* perf(mantis): parallelize proof builds and warm caches
* fix(mantis): size build image for the copied pnpm store
* chore(ui): refresh startup JS gzip baseline after streamed-markdown perf work
Identical source measures 345034-345058 B across builds while the
committed baseline left only a 9 B margin under the 512 B ratchet
tolerance, so build-artifacts flips on gzip nondeterminism (green on
main run 32559609413, red on PR run 32559442295, red locally).
Regenerated with scripts/check-control-ui-performance.mts
--update-baseline; the 350 KiB hard ceiling still bounds creep.
* fix(tui): own local shell process scopes
Keep local shell root results independent from authoritative process-scope
extinction, and make TUI shutdown cancel and join its ProcessSupervisor scope.
On Windows, create commands atomically inside a retained non-breakaway Job
Object and prove extinction from exact Job accounting.
Fixes#127488
* fix(process): preserve optional extinction waits
* fix(sessions): keep resolved skills out of durable state
Repair runtime-only skill persistence across SQLite, legacy stores, bounded Doctor cleanup, and lightweight health reads.
Refs #126663
Co-authored-by: ruel225 <ruel225@users.noreply.github.com>
* test(health): assert lightweight session list projection
---------
Co-authored-by: ruel225 <ruel225@users.noreply.github.com>
* fix(auth): create a fresh install with canonical shared-auth ownership
A brand-new install was born in the retired shape. `parseSharedAuthStoreOwnership(undefined)` returns
`legacy-main`, which is the correct compat answer for an existing install whose profiles really do
live in the main agent database -- but a new install has no ownership row and no legacy data either,
so onboarding wrote its first credential into `agents/main/agent/openclaw-agent.sqlite` and the
operator's very first `openclaw doctor` told them to run a migration for state OpenClaw had created
seconds earlier. The main agent also stayed undeletable until they did.
Record `auth.sharedStore = {"location":"state-db"}` when the shared store is first written and the
legacy source provably holds nothing: no `auth_profile_store` row, no `auth_profile_state` row, and
no unfinished cleanup ledger entry. Any legacy row, or any inspection error, leaves ownership alone
so doctor keeps owning the relocation. The check is memoized per ownership generation with a WeakSet
keyed on the process-stable ownership object, so a legacy root is inspected once per process and
doctor's committed flip naturally invalidates it.
The legacy row inspection moves out of `state-migrations.shared-auth-store.ts` into the auth-profiles
owner so doctor and runtime share one contract instead of runtime importing migration code. Explicit
main-agent credential writes now follow the shared target, which is a no-op on legacy roots where
both routes already resolve to the same file.
No SQLite schema change; the ownership row is data. Existing installs take exactly the path they take
today.
* fix(auth): preserve JSON-era shared credentials
* docs(auth): explain why doctor names the main agent dir during a shared JSON import
* test(auth): assert shared-owner runtime reads
* test(doctor): read migrated catalog credentials through the shared owner
A fresh root records state-db shared ownership, so the model-catalog credential migration persists into the shared store rather than the agent file. The assertion read the agent file directly and saw an empty store while all three credentials were present and correct in state/openclaw.sqlite. Read through the owner for that state root instead of pinning storage layout; the credential contents are still asserted exactly.
Host zlib versions compress identical startup assets into different bytes, causing false Linux budget failures. Emit canonical shipped pako gzip sidecars, restore the 512 B ratchet tolerance, and lower the startup baseline from 348351 B to 344531 B.
* fix(config): validate config writes against the config being written
writeConfigFileFromContext passed the pre-write snapshot's plugin
metadata into strict validation. During onboarding that snapshot belongs
to an intermediate config written by agent creation, which has no plugin
entries, so its scoped manifest registry is empty. Validating the final
candidate against it made every plugin entry added by the same write look
unknown, and non-interactive onboarding warned that the openai and codex
entries it had just written were stale or uninstalled.
Drop the stale snapshot so validation resolves the manifest registry from
the candidate it is actually validating. Strict semantic validation is
unchanged, and the registry load stays lazy.
* fix(ci): raise the Control UI startup JS baseline to unblock main
main is red on the Control UI startup-JS ratchet: unrelated PR #126725 measures 348289 B and this branch measures 348351 B against a 347023 B baseline + 1056 B tolerance. Twenty-six UI commits have landed since the last bump (#126474), none individually large. Baseline moves to the CI-measured 348351 B, well under the 358400 B maintainer-approved ceiling that still guards cumulative creep.
The OpenAI ChatGPT auth profile identity was derived in two drifted copies:
the plugin-sdk helper (used by Codex/Hermes import migrations) fell back to
bare JWT sub and leaked the workspace accountId into the user subject, while
the openai extension copy (used by login/refresh) composed the OIDC-stable
iss|sub pair and honored the credential email fallback.
Fold the extension's chain into the canonical SDK helper (credential-email
fallback, iss|sub before bare sub, no workspace-id subject), delete the
extension copy, and move all extension call sites onto
openclaw/plugin-sdk/provider-auth. Cross-checked against Codex OSS
(codex-rs/login/src/token_data.rs): chatgpt_account_id is workspace identity,
never user identity. No doctor migration: the divergent fallback branch only
fires for tokens lacking every user-id claim and email, which real ChatGPT
tokens do not produce, so no shipped install holds divergent keys from it.
* refactor(ui): unify sidebar alerts with custodian
* refactor(ui): move custodian alert state
* fix(ui): use the defined mono token in the custodian alert card
* fix(ui): break the update-watcher import cycle
Type the watcher against a structural leaf contract instead of
Pick<ApplicationContext,...>; context.ts reaches this module through
overlays-types.ts, so naming the context type closed a madge cycle.
* fix(ui): re-arm the alert explanation when an incident recurs
Failed-automation and model-auth alerts keep one incident id across
recover-then-fail-again, so an id-keyed dedupe showed the renewed alert
and never explained it. Scope ask-once to the presentation instead.
* fix: surface Codex input prompts across runtimes
Codex structured input now reaches bounded Gateway questions in native and ACP runs, with exact turn ownership, explicit unsupported outcomes, and cancellation fencing. Consume the published ACPX elicitation support.
* chore: align elicitation helpers with current guards
Use protocol-specific helper names required by current main and update the reservation regression to the generalized input owner.
* fix: formalize structured input SDK surface
Expose one documented, frozen agent-harness structured-input contract with runtime and subpath coverage, replacing the accidental function-property API.
* fix: satisfy elicitation architecture gates
Register the real-process ACPX fixture as an executable test root and move shared structured-input types into the boundary leaf to keep Knip and Madge clean.
* fix: remove structured input lint suppression
Preserve the rejected control and invisible-character ranges with an explicit code-point check so the production suppression inventory stays closed.
Ayaan Zaidi