chore: prepare v0.21.0-rc1 release candidate - #491
Conversation
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 15eefe9be8
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| 1. Back up the profile's `lcm.db` and its `-wal`/`-shm` companions, or run | ||
| `/lcm backup` from the old runtime. |
There was a problem hiding this comment.
When Hermes is still running in WAL mode, the first alternative permits operators to copy lcm.db, -wal, and -shm separately while writes continue; those non-atomic copies can represent different database states and leave the documented downgrade backup unusable. Require Hermes to be fully stopped before file-level copying, or make the SQLite backup API exposed by /lcm backup the sole supported online path.
Useful? React with 👍 / 👎.
| | `LCM_ADAPTIVE_RETRIEVAL_ENABLED` | `false` | Enable `lcm_retrieve` and bind query views for evidence reuse. Episodes are bounded to existing retrieval tools and store evidence/traces, never final prose. | | ||
| | `LCM_PREANSWER_EVIDENCE_ENABLED` | `false` | Enable the automatic pre-answer evidence hook. Disabled preserves the ordinary hook context and performs no retrieval or computation. | | ||
| | `LCM_PREANSWER_EVIDENCE_MODE` | empty | When the master flag is enabled, empty selects legacy selective behavior; explicit values are `off`, `legacy_selective`, or `requirements_v1`. | | ||
| | `LCM_SELECTIVE_COMPILER_ENABLED` | `false` | Separately opt into the semantic selector for code-derived closed operations. Pre-answer evidence alone remains provider-free. | |
There was a problem hiding this comment.
When LCM_PREANSWER_EVIDENCE_ENABLED=true is combined with an existing remote embedding configuration, a routed pre-answer hook without caller-supplied baseline refs calls _answer_ready_baseline, which dispatches lcm_recall with the user's question and can invoke that embedding provider even though the selective compiler remains disabled. The provider-free promise can therefore cause operators to enable this hook under a false data-egress assumption; document the inherited embedding behavior or force this baseline retrieval onto a local/provider-free path.
Useful? React with 👍 / 👎.
1 participant
Summary
v0.21.0-rc1identity across the manifest, operator examples, issue template, changelog, packaging docs, and version assertionsWhy
mainwould publish as a normal release, compare from a stale RC tag, and still report runtime version0.20.0Validation
476 passed2792 passed, 12 xfailed2792 passed, 12 xfailedpython -m compileall -q .python -m py_compile scripts/backfill_externalized_tool_outputs.py scripts/import_lossless_claw.py scripts/lcm_benchmark.py scripts/lcm_stress_check.pybash -n scripts/install.sh scripts/update.sh scripts/validate_release.shgit diff --check origin/main...HEAD && git diff --check && git diff --cached --checkactionlintand changed-fileruff,2 passedfailure_count: 00.21.0-rc1, all 15 tools are declared, and a second install is idempotent1 passedNotes
Refs #361, #436, #440, #446, #447