Five OpenClaw Main Fixes Protect Resets, Settings and Session State

Five merged fixes on OpenClaw main tighten one shared contract: state should not disappear merely because an interface changes mode, reconnects, carries an attachment or discovers invalid input.

The sharpest edge is openclaw onboard --reset in an interactive mode without a real terminal. Before the fix, OpenClaw could move configuration, credentials and sessions to Trash before the wizard failed or hung. The other changes protect transcript text, unsaved Control UI settings and live session continuity in Claude CLI and the terminal UI.

These are merged source changes, not proof of inclusion in an official package. Match your deployment to a containing release or exact source revision before treating any path as corrected.

Triage table

If none of these surfaces is in use, add their acceptance tests to the next controlled upgrade rather than moving to an unversioned main build solely for this digest.

Headless reset could mutate state before failing

PR #116491 fixes an ordering error in onboarding. Interactive guided/classic reset already required a real TTY, but the contract was checked too late. The reset path could snapshot and move configuration, credentials and sessions to Trash before the wizard discovered it could not interact.

The merged change runs the TTY preflight before any reset snapshot or mutation. Piped or headless interactive onboarding should now fail immediately with guidance and leave state untouched. Explicit non-interactive workflows retain their own supported path.

The operator rule remains stricter than the code fix: a reset is destructive maintenance even when recovery uses Trash rather than immediate deletion. Before running it:

1. record the intended OpenClaw home and profile;
2. take a restorable backup of config, credentials, sessions and databases;
3. confirm the command has the expected TTY and user identity;
4. stop automation from retrying a failed reset blindly; and
5. verify the resulting state before deleting any recovery copy.

Do not “test” the old bug against your only profile. Use a disposable OpenClaw home containing synthetic state.

Media replies could lose their text in history

PR #116225 covers a split between delivery and persistence. When an assistant reply paired commentary with an attachment, the shared transcript mirror could preserve the attachment name while dropping the reply text. The recipient saw the message, but later session history did not contain the same semantic record.

That matters beyond cosmetic history. Transcripts can feed continuity, memory, audits and incident reconstruction. A missing sentence can make a later model appear to ignore context that was never persisted.

Upstream says the fix preserves reply text alongside stable attachment names and adds formatter, SQLite persistence, outbound delivery and Gateway fallback coverage.

For potentially affected sessions, do not reconstruct text from memory and write it back as though it were original evidence. Preserve the delivered channel record, transcript and timestamps separately; label any reconstruction clearly.

Control UI drafts now fail visibly instead of vanishing

OpenClaw configuration is schema-validated, but the editor around that schema also has to preserve a user’s intent while input is temporarily invalid. PR #116282 addresses cases where JSON, scalar, collection or nested drafts could be discarded, optional structures could save partially, stale row state could be reused, or numeric coercion could disagree with the validator.

The corrected UI retains invalid drafts with validation, aligns controls with schema constraints, stages optional nested values atomically and stabilises collection-row identity.

Interpretation: validation failure should block a commit, not erase the proposed change. An editor must keep three states distinct: persisted configuration, valid unsaved input and invalid unsaved input.

After upgrading, test one harmless setting from each complex shape you rely on. Confirm that invalid input remains visible, cancel restores the persisted value, save commits the complete valid structure and a reload reads back exactly what was accepted.

Claude CLI prompt refresh should keep live context

PR #116292 fixes a different form of state loss. Claude CLI-backed sessions need updated system prompts when OpenClaw configuration changes, but deleting the prompt fingerprint to force a refresh also discarded the live harness context.

The merged replacement refreshes the system prompt without treating the existing conversation as disposable. Operators using long-lived Claude CLI sessions should verify both sides of the contract: the new prompt policy takes effect, and prior harmless context still exists exactly once.

This does not mean every configuration change is safe to apply mid-task. Tool-policy or authority changes should still be tested at an explicit boundary, and a stale model statement is not proof of the effective runtime policy.

TUI reconnects should not change the conversation underneath you

PR #116399 preserves TUI session state across switches and reconnects. For operators, the practical risk is attribution: continuing a command in the wrong or freshly reset context can apply an otherwise valid instruction to the wrong task.

After a network interruption or session switch, confirm the immutable session identity and a harmless recent marker before authorising file, shell or external side effects. Interface continuity is useful evidence, but it is not a substitute for checking the target.

Acceptance run

Use synthetic data and a disposable profile:

1. invoke interactive reset without a TTY and prove the state tree is byte-for-byte unchanged;
2. send harmless text plus an attachment and confirm both appear in delivery and SQLite-backed transcript history;
3. enter invalid then valid nested settings and test retain, cancel, save and reload behaviour;
4. refresh Claude CLI prompt configuration and verify new policy plus old context without duplication; and
5. switch TUI sessions, interrupt the connection and confirm each identity resumes its own state.

Preserve logs without credentials, message contents or raw configuration secrets.

Evidence and limits

OpenClaw Academy reviewed the official merged PR and commit records in the repository packet. Web search independently corroborated the headless-reset ordering fix and the schema-backed editor behaviour; the official collector supplied merge chronology and detailed summaries for the transcript and Claude changes. We did not execute these paths against a live installation or verify a packaged release containing all five fixes.

The evidence establishes defects and merged corrections, not a universal affected-install count. Treat upstream test claims as upstream results and validate the exact surfaces your deployment uses.

Bottom line

State protection has to happen before the risky transition, not after it: preflight before reset, persist text alongside media, retain invalid drafts, and refresh interfaces without throwing away live context. Upgrade through a confirmed containing build, then test the boundary—not merely whether the UI or Gateway still opens.