gwt-context

v0.3.0

v0.3.0 is the release candidate for making the GWT runtime usable as a Codex MCP memory system rather than only a benchmark/demo runtime.

Highlights

• MCP memory management surface: gwt_memory_profile, gwt_backup_memory, gwt_restore_memory, gwt_reset, gwt_export_memory, gwt_import_memory, and gwt_compact_working_memory.
• Structured runtime read models bootstrap from persisted memory after MCP startup, so exact gwt_collection_query survives restarts.
• Runtime collection parsing accepts key-value records, JSONL records, and Markdown tables with exact count/filter/top-k/average/sum/distinct/min/max operations.
• Stored records receive active namespace tags for project/global separation.
• Broadcast-bus arbitration records decision reason codes, summaries, and proposal grouping in traces and inspect output, with per-subscriber proposal budgets, payload budgets, and circuit-breaker state.
• Competition dry-run exposes admission reasons for admitted, evicted, kept, below-threshold, and eligible-but-not-selected candidates.
• Optional OpenAI-compatible external subscribers can be wired at the MCP composition root with separate bus and provider budgets.
• Integration coverage proves recurrent linked-memory activation and accepted subscriber queries can affect later workspace contents.
• Codex bootstrap, health, and cleanup helpers support split project/global MCP namespaces.
• docs/honest-gwt-report.md records the strongest defensible GWT claim and the remaining limits.
• docs/dogfood-report.md records the latest real local MCP, Codex MCP, and bounded Qwen sanity evidence.

Current Evidence

Observed on 2026-04-29:

Check	Result
pytest	206 passed
npm test	206 passed plus deterministic benchmark smoke 2/2
Local MCP usage loop	graph Paper Gamma, collection Ada, trace ok
Codex MCP temp smoke	returncode=0, tool count 21, trace ok
External subscriber POC	accepted flag_contradiction and query_memory, timeout/error 0 / 0
Qwen sanity --max-tasks 5	RULER and LongBench GWT bus on/off all 100%, baseline 100% except LongBench bus-off 80%, timeout/error 0 / 0

Verification

Run before tagging:

python scripts/release_gate.py
npm run memory:health -- --smoke
npm run qwen:sanity -- --run --max-tasks 5
python -m tests.benchmarks.bus_matrix --summarize .benchmarks/qwen-sanity/*.json
python -m tests.benchmarks.analyze_results .benchmarks/qwen-sanity/*.json

Release Gates

• pytest, ruff, mypy, npm verification, stdio MCP smoke, examples, build, and boundary checks pass.
• No tracked .env, benchmark result artifacts, supervisor folders, or real RunPod URLs.
• Bus-on accuracy is not below bus-off on the same benchmark slice.
• Bus-on average tool calls are no more than 0.25 above bus-off.
• Bus subscriber timeout/error count is 0 / 0.
• Benchmark summaries include bus inhibition reason counts.

Known Limits

• External subscribers are opt-in and should be treated as experimental until a larger bus on/off matrix is regenerated.
• The project/global split is convention-based through GWT_DATA_DIR; it is not a remote multi-tenant authorization layer.
• Current evidence is enough for local use and an RC, not for broad benchmark superiority claims.