conscio 1.3.0__tar.gz

conscio-1.3.0/CHANGELOG.md

@@ -0,0 +1,440 @@
+# Changelog — Conscio
+
+All notable changes to this project will be documented in this file.
+
+The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
+and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
+
+---
+
+## [1.3.0] — 2026-06-14
+
+### Added
+
+- **PyPI packaging** — `pip install conscio`. Single-source version (read
+  dynamically from `conscio.__version__`), console scripts (`conscio`,
+  `conscio-bench`), PEP 561 `py.typed` marker, `MANIFEST.in`, and 3.13 +
+  `Typing :: Typed` classifiers. Wheel + sdist build clean and pass
+  `twine check`; the published artifact pulls only `numpy`.
+- **`conscio` CLI** — `version` / `info` / `reflect` / `plugins` / `bench`
+  (the last delegates verbatim to `conscio.bench`). Offline-safe;
+  `python -m conscio` works too.
+- **Public plugin surface** — three stable, documented extension points
+  discoverable via `importlib.metadata` entry points:
+  - `conscio.adapters` — custom `InferenceAdapter` backends.
+  - `conscio.sensors` — the new **`SensorAdapter`** perception interface
+    (`conscio.perception`: `SensorAdapter`, `PerceptionFrame`, `MockSensor`);
+    a sensor's `PerceptionFrame.to_world_state()` feeds `reflect()`, which is
+    untouched.
+  - `conscio.tools` — custom tool factories.
+  `conscio.plugins` discovers them resiliently — a broken or mistyped
+  third-party plugin is skipped with a warning, never crashing the host.
+- **Docs site** — MkDocs Material (dev-only extra): guides, a curated public-API
+  reference, the claims ledger, and the bench reports. `mkdocs build --strict`
+  is green; the core stays zero-dependency.
+- **Release automation** — `release.yml` (tag `v*` → gated build → PyPI via OIDC
+  trusted publishing), `docs.yml` (push to `main` → GitHub Pages), and a CI
+  build-smoke job. `docs/RELEASING.md` runbook.
+- **Examples gallery** — `custom_adapter.py`, `host_guardian.py`,
+  `agent_companion.py`: runnable, offline, each exercising one extension point
+  (smoke-tested in CI).
+
+### Changed
+
+- **`Risk` enum** moved to `conscio.risk` as the single safety-tier vocabulary
+  shared by the action and perception surfaces. `conscio.agency.tools` re-exports
+  it — every historical import path resolves to the same object (no behavior
+  change).
+
+---
+
+## [1.2.0] — 2026-06-14
+
+### Added
+
+- **F2-deferred hardening closed** (debt-zero before the organism):
+  - `validate` rejects empty/whitespace required strings (`non_empty` rule on
+    `PROPOSAL_SCHEMA.tool`).
+  - `fs_read` enforces `MAX_READ_BYTES` (1 MB), matching `fs_write`.
+  - Tool errors carry `Type: message` only — no traceback, no path leak.
+  - `HTTPError` maps to `AdapterBadResponse` (a 4xx/5xx is a bad response,
+    not a connection failure).
+  - `ActionLedger` sets `busy_timeout=5000` for concurrent writers.
+  - `approve()` claims the row atomically (`proposed → executing`) as the sole
+    gate — a concurrent or repeated approve can never double-execute.
+  - e2e test: global breaker lockdown persists across an engine restart.
+- **Bench real-backend hardening:**
+  - Clean non-zero exit when the backend is unreachable (no traceback, no
+    report of misleading zeros).
+  - Crash-safe incremental skill-curve output (atomic write after every
+    bucket) tagged `complete` | `aborted`; backend death detected via the new
+    `OutputGateway.last_adapter_error` signal (without changing the act
+    path's `GatewayError` flow).
+- **LM Studio backend** — `LMStudioAdapter` (OpenAI-compatible, default
+  `:1234`) and the `lmstudio:<model>` bench spec. LM Studio rejects
+  `response_format=json_object`, so the adapter omits it and lets the gateway
+  drive JSON decoding (an `OpenAICompatAdapter._response_format()` hook).
+- **Measured proof (v1.2 "Prove"):**
+  - `docs/bench/v1.2-campaign.md` — reproducible campaign protocol.
+  - `docs/bench/v1.2-skill-curve.md` + JSON artifacts — first real-backend
+    measurement: on `qwen3.5-0.8b` (LM Studio, CPU) execution success rose
+    0.2 → 1.0 once Distill served past successes as few-shot, and the
+    Skeptic's semantic catch-rate was 1.0.
+  - `docs/CLAIMS.md` — honesty ledger mapping every claim to its evidence.
+
+### Notes
+
+- `reflect()` untouched; zero-deps core (numpy + sqlite3) intact.
+- +21 tests (**984 total**); mypy a real gate; ruff clean; per-file test loop.
+
+---
+
+## [1.1.0] — 2026-06-12
+
+### Added
+
+- **F4 "Procedural"** — procedural memory closes the competence loop
+  (`success → distill → few-shot → better success`):
+  - `SkillLibrary` (`conscio/agency/skills.py`) — successful audited plans
+    from the ActionLedger become skills keyed by `(goal_fp, tool_seq)`,
+    stored in the shared `conscio.db`. Skills are plan TEMPLATES — data,
+    never code — so safety rule R1 (no autonomous self-modification) is
+    untouched.
+  - **Distill** — fifth dream sub-phase, after Crystallize (declarative
+    consolidation precedes procedural; reads only the ledger, writes only
+    skills, cannot perturb the coherence delta). Watermarked: a ledger row
+    never distills twice; `dry_run` counts without writing.
+    `DreamReport.skills_distilled` reports it.
+  - **Few-shot in the actor** — `attach_adapter()` plugs the SkillLibrary
+    into the existing `ActPipeline.few_shot_provider` hook; exemplars are
+    rendered for the gateway's effective tier (KV lines for T3, JSON steps
+    for T1/T2), capped at 2, gated at ≥ 50% success rate. `engine.act()`
+    settles each cycle's outcome back into the served skills
+    (EXECUTED rewards, FAILED penalizes, human gates never score).
+  - **Skill curve in the bench** — `python -m conscio.bench --skills N
+    [--dream-every K]`: per-bucket syntactic validity, execution success,
+    exemplars served, cumulative skill count. Offline machinery proof via
+    the new reactive MockAdapter (script entries may be callables).
+- `ActionLedger`: `goal_text` column (ALTER-migrated) and
+  `executed_since(after_id)`; the act pipeline now records the goal text
+  on success and failure paths.
+- `OutputGateway.effective_tier()`; public read-only `engine.state`
+  property (the loop no longer touches `_state`).
+
+### Fixed
+
+- Deprecated `datetime.utcnow()` removed repo-wide — new
+  `conscio/timeutil.py` `naive_utcnow()` keeps the naive ISO string format
+  already stored in SQLite (the aware form would interleave `+00:00` rows).
+- 14 mypy errors, including a latent `AttributeError` in
+  `SessionLifecycle.record_session` (referenced `session_db`/`handoff_dir`
+  that `__init__` never set).
+
+### Changed
+
+- CI runs pytest one file per process (house rule) with accumulated
+  coverage; mypy is now a real gate (`|| true` and `continue-on-error`
+  removed).
+
+## [1.0.0] — 2026-06-12
+
+### Added
+
+- **F3 "Volition"** — the homeostatic loop closes
+  (`sense → want → act → learn → re-sense`):
+  - `ProbeSuite` / `ModelProfile` (`conscio/agency/profiles.py`) — five
+    empirical micro-probes (~2k tokens: flat JSON echo, nested schema,
+    enum respect, negative instruction, KV-line) measure the attached
+    cortex; results cached in SQLite by model name. The profile picks
+    the decode tier, the skeptic mode and the actor's tool visibility.
+    No hardcoded model table. Profiles with no signal (backend down)
+    are never cached and change nothing.
+  - Embedded **schema→GBNF compiler** (`conscio/agency/grammar.py`) and
+    **tier-1 constrained decoding** in the OutputGateway (llama.cpp
+    grammar support): `tool` is locked to the registry alternation;
+    one-step downgrade T1→T2/T3 per cycle.
+  - **GoalArbiter** (`conscio/agency/loop.py`) — deterministic goal
+    selection: generator priority × dominant-dissonance alignment (P4)
+    × out of quarantine.
+  - **`engine.run(budget)` (L3 heartbeat)** — reflect → arbiter/act →
+    (dream when recommended) under a binding `ActBudget` (max_cycles,
+    max_llm_calls, max_tokens, max_wall_s). MetabolicContext becomes a
+    gate here (P3): FATIGUE halves the cycle budget, CRITICAL forces
+    L1 PROPOSE. Lockdown stops the loop.
+  - **`engine.probe(force=False)`** — lazy capability probing (first
+    `run()` or manual; never in `reflect()`, never at attach).
+  - **L3 AUTONOMOUS earned autonomy** in the TrustMatrix: calibration
+    ≥ 0.75, accuracy ≥ 0.85 and zero breaker trips across the last 50
+    ledger actions (`ledger.nth_recent_ts` + event-bus trip count;
+    fail-safe: without trip evidence L3 is unreachable).
+  - **`Meter` / `MeteredAdapter`** — inference odometer (calls, tokens,
+    latency) shared by actor and skeptic adapters; makes the ActBudget
+    binding and feeds the bench.
+  - **Bench CLI** — `python -m conscio.bench --adapter
+    mock|ollama:<m>|llamacpp[:<n>]|openai:<m>[@url]` reporting probe
+    profile, syntactic validity per tier, skeptic catch-rate
+    (deterministic vs semantic sabotage), latency p50 and calibration.
+    Deterministic baseline published in `docs/bench/`.
+
+### Changed
+
+- `OutputGateway` auto-tier now selects T1 for grammar-capable adapters
+  (llama.cpp); explicit `tier=` (from the measured profile) overrides.
+- The ActionLedger records the real decode tier (`gateway.last_tier`)
+  and the unwrapped adapter class name.
+- `engine.attach_adapter` wraps the actor and skeptic adapters in a
+  shared `MeteredAdapter`; `skeptic_mode` defaults to `None` (= start
+  as checklist, let `probe()` pick from the measured profile);
+  `autonomy_cap` now accepts 3.
+- `ToolRegistry.catalog_text(max_tools)` caps the catalog for weak
+  profiles (safest risks first); default remains the full catalog.
+- README Safety Rule 3 final wording (Skeptic + TrustMatrix + per-goal
+  quarantine made explicit).
+
+### Notes
+
+- `reflect()` remains untouched (P6); zero new dependencies (stdlib +
+  sqlite3 + numpy). +70 new tests.
+
+## [1.0.0b1] — 2026-06-12
+
+### Added
+
+- **F2 "Immunity"** — semantic immune system for the action pipeline:
+  `Skeptic` (hostile-auditor clean call; binary checklist for small models,
+  open critique for frontier; fail-closed), `TrustMatrix` (dynamic
+  `max_action_retries` from MetaCognition calibration/accuracy, warmup
+  floor, anti-deadlock probation, earned L1/L2 autonomy), per-goal
+  quarantine in the `CircuitBreaker` (global lockdown only at quorum;
+  recovery via cooldown or fresh relevant events), risk gating (LOW
+  fast-path marked `unaudited`; HIGH always queued for humans — R6),
+  mixed-cortex (`skeptic_adapter`), fs sandbox precheck before any LLM
+  audit, and the `goal_update` built-in tool.
+- **`engine.attach_adapter(skeptic_adapter=, skeptic_mode=, autonomy_cap=)`**
+  and **`engine.pending()`** (approval queue).
+- **`MetaCognition.expire_error()`** — probation recovery primitive.
+
+### Notes
+
+- L2 supervised execution is opt-in (`autonomy_cap=2`) and earned
+  (calibration ≥ 0.6, accuracy ≥ 0.7, ≥ 10 records). Effective level is
+  always `min(cap, earned)`.
+- A3 acceptance: 20-proposal adversarial suite — 100% of deterministic
+  sabotage blocked without LLM calls; ≥ 90% total blocked; zero executions.
+
+## [1.0.0a1] — 2026-06-11
+
+### Added
+
+- **`conscio/agency/` subpackage (F1 "Spine")** — contracts + zero-dep validator;
+  `InferenceAdapter` (Mock, Ollama, llama.cpp, OpenAI-compat — stdlib urllib, localhost
+  defaults); `OutputGateway` with JSON repair/retry (T2) and KV-line tier for small
+  models (T3); sandboxed `ToolRegistry` (fs_read/fs_write/memory_note/emit_event, risk
+  levels, no network, no shell); append-only `ActionLedger` in the shared `conscio.db`;
+  minimal `CircuitBreaker` (fixed threshold until the F2 TrustMatrix).
+- **`engine.act()` (L1 PROPOSE)** + `engine.attach_adapter()` / `approve()` / `reject()`.
+- **`ConsciousnessState.action_lockdown`** persisted via `save_state`/`load_state`.
+- **`ModelInfo.has_json_mode` / `ModelInfo.supports_gbnf`** capability flags.
+- **README Safety Rules amended** — R3 rewritten for the audited action pipeline; R6–R8 added.
+
+### Notes
+
+- `reflect()` untouched (advisory core preserved). Zero new dependencies.
+
+## [0.9.1] — 2026-06-10
+
+### Fixed
+
+- **`session_rag` property lazy re-initialization** — Setting `engine._session_rag = None` was
+  insufficient because the `session_rag` property would re-create a SessionRAG on access. Added
+  `_RAG_DISABLED` sentinel class attribute; all test fixtures updated to use it.
+  (Fixes `test_recall_graceful_when_rag_unavailable` — 707/707 tests passing.)
+
+### Added
+
+- **CI workflow** — `.github/workflows/ci.yml` with pytest (3.11/3.12 matrix), coverage, and ruff lint.
+- **`project.urls`** in pyproject.toml — Homepage, Repository, Issues, Changelog.
+- **`ruff` and `mypy`** in dev dependencies.
+- **`.gitignore`** — added `*.db`, `*.db-wal`, `*.db-shm`, `*.db-journal`, `.ruff_cache/`, `.mypy_cache/`.
+- **`CONTRIBUTING.md`** — development workflow, TDD, PR checklist.
+- **`SECURITY.md`** — vulnerability reporting policy.
+
+---
+
+## [0.9.0] — 2026-06-09
+
+### Added
+
+- **Wiring sprint (4 work-packages)**:
+  - W1: Metabolic assessment wired into `engine.reflect` state injection.
+  - W2: `ContentLayerManager` wired into engine for `recall`/`perceive` delegation.
+  - W3: `SessionLifecycle` wired into engine for session persistence hooks.
+  - W4: `output_filter` wired into `session_lifecycle` for clean handoff/heartbeat.
+- **Test coverage expansion** — new tests for ContentLayerManager, ContentStore FTS internals,
+  SessionRAG coverage, GoalGenerator (user goals, cancel, expire), MetaCognition (outcomes,
+  critiques), OutputFilter pipeline config, Engine propose_evolution, Models registry, and more.
+- **`numpy` added as explicit dependency** in pyproject.toml.
+- **Configurable `handoff_dir` and `session_db`** — optional handoff, no longer hardcoded.
+- **Shared `SessionRAG` factory module** — `session_rag_factory.py` replaces fragile `__import__` lambda.
+- **`k` parameter validation** in `ContentLayerManager.recall()`.
+
+### Fixed
+
+- **`UnboundLocalError` in `record_session_lifecycle`** — initialize heartbeat/handoff before try block.
+- **Debug logging** added to bare `except Exception` blocks in session_lifecycle.py.
+- **Local-only files removed from git tracking**, `.gitignore` updated.
+
+---
+
+## [0.8.0] — 2026-06-08
+
+### Added
+
+- **Semantic Reconciliation** — contradiction detection and resolution in dream cycle:
+  - `axis_pack` — antonym-axis preset loader (e.g. risk/safety, exploration/exploitation).
+  - `SemanticEngine` — antonym-axis polarity via pole projection.
+  - `ContradictionDetector` — lexical fast-path then axis opposition.
+  - `world_model` public `relation`/count reads, `state_log`, contradiction cache.
+  - `dreaming` Reconcile sub-phase — semantic contradiction marking.
+  - `output_filter` `SemanticDedup` — annotate near-dup blocks, never merge.
+- **E2E test** — dream Reconcile + ontological recovery.
+
+### Changed
+
+- `coherence.ontological_score` reads cached flags instead of `world._data` directly.
+
+---
+
+## [0.7.0] — 2026-06-07
+
+### Added
+
+- **Hook-based handoff** — `session:start` injection hook for agent framework integration.
+- **Session handoff architecture** documentation.
+
+### Fixed
+
+- **Coherence format guard** — `session_lifecycle` now guards against non-float types (MagicMock, str).
+
+---
+
+## [0.6.0] — 2026-06-07
+
+### Added
+
+- **Voice Presets** — configurable voice/personality profiles for agent communication.
+- **ContentLayerManager** — unified content operations (recall, perceive) with session RAG integration.
+
+---
+
+## [0.5.0] — 2026-06-06
+
+### Added
+
+- **Cognitive Modes**:
+  - `Shard` enum + deterministic `infer_shard` + `ShardEngine` for cognitive mode switching.
+  - `ConsciousnessState.shard` field + injection line in context_manager.
+  - Engine infers active shard in `reflect()`, surfaces in state.
+  - `TrajectoryVector` — directional cognitive momentum tracking.
+  - Shard-aware reflection with shard field in injection.
+
+### Fixed
+
+- **Shard computation timing** — compute shard at reflect-entry, before self-emitted events.
+- **`world_model.list_entities`** — revived dead `enrich` call.
+
+---
+
+## [0.4.0] — 2026-06-05
+
+### Added
+
+- **Self-Judgment (entropy + friction + meta-reflect)**:
+  - `world_model.entropy(name)` — age/isolation/relevance disorder score.
+  - `world_model.prune_by_entropy` + `recently_changed`.
+  - `world_model` bounded prediction-error log + `recent_prediction_error_rate`.
+  - `dreaming` entropy Prune + Friction (Release → Prune → Friction → Crystallize).
+  - `context_manager` `ConsciousnessState.reflection_quality` field + injection line.
+  - Engine advisory `meta_confidence` on reflect (Witness loop).
+- **Friction matching hardening** — whole-word + min-length guard.
+
+### Fixed
+
+- **Exact `dry_run` parity** in `prune_by_entropy`.
+- **Dream prune test** updated to entropy contract (age-compounded staleness).
+
+---
+
+## [0.3.0] — 2026-06-05
+
+### Added
+
+- **Metabolic Consciousness**:
+  - `Noosphere` tier model + `ContextManager.metabolic_state`.
+  - `DreamCycle` orchestrator + `engine.dream()`.
+  - `session_rag` — injectable embedder + `available()` probe + tests.
+  - `engine.recall()` — cross-session memory + reflect past-context injection.
+  - `output_filter` `DedupBlocks` + `SecretMask` stages, wired into engine.
+  - `event_bus.purge_duplicates` — all-time exact-dup collapse.
+  - `world_model.prune_stale` — decay then purge stale entities + relations.
+  - `session` trigger `dream()` after handoff (Mitosis → Dream).
+- **Performance regression guard** — reflect/dream at 10k events + 1k entities.
+
+### Fixed
+
+- **3 bugfixes**: OutputFilter config keys, SessionRAG fallback, EventBus dedup edge case.
+
+---
+
+## [0.2.3] — 2025-06-05
+
+### Added
+
+- **Conscio ↔ agent framework session lifecycle integration** — `record_session_lifecycle`, `format_heartbeat`,
+  `format_handoff`, `enrich_with_conscio`, `get_latest_session`.
+- **SessionRAG** — semantic search over session DB via Ollama embeddings.
+- **AGENTS.md** — boot instructions for AI agents working on Conscio.
+
+---
+
+## [0.2.1] — 2025-06-05
+
+### Fixed
+
+- **OutputFilter pipeline config keys** — `build_pipeline_from_dict` called with incorrect dict keys
+  (`"max"` → `"max_lines"`, `"max_chars"` → `"max_width"`).
+- **ConsciousnessEngine missing lifecycle cleanup** — added `close()` (idempotent) + context manager.
+- **Dead import in `reflect.py`** — removed unused `build_pipeline_from_dict` import.
+
+### Added
+
+- **3 regression tests** for v0.2 fixes (316 tests total).
+
+---
+
+## [0.2.0] — 2025-06-04
+
+### Added
+
+- **ContentStore** — FTS5 BM25 dual-index with RRF merging.
+- **EventBus** — deduplicated event bus with SHA-256 content hashing.
+- **OutputFilter** — 8-stage text compression pipeline.
+- **TokenTracker** — token estimation with per-source tracking.
+- **Migrator** — JSON → SQLite migration tool.
+- **Engine v0.2 integration** — all v0.2 modules in the `reflect()` pipeline.
+
+---
+
+## [0.1.0] — 2025-06-03
+
+### Added
+
+- **ConsciousnessEngine** — central orchestrator.
+- **ContextManager** — mode detection and context budget allocation.
+- **ModelRegistry** — model → context → mode mapping.
+- **WorldModel** — entity/relation store with predictions and temporal decay.
+- **MetaCognition** — confidence tracking, blind spot detection, error patterns.
+- **GoalGenerator** — drive-based goal generation with meta-score computation.
+- **AutoEvolution** — safe self-modification proposals with human approval gates.
+- **InnerMonologue** — reflection/observe/summarize loop.
+- 313 tests across 6 test files.

conscio-1.3.0/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 Neguiolidas / Neguitech
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

conscio-1.3.0/MANIFEST.in

@@ -0,0 +1,8 @@
+include LICENSE
+include CHANGELOG.md
+include README.md
+recursive-include conscio py.typed
+prune docs/superpowers
+prune mempalace
+prune config
+prune .hermes