v0.3.0-rc1
v0.3.0-rc1 is the first release candidate for using gwt-context as a real
Codex/MCP memory runtime.
Highlights
- Package metadata is
0.3.0rc1.
- GitHub Pages has a demo landing page with current RC evidence and links to
the honest GWT report, dogfood report, benchmark report, runtime contracts,
and MCP tool contracts.
- MCP memory maintenance covers profile, backup, restore, reset, export,
import, dedupe, and working-memory compaction.
gwt_readiness_check and gwt://health expose compact readiness summaries
for agent clients.- Runtime collection parsing accepts key-value records, JSONL, and Markdown
tables.
- Broadcast-bus arbitration exposes proposal groups, inhibition reasons,
subscriber reports, budgets, and circuit-breaker state.
- Competition dry-run exposes admission reason breakdowns.
Current Evidence
Observed on 2026-04-29:
| Check |
Result |
pytest |
206 passed |
| Focused MCP/docs tests |
37 passed |
Qwen sanity --max-tasks 5 |
RULER and LongBench GWT bus on/off all 100% |
| LongBench baseline |
bus-on 100%, bus-off 80% |
| Subscriber timeout/error |
0 / 0 |
| Package build |
gwt_context-0.3.0rc1 sdist/wheel built |
See docs/benchmark-report-v0.3.md for the benchmark table.
Known Limits
- This is still a bounded RC sanity matrix, not a broad benchmark result.
- Built-in subscribers are deterministic local processors; fuller GWT-style
independent agent loops remain the next architecture frontier.
- The Pages landing is a demo surface and intentionally easy to replace with a
stronger product design.
Final Tag Checklist
python scripts/release_gate.py
npm run memory:health -- --smoke --json
npm run qwen:sanity -- --run --max-tasks 5