Repository navigation
Conversation
taras
marked this pull request as ready for review
November 18, 2026 09:20
Every established session — ordinary Session/Prompt, held conversations, native launch, client-native attachment and incomplete replay — now uses the executable the current environment selects, reopens the exact retained conversation, and requires the provider's live assertion of the same native identity. Version and digest observations are diagnostic history: a change warns once per logical session per invocation on stderr and never refuses. Missing or different assertions and real resume failures still refuse without creating a replacement. - Generalize strict reopening (expectedAgentSessionId) beyond held Codex. - Select the host executable once, observe that same file, execute it. - Confirm client-native handoffs over ACP: before a native resume, and after a fresh creation once the child joins; exit status is never identity. - Retire build-drift admission and legacy refusals based only on missing historical observation; keep original-record agreement. - Add AcpxProviderDependencies.compatibilityWarnings, shared per CLI invocation; show held Sidekick stderr beside the reply. - Amend ACP, native launch, workflow and REPL specs, architecture, and decision records 004/005; proof fixtures record versions as evidence.
taras
force-pushed
the
codex/issue-891-unified-resume
branch
from
November 18, 2026 09:22
1ee7919 to
b283f08
Compare
taras
marked this pull request as draft
November 18, 2026 09:32
A configured prepared-only replay reopened its conversation, reconfigured it, and then could not give ACP up. Its scope recorded the close failure for teardown and answered success, so the detach retried the close and, when that retry succeeded, recorded a detach and started the native UI. The replay now stops with detach-failed before any detach or spawn; teardown still owes and reports the original failure. NR9 reproduces the reviewed probe: the first close fails, later closes would succeed, and no native launch is requested. The native launch spec no longer says a build-bound agent requires an executable observer.
The user withdrew version history and compatibility notices: a session continues on the executable the current environment selects, under the provider's positive confirmation of its retained native identity, and says nothing about the change. - Remove warnCompatibility, AcpxProviderDependencies.compatibilityWarnings and executableObserver, the CLI invocation set and its forwarding, the workflow diagnostic observer, MachineSessionAssembly's observer, writeDiagnostic, and the Sidekick held-stderr reading and rendering. - Production no longer hashes or parses executables. Selection stays with the host selectors and their launchability check; custom commands are untouched. - Fresh routes are released unbound V1. Existing V2/V3 routes and journals stay byte-preserved; their original bindings are still compared when both exist. - Launch detach/exit ACP work (reconfiguration and native confirmation) now runs under the launch's session route pin, as preparation already did, so a host that pins routes places the post-child confirmation correctly (TL11). - Remove the Test Agent controlled observer and build version; replace test fake observers with a controlled host-selection double. - Amend architecture, ACP, native, workflow and REPL specs and decisions 004/005.
The opt-in live proofs still required bound fresh routes and pinned CLI releases, and the specs still described version reporting and bound V2/V3 fresh construction. - ClaudeNativeToAcp: assert the unbound V1 fresh route and an absent journal binding, check only that route and journal agree, and keep the observed version and digest as recorded evidence. The absent-identity case still publishes a historical V2 route. - Codex native and zero-turn documents: Codex release, digest and platform are evidence; only the carried adapter identity is asserted; fresh routes carry no build. - Claude launch documents: version and platform are evidence. The zero-turn journey accepts XMD's own confirmation refusal (identity-unavailable, no re-entry child) as the exact-identity refusal it is. - Native and ACP specs: fresh construction is unbound V1 and nothing about an executable's version is compared or reported. - Remove the unreadable NativeBinding.versionArgs, the stale boundSessions comment and leftover blank lines.
This branch has not been deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
REPL UI evidence
Mounted typing latency
Viewport 200 × 52. Milliseconds from character dispatch to the completed emulator write containing the changed Message value.
Why
Closes #891. Upgrading Codex or Claude could stop an existing XMD conversation because the executable version or digest changed.
What changes
XMD selects the executable from the current environment and continues the retained conversation after the provider confirms the same native identity. Successful continuation is quiet. Real executable, resume and identity failures still stop the operation.
Before, an upgrade could refuse the prompt (generated session key abbreviated):
After, the prompt completes in the retained conversation. The controlled fixture in the captures returns:
That reply is fixture output. XMD adds no version-change or undetermined-version notice. The same synthetic fixture establishes a session under emulated Codex 0.167.1, then resumes under 0.169.1, checking the exact native assertion, explicit resume target, selected executable environment and requested turn count. Separate integration regressions recover context supplied only to the original conversation.
The paired captures use baseline
76535a3ab46c32645a4431bce9b56b8297bc0017and candidated7f65f3938a53b9ac7796ba56eed91b1dd884f1f, both at 160×36.How it works
The host selects the current executable and carries it into ACP and native execution. The provider uses existing ACPX exact-resume and expected-identity capabilities. Confirmation precedes settings, prompts and established native handoff. Fresh native creation confirms its allocated identity after the child joins.
ACP release joins before native spawn. A failed release stops handoff even if later cleanup succeeds. Delayed confirmation and replay configuration re-enter the host's existing
withSessionRoutepin; the native child runs outside it.Fresh construction uses existing unbound V1 records. Released V2/V3 records and journals retain their original facts. Original accounts still have to agree where both supply bindings; today's version and digest never decide admission.
Review guide
Start with
specs/decisions/005-current-executable-session-continuation.md, then the ACP/native specs andpackages/acp/src/provider.ts.Trace strict reopening, native confirmation/release order,
LaunchInvocation.route, historical record agreement and completed replay. Native NR9 protects failed-release handoff; Test Agent TL11 protects delayed ACP work using the correct host route. Review context-recovery/refusal regressions alongside executable-selection cases.What must stay true
How to verify it
The runtime change at signed
35ea371623e985f3c2facf1199eb867e2831c481passed:The later proof/spec correction at
996e9b9apassed 22 groups/180 steps covering native launch and Test Agent, five executable proof documents in unset-opt-in refusal mode, Deno/Node typing, documentation validation and repository lint. Its production source edits remove an unread member and update comments/whitespace.The fourth correction at signed
d7f65f3938a53b9ac7796ba56eed91b1dd884f1fchanges only the Claude attachment proof fixture. Current version measurement is optional evidence and runs after the negative control. The absent-identity control uses fixed historical V2 metadata. Complete static branch traces cover unfamiliar output and thrown observation; neither can gate the verdict or control. Fixture Deno/Node typing, the document's unset-opt-in branch, documentation validation, formatting and lint passed. No opted-in paid/live proof was run.The synthetic REPL continuation journey and preflight at
d7f65f39, at 160×36 and 72×20, passed. The publishing workflow supplies the exact capture hashes and native paired typing measurements above.Focused commands
Final proof-only checks:
The proof command ran with both attachment opt-ins unset. The final lint command exited 0 with one existing control-regex warning on untouched line 312.
Scope
Current executable selection, confirmation of retained conversations, native attachment/replay, deterministic regressions and coherent specifications/proofs. Warning-only APIs, state and UI plumbing are removed. No new version-history store or journal field is added.
No ACPX vendored runtime, Claude adapter/SDK, producer recipe, generated snapshot, dependency or lockfile changes are required.
Review and delivery status
Current signed head:
bbff2453750997cd5241a496987975eb33d03615, tested treedbdfe69a2116e030bd1fd7f5893e8416f0d8b515. Correction 6 implements the two residual findings below, following correction 5's five fixes. Changed workflow selection joins close and strictly reopens the original conversation; compatible live selection still reuses its handle. Proof addressing, subprocess submission counting, child classification, integration references and optional observations are corrected.All 18 focused final checks passed: 33 groups/186 steps for workflow/provider, 29 groups/247 steps for native/legacy/replay, and four proof cases each under Deno, Node and Bun, plus typing, five unset-opt-in proof documents, documentation validation and affected lint/format checks. The new real-workflow test first reproduced A→A on the unchanged parent for both providers. No opted-in live provider or paid turn ran.
bbff2453is running; prior reviews remain recorded.bbff2453follows.bbff2453. Its prior CHANGES REQUIRED finding remains historical until closed by actual feedback.d7f65f39is published and both attachment files verified; current-head refresh is pending.Findings awaiting correction-6 re-review
session(name)caller produces a details preparation that bypasses the new existing-entry guard. The two-Prompt workflow switch regression passes; named eager resolution still needs the same joined-release guarantee.claude --session-id differentwhen it lacks the expected private instruction-file flag. The actual parsers preserve the call, but the filter hides it from no-child/no-substitution controls. It must count or explicitly reject unexpected interactive argv while excluding known non-UI operations.Controlled real host-selector/provider/wire probes and a synthetic shim trace substantiate these findings without a model turn. The other prior runtime/proof findings are closed by the independent trace. The original Planner froze the finite supplement for source round 6/6, and the retained Astra 6.2 Implementor signed
bbff2453after all 18 focused checks passed. The new permanent tests first reproduced both residual defects on source 5. Exact-commit internal and independent reviews remain due. No further permission is pending. Both prior internal PASS receipts remain preserved as their actual reviews.The PR remains a draft. These feedback reviews do not establish CI completion or authorize merge.