Skip to content

feat(browser): improve browser session and crash recovery #20

Closed

Description

@trmosala

Problem

Recovery stores URLs and selected tabs, not navigation stacks or form state.

Implementation order

Browser backlog item 16 of 29. Phase: Compatibility and recovery. This is ordered planning, not a claim that Codex implements every item. Items 24–29 are optional additions.

Acceptance criteria

  • Restore navigation history where supported and provide actionable failed-restore recovery.
  • Preserve private-by-default grants and a locked vault on restoration.
  • Test crash/relaunch and schema migration; do not persist sensitive form values.

Dependencies

No prerequisite backlog issue.

Engineering and validation

Reuse existing Chromium/Electron/OS facilities where available. Preserve main-process ownership and permission enforcement, add no placeholder controls, use existing i18n patterns, and test behavior with synthetic fixtures. Document platform limitations. Baseline: browser-tabs at 47fef3e; see packages/cm-browser/README.md and password-security.md.

Activity

  1. trmosala commented on Sep 16, 2026

    @trmosala
    OwnerAuthor

    Concrete recovery bug, user-confirmed 2026-09-16: Windows Update rebooted the machine during the password-generation publishing subtask. After relaunch the UI still indicated that the old task was running, although that execution had ended. The local commit c148c22 existed, with origin still one commit behind; no active Git publisher was found during reconciliation. Recovery resumed by inspecting local/remote state and pushing the existing commit, not repeating implementation or replaying the interrupted side effect.

    Include this in #20: reconcile restored running indicators with live execution ownership after crash/reboot, show interrupted/recoverable rather than falsely running, and provide an explicit continuation path. Add an isolated crash/relaunch regression, including interrupted tool publication and no automatic duplicate side-effect replay. Preserve private-by-default browser grants, locked vault restoration, and no persisted sensitive form values. Root cause is not yet diagnosed; this records the observed incident, not a completed fix. Do not restart the installed app/bridge driving the active session for validation.

  2. trmosala commented on Sep 16, 2026

    @trmosala
    OwnerAuthor

    Read-only diagnosis of the reboot stale-running incident: likely durable unfinished transcript state outliving process-local execution owners. This mechanism is supported by source, but affected session records were not inspected, so incident attribution is not proven.

    Evidence: packages/core/src/session/projector.ts persists tool/assistant updates; packages/opencode/src/session/processor.ts interruption cleanup requires a live process. Legacy run-state/status and V2 run-coordinator owners are process-local maps. packages/session-ui/src/components/message-part.tsx derives task running directly from pending/running tool state, and streaming from absent completion timestamps. V2 runner/llm.ts has stale-tool settlement when execution starts, but an advisory wake without pending input returns before that cleanup. This extends #20 beyond browser tab restoration into session transcript recovery.

    Repair direction: idempotent settlement-only reconciliation under the real storage/ownership boundary, serialized before new execution. Preserve partial outputs, child links, completed results and pending admissions. Mark interrupted with external outcome unknown; never infer rollback or replay tools/provider calls. Separately audit server-sync activity refresh on backend reconnect with race protection; that surviving-renderer path alone does not explain a full relaunch.

    Required isolated regression: crash-shaped unfinished parent/child/tool records plus completed side-effect sentinel and pending admission, fresh owners, reconcile twice. Assert terminal interrupted records, retained evidence/admissions, no false-running UI, no duplicate settlement and ZERO tool/provider invocations or sentinel changes. Real browser restoration must still be private-by-default with locked vault and no sensitive form persistence.

    Read-only exploration only; no fix or native crash/relaunch proof yet. Sanitized parent/child IDs, installed build/protocol and exact stale indicator would strengthen incident attribution, but are not required to start the isolated regression.

  3. trmosala commented on Sep 16, 2026

    @trmosala
    OwnerAuthor

    Additional recovery evidence, 2026-09-16: the pending read-only #12 review (task ses_f5905c6a3ffeYlhAvgmZBvPSmg, call cm12_complete_review_107) returned wpp_auth_required with: "WPP login required: WPP page is on a login or identity-provider URL. Original failure: Timed out waiting for O1-Code worker result after 900000 ms." The user reported that no bridge job remained active and the visible WPP session was authenticated; Okta sign-in was subsequently refreshed successfully. They identified the timeout as misclassified login-required and requested explicit continuation of the saved work, not a replay of completed implementation/tests/publication.

    Observed tool output establishes the timeout plus auth reclassification. Authenticated visible-session/no-active-job state and misclassification are user-reported, not independent evidence of the failed worker page state at the original failure. Together with the earlier reboot stale-running incident, #20 should distinguish persisted UI running state from live job ownership and preserve original timeout/capture causes instead of presenting login as proven cause. A live-session probe may refer to a different or stale worker; causality remains unproven without sanitized job/worker state.

    Recovery here inspected the saved uncommitted diff and checkpoint, then retried only the unfinished source review. Both fresh bounded reviewers returned successfully; no completed implementation/tests or side effects were automatically replayed. Add isolated coverage for timeout settlement, stale running indicators, authenticated-visible-session versus worker-auth disagreement, and explicit continuation with unknown external outcome. Preserve settlement-only/idempotent recovery, partial results/child links/pending admission, zero automatic tool/provider replay, and private grants/locked vault restoration. This records evidence, not a #20 fix. Installed app/bridge and CM_GPT6_Astra_High were not changed or restarted.

  4. trmosala commented on Sep 16, 2026

    @trmosala
    OwnerAuthor

    Additional continuation evidence: #14 review-fix call cm14_review_fixes_173 (child ses_f5527cbdcffepUmzUuhPN27mLT) timed out after 900000 ms with type/code o1_code_extension_timeout. Unlike the earlier #12 incident, this result was not reclassified as login-required; no auth change was requested.

    Reconciliation found partially saved changes, not a rolled-back task. A bounded process query found no matching fixture/test/publication processes, and a report-only child continuation returned successfully, identifying completed HTTP red/green evidence and saved but not yet verified fill guards. It reported no outstanding tool calls and explicitly relinquished editing. Slot naming and parser cleanup remained unfinished. Recovery continues from those saved steps, without replaying completed work or assuming the timeout undid side effects. The process query is not proof about every worker or the UI running indicator.

    This supports #20 settlement-only recovery with preserved partial output/child links and original timeout cause. No new stale-UI or auth-misclassification claim is established by this incident. No #20 fix, automatic side-effect replay, or installed app/bridge/model changes.

  5. trmosala commented on Sep 16, 2026

    @trmosala
    OwnerAuthor

    Additional bounded recovery evidence: #16 implementation call cm16_implement_321, child ses_f53bb78fdffeeEfs6Ob6ZdREyc, returned Timed out waiting for O1-Code worker result after 900000 ms., type/code o1_code_extension_timeout (not auth reclassification). Saved work was one protocol test only, observed red; no native production changes. Process scan found no matching Bun/native/publisher execution.

    Report-only continuation cm16_recover_report_322 failed with o1_code_required_tool_not_called: The WPP agent returned two terminal responses without the tool call required by this turn. A subsequent narrowly authorized read-only continuation cm16_recover_readonly_327 succeeded, confirmed exact saved state and no outstanding tools, and explicitly relinquished checkout ownership. Parent preserved saved work and assigned a fresh smaller protocol/plugin slice; no automatic replay or repeated external side effects. This records the recovery/report-only transport limitation, not a #20 implementation fix. CM_GPT6_Astra_High and installed processes remained unchanged.

  6. trmosala commented on Sep 17, 2026

    @trmosala
    OwnerAuthor

    Browser recovery slice published

    Commit: 913eb4f on browser-tabs, based on b69d5e0. Local and remote SHAs confirmed equal. Exactly 12 files, 532 insertions / 60 deletions. Normal pre-push typechecks passed 31/31 tasks, 28 cached. No release, deployment, dependency/configuration/model changes or installed-process restart.

    Electron navigationHistory.restore now receives URL/title-only entries and the active index. Recovery retains up to 20 validated entries around the selected page; legacy URL-only records, including long URLs, remain supported. Invalid navigation metadata falls back to the saved URL. No pageState, history.state, form values, POST bodies or agent grants are persisted by this recovery path. Restored tabs receive fresh IDs with agent access off; the vault starts locked on relaunch.

    Pending/failed restoration retains its snapshot. Failed Reload, HTTP 204 and Stop do not discard history merely on request start. Replacement/closure invalidates stale callbacks. Restoration failure explains existing Reload/address/close actions; ordinary load-error copy remains unchanged. Malformed stores are not overwritten or cached as incomplete groups.

    Verification on Windows, Electron 42.3.3 / Chromium 148.0.7778.218:

    • Desktop typecheck and 155 browser tests / 2276 expectations passed.
    • App typecheck and 24 scoped browser/i18n tests / 506 expectations passed.
    • Separate serial native default, --recovery, --tab-lifecycle, --frames and --account-fill runs passed.
    • --recovery kills only its owned seed process at a checkpoint, relaunches the same temporary profile and checks selected history index, Back/Forward, recently closed history, fresh private IDs, vault locked before unlock, empty form/history state and zero POST requests on relaunch.
    • Redirects, failed Reload, HTTP 204, Stop, pending replacement/close and malformed-store preservation passed.
    • Changed-path Prettier and staged whitespace passed; scoped lint of ten TypeScript files: 84 warnings, zero errors.
    • Final source-only correctness and standards reviews clear; not independent test execution.

    Failed attempts retained in password-security.md: incorrect Electron index argument, premature fixture close before intended navigation committed, fixture typing/destroyed-handle errors; failed-Reload history loss and legacy-long-URL rejection reproduced before correction. Malformed-group and ordinary-wording checks are GREEN coverage only. Lifecycle initially timed out because its held-key fixture had no focused main frame; explicit fixture focus and fail-fast dispatch observation fixed the check without weakening production guards. The crash seed has one open tab, not new multi-open-tab selection evidence.

    Limits: URL/title metadata is plaintext and can contain site-reflected secrets. Restored pages execute scripts and authenticated requests using existing cookies/storage; excluding POST bodies does not make page loads side-effect-free. Authentication/dialog answers are fixture seams. Power-loss, packaged builds, native accessibility, reviewed translations, macOS/Linux, real providers and #34 assurance remain unverified.

    Task recovery remains open

    This does NOT fix the reboot stale-running transcript incident, worker/auth disagreement or timeout settlement. No persisted transcript is settled and no provider/tool/publication is replayed. Investigation found no reliable abandoned-owner identity: legacy and V2 runners are process-local; SQLite WAL/busy timeout, Electron single-instance lock and daemon discovery do not enforce exclusive execution across all database writers. An empty local map cannot prove another process is not working.

    Next decision: agree an enforceable cross-process execution ownership/fencing boundary before settlement-only recovery. Preserve partial output, child links, completed effects and pending admissions; test idempotent reconciliation with zero tool/provider calls. Root AGENTS.md also requires explicit post-crash continuation design. This is the stopping gate for the remaining #20 work, not a claim that the entire issue is complete. #19 real-account/platform validation and #34 remain open. Excluded files untouched; CM_GPT6_Astra_High preserved.

  7. trmosala commented on Sep 17, 2026

    @trmosala
    OwnerAuthor

    Concurrent-host recovery: conditional design proposal

    Status: DESIGN FOR USER REVIEW, not implementation or activation approval. Baseline 913eb4f. The user selected preserving concurrent desktop/CLI/server hosts rather than a single execution host. The published browser-recovery slice remains unchanged; the stale-running incident is not fixed.

    Scope and invariants

    Cooperating process hosts share one local SQLite database. Different Sessions may execute concurrently; one Session has one execution authority. This is not multi-machine clustering or network-filesystem support. Keep Session Drain process-local, EventV2 replay ownership separate, and existing process-local active/interruption contracts unchanged. Locally unowned interruption remains a no-op.

    No automatic provider, tool, publication or interrupted-operation replay. Old untracked records remain untouched and unverified. Recovery preserves partial output, child links, completed results, pending admissions and original failure causes. Controller interruption does not prove that a subprocess stopped or an external operation failed.

    Ownership and fencing

    A shared Core SessionAuthority module serves legacy and V2 execution adapters. Candidate lifetime guard: a dedicated native SQLite connection holding BEGIN IMMEDIATE in a separate private lock database per Session. Never hold the MAIN database transaction over execution. This uses an installed facility, but its safety across Bun/Node bindings and platforms is UNPROVEN.

    No heartbeat expiry, stale-lock deletion or timeout takeover. Existing Flock/EffectFlock are unsuitable as execution lifetime guards because they break aged locks. Canonical database/lock identity must be shared across path aliases. Existing tracked claims with missing or malformed lock resources fail closed. Do not delete, replace or prune lock files while hosts may exist. Validate same-process contender/probe open-close behavior, multiple embedded hosts and missing-file refusal; the existing general Database layer is not a ready-made lock adapter.

    After obtaining the lifetime lock, use short main-database transactions for retained authority epochs and a random runtime-instance identity, not PID-based identity. Work and late finalizers retain their original capability; they cannot borrow the newest epoch. Execution-owned publication validates authority in the SAME transaction as projection, event insertion and sequence advancement. Cover direct Context Epoch/todo SQL and live deltas separately. Pure admission cannot be blanket-fenced as execution.

    Admission and dispatch

    Only complete, pure durable input is owner-independent. Legacy message and part publication needs atomic execution eligibility; another executor must not consume a partially prepared message. Command substitution, destructive revert cleanup and tool-based preparation must execute under captured authority and recorded attempt facts, or fail BEFORE effects on activated storage until integrated. This changes existing behavior and needs compatibility approval.

    Allocate a private unique physical-attempt ID before every provider/tool/preparation dispatch, including retries, title/compaction, nested MCP, shell and child launch. Associate existing Session/assistant/part/tool identifiers where applicable; no artificial public empty assistant message or durable enclosing run is required. Persist intent as MAY-HAVE-DISPATCHED, never proof of sending or non-sending. Settlement attribution commits with actual transcript facts. Store IDs/status rather than duplicated payloads or secrets. Unobservable retry paths must be adapted or explicitly disabled. Revalidate after permission waits and before effects; neither this nor database fencing guarantees atomic remote dispatch or exactly-once external effects.

    Recovery transitions

    1. Under the lifetime lock, classify the control row and attributed outstanding work BEFORE replacing authority. Clean released tracked state may grant execution for requested work. Busy, uncertain lock/identity, incompatible storage or ambiguous legacy evidence cannot grant execution or mutate the transcript.
    2. For an unreleased tracked execution/preparation/mutation claim, atomically retain its exact recovery target and evidence revision and install reconciliation-only authority. A restarted reconciler retains that target rather than substituting its own newer epoch.
    3. Reconcile only exact unfinished attributed facts, with conditional updates. Commit matching transcript changes and classification together, or use idempotent bounded batches with a durable target. Completed results and independently running children are untouched. Finish in review-required state, not executable state. Recovery does not initialize provider/tool/plugin execution. Promotion/preparation without a terminal fact is also review-required.
    4. Normal release first closes dispatch admission, drains owned callbacks, conditionally records clean release with the captured epoch, and closes the native lock LAST. Uncertain cleanup stays quarantined; a local Idle flag or completed cancellation waiter is insufficient.
    5. Explicit continuation requires review of external outcomes and surviving processes. Acquire the same lifetime lock, re-read the recovery revision, and atomically record an idempotent confirmation receipt AND a new execution claim. Hold the lock through handoff: no cleared-but-unowned gap. Crash after confirmation leaves a tracked dirty claim. Duplicate confirmation returns its prior receipt, never restarts work. New prompts, force-resume and advisory wakes cannot bypass recovery review. Confirmation does not authorize replay of old invocations.

    Mutations and children

    Supported revert/rewind/move/delete/context reset must serialize on the same Session lifetime lock through filesystem effects and durable changes. Preserve authority tombstones independently of transcript deletion. Event replay touching execution state also needs coordination; replay ownerID is not execution authority.

    Recursive subtree mutation and any unintegrated mutation/replay path must fail before effects on activated storage. This is a proposed visible restriction requiring product approval, not silently authorized by the concurrent-host choice. No multi-session lock manager is proposed for the first slice.

    Children acquire their own Session authority. Parent authority gates launch/admission, not child execution lifetime. Commit child completion with a durable idempotent parent-notification intent, then deliver pure notification admission independently. This closes the crash gap between child output and parent delivery without stale parent writes or automatic recovery bypass.

    Cross-host liveness and observation

    Pending inbox data is not wake demand. Complete resume-enabled admission atomically advances durable requested demand; admit-only/noReply does not create demand. A live requester that encounters contention retains a bounded-rate asynchronous recheck until its demand is satisfied or explicitly cancelled.

    Acknowledge demand only with the quiescence check under ownership in a main transaction; later racing admission remains unsatisfied. Preserve local joining and steer/queue semantics. A locally owned interruption suppresses demand only through its observed cutoff, not a later request, and preserves pending input. Locally unowned interruption changes nothing.

    Recovery gates override all demand. Startup discovery settles/reports only; it does not schedule provider work. Orphaned demand remains visible pending until explicit continuation, a deliberate liveness limitation. Active durable-tail subscribers periodically refresh durable sequences as well as listening to local PubSub, deduplicating committed events; live-only fragments remain ephemeral. No IPC broker is required.

    Review and acceptance gates

    Read-only mapping: ses_f50e26603ffe9L4jFNoFgWpXV6. Design reviewer: ses_f50d8fbd8ffejJgiF0rFubh9Z0. Initial review found five gaps: legacy effectful preparation/incomplete admission, restartable recovery transitions, destructive mutations, cross-host wake/tail delivery, and physical-attempt identity. The revised proposal above incorporates those corrections. Follow-up cleared RECORDING FOR USER CONSIDERATION ONLY, not implementation or platform validation. No tests were run for this design.

    Required isolated tests include real two-process legacy/V2 contention, independent Sessions, paused owner without takeover, owned kill/reopen, old-generation late cleanup, permission-held dispatch, admission eligibility, reconcile twice with zero provider/tool/plugin calls, crash around promotion/intent/output/release/confirmation, completed external effect with missing response remaining unknown, child delivery crash recovery, demand handoff/interruption races, remote event refresh, unchanged untracked legacy records, aliases/missing locks, and Bun/Node/platform lock semantics.

    Mixed-version writers ignore new locks and metadata. An additive migration or version flag cannot fence them. No automatic activation on existing storage: separately decide coordinated upgrade versus an explicit upgraded database namespace, compatibility restrictions and rollback that never reopens activated storage with old writers. No namespace migration is chosen here.

    Public recovery status/confirmation and any new contention/unsupported outcomes need separate contract approval and the prescribed Protocol/HttpApi generation. Do not reinterpret active() as a cross-host status endpoint. Canonical glossary/spec alignment belongs to approved implementation, not this candidate.

    Source anchors

    • packages/opencode/src/session/run-state.ts:88-105; prompt.ts:1047-1071,1398-1405,1461-1474; effect/runner.ts:171-202.
    • packages/core/src/session/execution/local.ts:16-28; runner/llm.ts:119-139,390-398; runner/publish-llm-event.ts:74-84.
    • packages/core/src/event.ts:239-351,514-520,565-602; session/context-epoch.ts:72-76,122-173; session/todo.ts:32-56.
    • packages/opencode/src/session/revert.ts:38-79; session/session.ts:606-623; tool/task.ts:233-269.
    • CONTEXT.md:51-52,104,215; specs/v2/session.md:50,162-185.

    Next recommended slice, subject to user review: isolated lifetime-lock feasibility tests before production integration. Native-lock acceptance, integration tests, visible restrictions, activation/rollback and public contracts remain gates. No new source changes, commits, deployment, dependencies, model/configuration changes, live-database inspection or installed-process restarts in this design work. Excluded files remain untouched; #19/#34 and the remainder of #20 stay open.

  8. trmosala commented on Sep 22, 2026

    @trmosala
    OwnerAuthor

    Approved isolated lifetime-lock feasibility experiment published on sqlite-lock-probe at ddaa967: ddaa967c48 . Run: node script/sqlite-lock-probe.mjs. Throwaway code only; no production integration, live storage, installed-process restart, dependencies or recovery activation.

    macOS arm64 results: 84 assertions, 82 passed, 2 failed; exit 1 deliberately preserves failures. Bun 1.3.14 / bun:sqlite / SQLite 3.51.0 and Node v25.9.0 / node:sqlite / SQLite 3.53.4. WAL and DELETE modes exercised the Bun/Node owner-contender matrix, independent databases, same-binding in-process contender/probe close followed by external contention, bounded SIGSTOP pause, owned kill/awaited exit/reopen, and canonicalized symlink aliases.

    Both failures: default opening of a missing resource creates a new database and acquires BEGIN IMMEDIATE. Thus bare default open plus transaction is NOT a fail-closed lifetime-lock implementation. Experimental realpath/file/schema guards reject the tested missing, malformed, empty and directory resources but do not prove race-free identity between validation and open. SIGSTOP delivery is requested but the probe does not independently witness stopped process state. Busy classification uses error text, not native codes.

    All owned children exited before scratch cleanup. Parent source inspection completed; no independent safety review or production acceptance claim. Staged whitespace passed; normal pre-push typechecks passed 31/31, all cached (not proof that this standalone mjs is typechecked).

    Untested: Windows/Linux, bundled Electron Node, hard links/raw aliases, multiple independently loaded SQLite libraries in one process, long-duration/stress behavior, provisioning/replacement races and production lifecycle/fencing. These and the existing activation/rollback/public-contract decisions remain gates. No transcript settlement or provider/tool replay was added. Issue remains open; this completes only the approved experiment, not crash recovery.

  9. trmosala commented on Sep 22, 2026

    @trmosala
    OwnerAuthor

    Further isolated #20 feasibility evidence pushed to sqlite-lock-probe: 24e0880 (test(core): expose SQLite lifetime lock loss). Report: docs/sqlite-lock-probe.md. No production integration, live storage changes, installed-process restart, recovery activation or tool/provider replay.

    The expanded probe ran on macOS arm64 with Bun 1.3.14 / SQLite 3.51.0 and Node 25.9.0 / SQLite 3.53.4: 128 assertions, 122 passed, 6 failed; exit 1 intentionally retains the safety failures.

    • Bun readwrite:true/create:false and Node file URI mode=rw both refuse missing paths without creating files. Positive contention and post-kill acquisition checks pass for those openers.
    • Two prior default-create failures remain.
    • Four NEW failures: a live Node owner in DELETE mode loses exclusion after an unrelated raw open/close of the database file; in WAL mode the corresponding raw open/close of the -shm file has the same effect. Both external Bun and Node contenders acquire. The owner reports an active SQLite transaction after the raw close and is killed only afterward. Transaction status is not proof of continued exclusion.
    • Bun-owner descriptor cases remained blocked in this matrix; this is bounded evidence, not a general safety claim.

    This reproduces SQLite's documented process-wide POSIX descriptor-close hazard. The existing-file-only open modes address default creation but do not establish race-free resource identity or protect against descriptor interference. The proposed in-process lifetime-lock candidate is not accepted for production. Dedicated-holder isolation is a possible next investigation, requiring parent-death/orphan/pipe/fencing/shutdown coverage, not an implemented solution.

    Independent source reviews found no actionable issues in the experiment/report. node --check, formatting and whitespace passed; scoped lint 1 warning / 0 errors. Normal push hooks passed 31/31 typecheck tasks, all cached; that is not proof the standalone mjs was typechecked. Every owned child exited before scratch cleanup.

    Separately, browser recovery passed on macOS arm64 / Electron 44.3.0 / Chromium 152.0.7977.78: --recovery owned crash/relaunch, selected navigation history/Back/Forward/recently closed, fresh private IDs, locked vault, absent sensitive form/history state, zero POST, failed restore/reload/204/Stop/stale callbacks/malformed-store checks. tab-recovery.test.ts passed 1 test / 26 expectations.

    #20 remains OPEN. The storage-activation choice is pending with the user (separate explicitly upgraded database versus coordinated upgrade of every writer). Cross-platform/runtime identity and ownership, settlement-only transcript recovery, explicit continuation, attribution/fencing and public-contract integration remain incomplete. No stale-running incident is claimed fixed.

  10. trmosala commented on Sep 22, 2026

    @trmosala
    OwnerAuthor

    Approved direction: a separate opt-in upgraded database namespace, preserving original storage for rollback; unfinished tasks require review and never resume automatically.

    Pushed c740bd7 on sqlite-lock-probe: a disposable dedicated SQLite holder experiment. All 534 assertions passed on macOS arm64 across Bun 1.3.14, Node 25.9.0 and bundled Electron 44.3.0, DELETE/WAL modes. Covers cross-runtime exclusion, unrelated host descriptor closure, independent Sessions, host/helper crashes, witnessed paused processes, private pipe lifetime, late acknowledgements and stale cached authority after helper death. The latter two first failed and drove experimental protocol fixes.

    Validation: node syntax check, Prettier, scoped oxlint (zero warnings/errors), two-axis review, and push hook (31 cached typechecks). Details and limitations: docs/sqlite-lock-probe.md. Original in-process probe's six failures remain preserved.

    This is not production recovery. All test hosts are external Node; Electron is run-as-node, not utilityProcess. Synthetic dispatch preflight is not atomic with an external effect. Production attempt intents, publication fencing, settlement/explicit continuation, provisioning/path identity, desktop lifecycle, and Windows/Linux validation remain outstanding. Keeping #20 open.

  11. trmosala commented on Sep 22, 2026

    @trmosala
    OwnerAuthor

    Closing at the owner’s explicit request. macOS is the sole target; Windows and Linux are no longer targets, so their validation is not a closure requirement. The browser recovery validation and isolated lock-holder experiment passed on macOS as documented above. This closure does not imply that the experimental holder has been integrated into production: the previously documented production recovery, fencing, and migration work remains unimplemented.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions