mirror of
https://github.com/nearai/ironclaw.git
synced 2026-09-03 08:06:01 +08:00
89443ace3a
* fix(config): keep startup LLM fallback in-memory only (#3229) A transient hydration failure (DB read race, secrets decryption hiccup) during startup used to overwrite the user's persisted llm_backend with "nearai" and delete selected_model from the DB, turning a one-off fallback into a permanent reversion of the configured provider. The fallback now stays in-memory: the runtime still demotes to NearAI to keep the instance usable (#2514), but the user's DB-persisted selection is preserved so the next restart with credentials available picks up their original config. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * ci: re-trigger workflows for #3229 Empty commit to refresh GitHub's pull/3324/head ref, which was not published when the PR was opened, blocking the regression-test enforcement workflow's git fetch. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * ci(regression-test-check): derive PR head from event payload Stop fetching `refs/pull/${PR_NUMBER}/head:pr-head` and source the head SHA from `github.event.pull_request.head.sha` instead. GitHub occasionally fails to publish `refs/pull/N/head` for in-repo PRs (#3324 hit this — the workflow could not get past its first git command, blocking the regression-test enforcement check before it ran any of its actual heuristics). The merge commit produced by actions/checkout has the head SHA as its second parent and is fetched in full by fetch-depth: 0, so the SHA is always locally reachable. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>