v0.8.0: decay + provenance schema. Slash write ops. Antigravity held.

The schema cut promised to the working group on anthropics/claude-code#47023 and openai/codex#19195.

What is new

• Domain-aware confidence decay -- new world_model_server/decay.py module. Per-evidence-type half-lives (source_code 365d, test 180d, session 14d, user_correction 730d, bug_fix 365d). Decay applies on read in query_facts. Settled facts (canonical, or any fact with confirmer != NULL) never auto-transition. Synthesized facts under 0.20 confidence and corroborated facts under 0.10 confidence auto-supersede on read.
• Per-item provenance fields on facts -- three additive NULL-defaulted columns: source_tool, confirmer, last_decay_at. The Fact model gains matching optional fields. create_fact persists both. Honors the public commitment to Patdolitse and ferhimedamine.
• Slash command write operations -- /world-model resolve <id> marks a contradiction as resolved; /world-model forget <id> sets invalid_at on a fact (preserved in audit log). Both are idempotent.
• resolve_contradiction accepts confirmer -- when set, the winning fact gets stamped, marking it settled.

Antigravity adapter held to 2026-06-27

The 2026-06-13 re-verification found OnCompactionHook declared as InspectHook in google-antigravity/antigravity-sdk-python HEAD, with no TransformCompactionHook subclass and no additional_context return field. The architectural gap that blocked v0.7.5 and v0.7.6 has not closed. Next re-verify 2026-06-27.

Quick start

pip install -U world-model-mcp
# Schema migration runs automatically on first KnowledgeGraph.initialize()
# Existing facts get NULL provenance. New facts can carry source_tool / confirmer.

Numbers

• 26 MCP tools (unchanged; resolve_contradiction gains optional kwarg)
• 19 CLI subcommands (unchanged)
• Slash subcommands: 4 read + 2 write (was 4 read in v0.7.6)
• 375 tests passing (was 333): 42 new in tests/test_v080_features.py

Backward compatibility

• All v0.7.6 surface unchanged
• Schema migration is additive, idempotent, NULL-defaulted, no backfill
• inject_helper and hook_helper unchanged
• No new required dependencies
• All four adapters (Claude Code, Cursor, Codex, pi) unaffected

What is next (v0.8.1)

The benchmark publication arc: LoCoMo confidence-on/off comparison, expanded 200-pair contradiction-resolution benchmark on HuggingFace, contradiction-recall benchmark methodology. End of June target.

Full notes in RELEASE_NOTES.md.