    • entropy_drop — hardcoded avg_gap < 3.0 panic threshold fired during fast Read/Glob exploration loops; historical precision <20%. Underlying entropy signal stays.
    • context — helped was structurally biased toward 0% because it required the agent to literally write next.md or run git compact. Underlying context-usage signal stays.

[2026.5.5] — 2026-04-23

Block-randomized A/B + data hygiene.

Added

  • ab: block-randomized A/B harness — per-firing assignment to treatment / control, randomization keyed on firing_id (not session_id) to prevent intra-session bleed.

Fixed

  • Silence-cache reset that was blocking the A/B data pipeline.
  • Silence refresh now filters pre-reset guidance_outcomes so resets actually clear pattern-silence state.
  • Test-agent write guard prevents test_* agents from polluting analytics; purge migration in analytics removes prior pollution.

[2026.5.4] — 2026-04-19

A/B measurement bias + control-arm contamination fix.

Fixed

  • A/B measurement bias caused by control arm receiving the same context-injection as treatment under a specific code path.
  • Control-arm cleanup migration to remove contaminated rows.

[2026.5.3] — 2026-04-19

A/B proof pipeline + stricter followthrough.

Added

  • A/B proof pipeline: ab_outcomes table, control-arm sham firings, paired-difference reporting, and the validation CLI to surface gate status per pattern.

Changed

  • Followthrough scoring now requires stricter behavioural evidence to count as "helped".

[2026.5.2] — 2026-04-19

Round-2 audit fixes — second-pass corrections after independent review of 2026.5.1.

Fixed

  • A series of correctness fixes from the round-2 audit (specific issue ids in commit messages).

[2026.5.1] — 2026-04-19

Audit-fix release — first-pass corrections after self-review of 2026.5.0.

Fixed

  • Correctness regressions surfaced by the post-release audit.

[2026.5.0] — 2026-04-19

Self-Calibration + Strict Mode. Per-agent warmup, adaptive thresholds, and a hard-block tier for destructive ops.

Added

  • calibration: CalibrationProfile + persistence — per-agent baselines persist across sessions.
  • calibration: warmup gate — guidance disabled until enough actions have been observed for the per-agent baseline to be meaningful.
  • calibration: personal thresholds in pattern checks — pattern triggers consult the per-agent profile, not a global constant.
  • calibration: adaptive auto-silence refresh loop — patterns that consistently fail to help are silenced automatically; silence is re-evaluated on a refresh interval.
  • blocks: BLOCK mode state + soma unblock CLI.
  • strict-mode: PreToolUse enforcement + block lifecycle (issue, expire, clear).
  • signal-pruning: dropped _stats user-facing emission. Highest fatigue source in 5.x; underlying signal collection unchanged.
  • visibility: surfaced SOMA's work to the user explicitly via the status line and dashboard.

[2026.4.3] — 2026-04-18

Clean replacement for the yanked 2026.4.2.

Fixed

  • Drift pattern precision: removed false-positive triggers on intentional context shifts.

[2026.4.2] — yanked

Pulled from PyPI immediately after release; replaced by 2026.4.3.

[2026.4.0] — 2026-04-17

Switch to CalVer. Same wire format as 0.7.x; new versioning scheme.

Changed

  • Version scheme: SemVer → CalVer (YYYY.M.PATCH). Reset patch numbering.

[0.7.0] — 2026-04-15

Dashboard rebuild + Smart Guidance v2.

Added

  • Dashboard: full rebuild on FastAPI + SSE with new pattern-card surface, ROI panel, replay tab.
  • Smart Guidance v2: pattern engine refactor — priority dict, persistent cooldowns, per-pattern severity, healing-suggestion dictionary.
  • Audit logging: every guidance firing recorded with evidence + outcome.
  • Cost-template guidance.

[0.6.3] — 2026-04-14

Changed

  • Pressure system recalibrated: tighter thresholds, less false-positive guidance.

[0.6.2] — 2026-04-12

Changed

  • Repo cleanup; docs overhaul.

[0.6.1] — 2026-04-12

Added

  • Dashboard v2 (predecessor to the 0.7.0 rebuild).

Changed

  • Repo cleanup pass.

[0.6.0] — 2026-04-02

Mirror: proprioceptive behavioural feedback. First version where the agent reads its own state and changes course mid-session — the closed-loop primitive that everything since has built on.

Added

  • Mirror loop: state read → guidance → context injection → next action.

[0.5.0] — 2026-03 (mid)

Added

  • Documentation overhaul.
  • Async/streaming hardening.

[0.4.0] – [0.4.12]

Added

  • 0.4.0 — initial guidance system release.
  • 0.4.6 — workflow-aware severity.
  • 0.4.7 — positive feedback (reinforce when the agent does the right thing).
  • 0.4.9 — directive injections + audit fixes + coverage.
  • 0.4.11 – 0.4.12 — core polish, multi-agent ready.

[0.3.x] and earlier

  • 0.3.1 — auto-install slash commands via setup-claude.
  • 0.3.0 — Claude Code plugin: slash commands and auto-registered hooks; PyPI publish workflow with trusted publishing.
  • 0.2.x — 9 new CLI commands + autonomy modes fully connected.
  • 0.1.x — project skeleton: types, ring buffer, first 11 tests.

Versioning policy: CalVer minor (YYYY.M.X) bumps for any user-visible change, patch for hotfixes, retroactive yanks for shipped regressions (see 2026.4.2).