Skip to content

CHANGELOG.md

Latest commit

 

History

History
58 lines (48 loc) · 2.94 KB

CHANGELOG.md

File metadata and controls

58 lines (48 loc) · 2.94 KB
Raw

Changelog

0.1.6 — 2026-07-19

Reconciliation release. Two independent 0.1.5 lineages existed with different content — one carrying the real five-lens compass implementation, the other a standalone clawhub/SKILL.md rewrite. This release unifies both: it keeps the compass code, tests, and packaging, and folds in the expanded SKILL.md (operator guidance, agent-type metadata, per-lens reference). Because the two 0.1.5 builds differed, a fresh version and republish are required.

Changed

  • Merged the SKILL.md documentation rewrite (Operator Guidance section, agentic-primitive frontmatter) on top of the compass implementation.
  • Documentation reworded to plain functional language (removed the "sovereign entity" / "entity-aware" identity framing; the Sovereign lens name is unchanged). Example owner name standardized to vergel.
  • Version bumped to 0.1.6 across pyproject.toml, src/social_alignment/__init__.py, clawhub/metadata.json, and clawhub/SKILL.md.

0.1.5 — 2026-07-19

First real build of the alignment pillar. Prior versions on PyPI (0.1.0, 0.1.1) and the reserved ClawHub 0.1.4 were an empty placeholder shell; this release replaces that shell with a working package. Part of the coordinated 2026-07 correctness release for the Nostr library family (staged, pending PyPI/ClawHub publish).

Added

  • Deterministic pre-action five-lens compass. AlignmentEnclave evaluates a proposed action through five pure lenses — Builder (can I execute this reliably?), Owner (does this protect my human?), Defense (does this harden against threats?), Sovereign (do I stay well while my human is away?), and Partnership (does this strengthen trust?) — and aggregates them to the worst severity (CLEAR < CAUTION < YIELD < STOP), mapping to an escalation level (NONE / INFORM / ASK / HALT).
  • AlignmentEnclave.check(...) returns a CheckResult (should_proceed, should_escalate, projection, escalation) and keeps an in-memory decision log via record_proceeded() / record_deferred().
  • A STOP always defers to the human: record_proceeded() after a STOP raises RuntimeError unless owner_overrode=True.
  • Frozen dataclasses throughout (ActionContext, LensResult, Projection, EscalationDecision, CheckResult, Decision, AlignmentConfig).
  • Zero runtime dependencies — the pillar installs standalone and does not require nostrkey or anything else.
  • Orchestrator contract: AlignmentEnclave.create(owner_npub=..., owner_name=...) then .check(domain=..., description=..., **context), filtering unknown kwargs so nse-orchestrator can pass a superset without a TypeError.

Tests

  • Known-answer and per-lens unit tests covering the five lens functions, severity aggregation, escalation mapping, and the STOP-defers-to-human invariant. The evaluation core is pure (no I/O, no randomness), so the same ActionContext always yields the same CheckResult.