[ytvee]

Result

Prepare one v1.9.0 release on the current Kit base: retain the transferable STEM rules and learning workflow, then apply the user's chat fixes. STEM/ASI installations are unchanged. Full source coverage (243 reusable files plus root policy) and original 42-file learning provenance are inventoried with hashes and explicit exceptions.

Behavior

• New project memory is local by default, with capture, telemetry and exchange off and no preset remote. CLI/hooks resolve both project flag forms consistently; explicit state overrides and project-only legacy migration remain supported.
• Natural-language requests to save experience write structured memory, not project rules. Triage skips curated event IDs; explicit reconsideration and stable paged context cover partial/long sessions.
• Snapshot exchange uses the verified SQLite destination, blocks conflicting legacy configuration and requires separate completion-stage consent. No background learning or implicit upload.
• Adaptation offers Langfuse connection/decline, reuses keys or opens a private credential file and waits. Scoped authentication, environment checks, stable project tags and current-probe readback distinguish setup from ingestion. Other projects and saved decisions are preserved.
• Selected incident dates drive bounded, resumable telemetry queries and filtered archive context; missing evidence falls back to SQLite.
• Routine model/version/effort announcements are omitted. Kimi support and update protections remain included.

Verification

Local final source: 91 learning tests passed; full source/generated-target/archive validation, Ruff, Markdown and link checks passed. Current head: 54d1f26a70cd662e8934af1a5c02734dc39df4f5. All five GitHub checks pass on this head: links, lint, Linux helpers, Windows helpers and source/portable targets. Old green checks were not reused.

Five retained Codex CLI trials include baseline/candidate memory and candidate onboarding/output cases. Memory persistence was blocked by the client's protected .agents policy; the candidate chose the intended workflow, reported the write failure and made no substitute rule edits. Read-only key/decline responses matched the contract; output omitted model boilerplate. These do not prove native file opening, hook activation, real Langfuse ingestion or general agent improvement. Exact prepared fixture hashes and limitations are recorded in docs/release/1.9.0-verification.json; private traces remain local. Later helper fixes have regression coverage, not a final-head live certification.

Release boundary

One PR, one v1.9.0 release. The user separately authorized merging and publication on 2026-10-29. The release workflow now publishes both Kimi archive names. Remote draft assets built from e48ea26 are STALE and must not be published. Twelve local archives were rebuilt from this candidate, archive validation passed and every SHA256SUMS entry was verified. All 244 committed source-inventory normalized hashes match. The release process replaces stale draft assets with a fresh build of the merged commit before publication.

Fixes #84
Fixes #85