Skip to content

Pair gateways as the same person from Devices (#893) - #1296

Open
Varuas37 wants to merge 105 commits into
mainfrom
codex/893-desktop-same-person
Open

Varuas37 wants to merge 105 commits into
mainfrom
codex/893-desktop-same-person

Conversation

@Varuas37

@Varuas37 Varuas37 commented Nov 10, 2026 •

Copy link
Copy Markdown
Contributor

Before: Devices paired phones and agent computers, but had no read-gateway invitation or same-person choice.

After: Devices → Add device → Gateway offers the immutable creation choice. True uses explicit Human reauthentication on a separate pairing-only session; false keeps the panel path. The existing native authority owns activation, attribution and cleanup.

Part of #893. API/CLI, principal/grant persistence and link lifecycle support already merged in #1285. This PR remains draft: broader UI qualification is red, so #893 is not declared complete.

Background

PeerRead gateways require separate conversation share grants. Phones retain OwnerPhone consent and agent computers retain Environment consent. The panel stays Integration. Only the credential registry supplies Human provenance; an administrator role, caller label, namespace or placement's internal Creation::Person category cannot supply it.

#1282/#1286 UI semantics are preserved. Main's #1291 ambient renderer, #1293 Place type and #1301 placement/stress work are integrated without replacing their hunks.

The problem

sequenceDiagram
 participant P as Person
 participant D as Desktop
 participant R as Credential registry
 participant G as Grant store
 P->>D: opens Add device
 D-->>P: offers computers, phones and cloud
 Note over P,D: No gateway invitation entry
Loading

The desktop lacked both the gateway entry and a specified Human proof path. Adding this choice to phone/computer pairing would change their consent classes.

The solution

sequenceDiagram
 participant P as Person
 participant D as Desktop
 participant R as Credential registry
 participant G as Grant store
 P->>D: chooses same person and signs in explicitly
 D->>R: authenticates the Human on a separate session
 R-->>D: confirms the current gateway and organization
 D->>R: creates the immutable PeerRead invitation
 P->>D: compares the claimed fingerprint and approves
 D->>R: approves that exact key
 R->>G: commits attribution after credential activation
 G-->>R: commits the link and its audit
 R-->>D: returns the existing activation receipt
 D-->>P: closes pairing sign-in and explains the separate share
Loading
  • Add Gateway creates PeerRead invitations through the current protocol. Same-person true authenticates the existing Human credential at the panel's pinned endpoint, then compares gateway, organization and audience before exposing pairing methods. The normal panel connection remains separate.
  • Canonical pending/status reads recover uncertain replies without replaying mutations or reconstructing a lost code. One immutable selected value correlates status, code, exact approval outcome and fingerprint with the invitation/key. Retryable, permanent and unfinished-removal outcomes remain distinct.
  • End, terminal status, navigation, pagehide and the scheduled UI deadline retire owned work and late clients. Polling preserves focus. Native enrollment guidance safely quotes POSIX arguments. Credential class, hold.peer, PeerRead, DeviceRead, grant subject and phone consent remain with their existing owners.

How we checked it

Published head: 8fc4301afed597d998d773453d9d9fed9e3501e3. Assembled source checkpoint: 3cb4eecdb524757253fe786152f8a89b54436d52, with main 6c70aeb9 integrated.

Check Result and source boundary
Exact-head CI42856360767 Red: stress 62/71, first failure059 OwnerPhone cancellation; Windows conversation fetch-cancellation join timed out at leases.rs5800. Gateway/native pairing4/4, frontend, contract, coverage, release compile, Linux and Windows server-other passed; Windows and macOS local-auth passed. The final required-check job remains queued without a runner; the failed-job retry request was rejected403 because the workflow is still running. No all-platform pass.
Native gateway/MCP compile Passed at source 3cb4eecd; native/server/Cargo trees match main exactly.
Stress syntax/format and verification units Passed; 682 tests, 2 skips. Gateway/Devices and upstream await/fork/channel hunks are preserved together.
State documentation and independent review 180 documents, zero issues; fresh integration closure has no findings. Desktop checkpoint advances to 3cb4eecd; native and OwnerPhone checkpoints retain their scope. Diagram body unchanged; no TLA/TLC claim.
Complete functional browser matrix Source 88bc8b58: 1,530/1,620 held, exit1, 18 cases unable to run, 7,798.3s. 27 families held, 10 failed and command-order could not fully run. Both engines/layouts/default sizes; separate perf budget omitted. Frontend/SDK/functional check sources remain unchanged; this does not establish new native assembly behavior.
Fresh main controls Source 19954d1f, same functional frontend as 6c70aeb9: 51/133 held in 36 diagnostic groups. They reproduce the red families, with individual-case variation. These are selected controls, not a complete main matrix or performance-rate estimate.
Integrated native/scripted, retained-app and Gateway stress Gateway stress 10/10 default1 / 14/14 raised3 per layout, both engines; retained-app dev 4/4. Initial production aggregate exit2 (WebKit closed before feature entry); isolated native reruns1/1,1/1,2/2 passed. Full production retry exit0, 389.7s, native 4/4. Initial failure remains unexplained. Raised also parallel6, turns4, followers4, rounds40, seed7.
Bounded paired repeats WebKit sidebar focus: main14/23,14/23 vs candidate15/23,14/23; burst fails in all four. Stationary drag all four pass. Reduced motion main pass/fail vs candidate fail/fail. Phone probes candidate5/5 twice, main5/5 twice, Gateway→phone6/6. Complete Ubuntu CI scope: 71/71 uninstrumented, exit0, 546.6s; earlier diagnostic 71/71, exit0, 807.4s; Cancel click→request→positive response→modal removed. Observer overhead can perturb timing; no CI failure explanation or fix is claimed.

At the preceding 88bc8b58 checkpoint, CI42824610437 passed 13/13 after the initial macOS runner-capacity failure was retried. It includes frontend 455 files / 5,802 tests + 44 webhook tests, native pairing 4/4, default full stress 71/71, Windows and Linux/macOS qualification. Local production scripted passed, with Gateway stress 10/10 default and 14/14 raised per layout. These retain their source cut; they do not substitute for final-head native/CI results.

Negative controls cover 45 deliberate unit mutations at the behavioral checkpoint, restored after each; unchanged reducer relocation is recorded by hash. Native probes exercise actual audit-write failure/restoration, stable credential/receiver retry, no implicit share, Integration-attributed revoke and later false enrollment without stale attribution. The obsolete UI and wrapping controls fail their intended assertions and pass restored. Mandatory harness corrections also bite: unexpected masthead copy fails4 assertions; removing only MCP-list whitelist support fails16, restored to smoke32/32 and linked18/18.

Live receipts, exact main-control configurations and unresolved gates. Committed implementation/source checkpoints and control evidence.

Explicit Human sign-in

Immutable gateway invitation

Tradeoffs and follow-ups

Explicit Human token entry is the selected proof path. It adds no durable Integration-to-Human relation or owner-secret fallback. The token retains its existing authority, while only the pairing port reaches its private client. The input clears; no application storage or URL retains it; reconnect is disabled. The ten-minute scheduled UI timer depends on the event loop and is not credential expiry or a server session TTL. JavaScript string zeroization is not promised.

A saved choice or Active phase does not prove that the person link committed or that reads are allowed. The separate share grant and native admission remain canonical. Real devices, provider credentials, native WKWebView/host IPC and the separate performance budget are outside these browser receipts.

Broader failures remain blocking. Static review identifies inherited button/sibling hover styling and stale inline-iframe reads; main controls also reproduce overview/motion, notice, background, bidi and delivery failures. The CI phone-cancel failure is not reproduced by the five selected controls and remains unresolved; the complete local Ubuntu CI-scope diagnostic passed71/71 without reproducing it. This retains the original red CI outcome. The source-identical Windows lease-test timeout also remains unresolved. Main's newer #1299 tooling/lint commit bc8cb5109 was reviewed and textually merges cleanly; it is not integrated into the frozen head. Active #1294 head 945f2ca0 retains sidebar ownership. No assertion or deadline is relaxed and no overall UI pass or merge readiness is claimed.

Rollout

No persistence migration. Reverting the desktop entry leaves stored invitations and their lifecycle with the existing registry.

@coderabbitai

coderabbitai Bot commented Nov 10, 2026 •

Copy link
Copy Markdown

Important

  • Trigger review

This repository does not receive automatic reviews because it has fewer than 10 stars.

Run configuration

  • Source mode: Auto
  • Repository: Unavailable
  • Branch: Unavailable
  • File: No configuration file from this repository applied
  • Commit: Unavailable
  • Layers: defaults
  • Review profile: CHILL
  • Plan: Advanced
  • Run ID: 9b6bc647-752c-4c3f-a557-d36d12c1b8e5
  • Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

Share

Varuas37 commented Nov 10, 2026 •

Copy link
Copy Markdown
Contributor Author

Current qualification at published head 8fc4301afed597d998d773453d9d9fed9e3501e3, assembled source 3cb4eecdb524757253fe786152f8a89b54436d52, main 6c70aeb9 integrated. Implementation is published; qualification is incomplete. PR #1296 remains draft and #893 is not declared complete.

Exact-head CI

Run 42856360767 is red:

  • Conversation stress: 62/71 held, first failure 059 linking: “the cancelled code's modal stayed open” and “Pair a device was not offered again after cancelling.” Later huddle–handoff cases did not run after this failure. Native gateway pairing 4/4, Gateway 016, and Devices 037 passed first. The failed screenshot shows the OwnerPhone modal still carrying its live code, no refusal notice.
  • Windows server-conversation: 1,454 passed, 1 failed, 6 ignored. a_fetch_timed_out_or_cancelled_is_recorded_ended_timed_out_or_cancelled times out joining cancellation at tests/conversation/leases.rs:5800. This Rust source is identical to main; its cause is not established by that equality.
  • Passed: discovery, workflow validity, gateway contract, SDK domain coverage, frontend, desktop release compile, Windows server-other and Linux local-auth. Windows and macOS local-auth both passed. The final Required checks job is queued on ubuntu-latest, with runner_id0 and no steps; the original run is not yet complete. No all-platform pass is claimed. Checkout logs identify synthetic merge f57060972f3299c6ed30f61a56c7c10c23ad2e29 (8fc into6c70); its tree 49373734419afc2507bf051c3eb7f70ef1768249 exactly equals the published head.

Fresh local integrated checks

  • Native gateway/MCP build passed. Verification units 682 passed / 2 skipped, format/syntax and 180 state documents / zero issues passed. Independent source/integration closure found no issue.
  • Initial full production scripted attempt: exit2, native pairing 3/4; WebKit sidebar closed during the unchanged settling wait, after navigation and before entering the pairing feature. No attributable teardown was found. Bounded lifecycle-instrumented WebKit reruns passed 1/1, 1/1, 2/2, with ordinary explicit cleanup and no page-crash events.
  • One complete production scripted retry: exit0, 389.7s. Resize 4/4, gateway-window 11/11, native gateway pairing 4/4, scripted scenarios 8/8, parent-stop passed. Production intentionally does not run the dev-only MCP gateway family. The initial closure failure remains unexplained; it is not erased by the retry.
  • Separate development retained-app check: 4/4, both engines/layouts.
  • Selected Gateway stress: 10/10 default1 and 14/14 raised3 per layout, both engines/layouts. Raised also uses parallel6, turns4, followers4, rounds40, seed7. These selected cohorts are not the complete stress cohort.
  • Focused phone/linking probes, unchanged deadlines: candidate alone twice 5/5 each, main alone twice 5/5 each, candidate Gateway then linking 6/6. No cancellation failure reproduced. UI/reducer/adapter and OwnerPhone class/consent are byte-identical to main. A successful matching cancel response clears the modal before the follow-up read, so the CI screenshot implies no successful acknowledgement reached that transition. A four-second poll/actionability race is only a hypothesis.
  • The complete local reproduction of the checked-in Ubuntu CI scope passed 71/71, exit0, 807.4s including native fixture compilation, with passive Cancel event/disabled/busy and request/response-success diagnostics. Linking059 recorded a trusted enabled/connected click; cancel sent11ms later; positive response35ms after send; canonical read succeeded; dialog gone146ms after click. This is correct ordering in this run, not a fix or an explanation of CI. Independent review found no behavior/deadline/protocol mutation in the observer; its timing overhead can perturb a race. It logs no raw frames, parameters, tokens, codes, keys or wire request IDs (only local serials). Its refused flag means a negative response, not specifically authorization refusal. The workflow explicitly uses --inputFiles 0; the private Swift phone-files product is separately exercised on macOS. The no-flag attempt stopped at the missing Swift verifier before browser execution. A subsequent CI-scope attempt omitted the shared Cargo target and was stopped during duplicate compilation, also before browser execution. Both remain setup attempts, not test passes. The actual complete run used the established shared target; newly created duplicate debug files were removed, preserving older contract targets.

Further cancellation investigation

  • Uninstrumented candidate reproduction at the unchanged published head: complete Ubuntu CI scope 71/71, exit0, 546.6s. This is separate from the earlier instrumented71/71 run. Linking059 passed in both; the original CI failure remains unexplained.
  • macOS completed successfully and its log explicitly passes the exact cancellation test that failed on Windows.
  • Windows baseline #1301 job129905778453 passed that exact test and 1,455/1,455 selected tests (6ignored). Its Rust/Cargo tree equals the candidate's. #1299 job129970348718 also passed that exact test. Current main bc8cb5109 changes manifests/lints, not the relevant production or fixture code.
  • The failing Windows assertion is specifically after exact host arrival, cancellation signaling and successful clock.finish(Duration::ZERO). It then times out joining within min(remaining wall watchdog, 5s); the actual remaining budget is not logged. Final FetchEnded recording deliberately awaits evidence/storage locks and durable save outside the cancellation selection. No obvious lost wake was found. Existing saved_fetch_ends() records save entry, not physical durability. Next Windows evidence must capture remaining budget, cancellation, save entry/completion and task completion without extending either bound.
  • A corrected synthetic gesture probe on current main and candidate records enabled/busy-false at actual pointerdown, then an observed disabled interval, no cancel request and the modal still open in Chromium on both revisions; subsequent normal-click recovery passes. It holds a pointer and delays a scripted poll reply250ms. This demonstrates a shared mechanism under a forced interleaving, not CI's cause. WebKit's corrected probe times out observing the disabled interval after5s on both revisions; its result is not interleaving proof. The earlier probe without the initial-state capture proves only suppression while disabled. The fixture supplies scripted PeerRead status: these probes exercise the actual phone UI's class-independent guard, not native OwnerPhone protocol/consent. State was not captured at pointerup, so no stronger release-state claim is made.
  • Main's green default stress linked at026, while candidate links at059 because the new knob reshuffles the seeded plan. A current-main control retaining candidate-relative payloads/indices/order, omitting unavailable Gateway and expecting main's three tiles, is preparing with all native fixtures consistently from main. It is not identical load: Gateway's messages, restarts, timing and side effects are absent. Temporary relocation initially missed a dynamic SDK import; another preparation mixed checkout roots in the Cargo cache and was stopped before scenarios. Those setup attempts are excluded, not functional failures/passes.
  • One repeat of only the failed remote jobs was requested through the GitHub connector. GitHub returned HTTP403, “This workflow is already running.” No rerun executed. The request is held until that workflow state materially changes; no alternate route is used. The denied Declare two gateways the same person at pairing #893 description update was not retried.

Broad UI gate and main controls

The complete default development functional matrix at 88bc8b58 held 1,530/1,620, exit1, 18 cases unable to run, 7,798.3s: 27 families held, 10 failed and command-order could not fully run. Both engines/layouts/default sizes; separate performance budget omitted. Frontend/SDK/functional run-all sources remain unchanged at the published head; this does not establish new native Rust behavior.

Fresh selected controls at main 19954d1f (same functional source as 6c70aeb9) held 51/133 in 36 diagnostic groups. This is not a complete main matrix or a performance-rate comparison. They reproduce failures in shared hover, focus, drag, responsive/columns, settings notices, subagents, MCP apps, bidi app review, message delivery and command order, with individual-case variation.

Bounded paired repeats on main 6c70aeb9 and candidate 8fc4301a:

Profile Main Candidate
WebKit sidebar full focus 14/23, 14/23 15/23, 14/23
Focus burst within those profiles failed twice failed twice
Chromium sidebar stationary drag target, 1440×900 passed twice passed twice
WebKit sidebar reduced motion, 1440×900 pass, fail fail, fail

Static review locates inherited hover ownership in button/sibling styling and the MCP fixture's stale inline-iframe read before its window-button action. Pairing is not mounted in those seeded flows. Timing/frame assertions and all deadlines remain unchanged. Main reproductions and selected passes establish boundaries, not an overall UI pass.

Disposition and next steps

  1. Keep this draft unmerged until the red UI/stress/native gates have an explicit resolution. Do not infer an attributable fix from a retry passing.
  2. Phone cancellation: the bounded diagnostic cohort completed without reproduction. The local successful request/acknowledgment falsifies a lost click for that run only. The CI failure still needs a matched capture at its failing execution: click/disabled transition, request sent or absent, response and reducer outcome. Do not call it inherited solely because the phone source is unchanged. If reproduced, fix at its existing linked-device owner and prove the unchanged OwnerPhone consent boundary.
  3. Broad UI: preserve active #1294, currently head 945f2ca0, which owns sidebar parking/focus work. Hover ownership and stale MCP frame observations need separate scoped dispositions. Do not absorb those repairs into same-person authority.
  4. Windows cancellation timeout: retain the precise failed test/join evidence and diagnose on Windows without lengthening the watchdog to pass.
  5. Current main advanced to bc8cb5109 (Run changed checks on push, and share one Rust lints table #1299 pre-push/workspace lints). Read-only review found no runtime/protocol/pairing/browser fixture change; merge-tree is textually clean. It has not been integrated into this frozen qualification head. New manifest/tooling checks would need qualification after integration.

The native/API prerequisite is already merged #1285. Desktop Human proof, immutable choice, private-session cleanup and no implicit read grant are implemented here. Committed source checkpoints and fail/restore evidence remain accurate. No real devices, live account grants, provider credentials, native WKWebView/host IPC or deployment were exercised.

Historical receipts at earlier source cuts (not final-head qualification)

Current qualification at published head 88bc8b58df613e90dfb8142cf7d0ff53f052ac91 (main f4c1ba363 integrated unchanged). #893 remains incomplete and this PR remains draft.

The full default development functional matrix finished in 7,798.3s, exit 1: 1,530/1,620 attempted cases held, with 18 cases unable to run. Of 38 families, 27 held, 10 failed and command-order could not fully run. Both Chromium/WebKit, columns/sidebar and all default sizes ran. --skip-perf excludes the separate performance-budget suite; no deadline or assertion was relaxed.

Red family Held / attempted
Shared controls 28/32
Focus 77/92
Drag 173/188
Responsive 52/58
Columns 251/270
Settings 23/24
Subagents 8/12
MCP Apps 88/100
App review 32/36
Message delivery 8/14
Command order 4/8; could not fully run

Passed families include smoke32/32, ambient8/8, safe-area90/90, committed-transcript120/120, tab lifetimes24/24, lease details108/108, gateway states14/14, linked devices18/18, shares26/26, session rows10/10, widgets48/48, approval deadlines2/2, clock snapshots4/4, published files24/24, attachments144/144 and the remaining transcript/turn/send checks. These are current frontend observations, not a global pass.

CI42824610437 passed every executed qualification job, including frontend 455 files / 5,802 tests + 44 webhook tests, native pairing 4/4, full default stress 71/71, Windows and Linux. Initial macOS never acquired a hosted runner (capacity annotations, no steps), which made the required-check gate red. The focused macOS retry is now executing; all-platform qualification is pending. CI checkout cc4921d was verified tree-identical to 88bc8b58.

Production scripted/default/raised Gateway reruns are following the functional matrix serially. Fresh exact-configuration main controls are prepared (36 groups so far), and are not yet results. Main has since advanced to 19954d1f (#1293, Place refactor). Independent source review finds no overlapping/auth/grant/protocol/frontend changes; native assembly against that upstream must be qualified after the frozen queue finishes. Browser frontend/harness source is unchanged by #1293.

Static attribution review found no new pairing startup/iframe route: normal session composition already eagerly imports the SDK; private pairing operations are lazy. The row-hover path is unchanged main CSS filling only the button while its sibling action owns the pointer. MCP openPanel reads its captured inline iframe after fullscreen replaces it with a row; its panel/window/teardown failures follow that unchanged path. Pairing's pagehide handler exists only inside Add Gateway and is absent from that sample flow. Runtime main controls remain necessary; these reviews do not establish timing equivalence. Headed WebKit/WKWebView and the separate perf budget are unrun.

The two corrected verification prerequisites pass and their probes bite: unexpected masthead copy fails4 assertions; removing only MCP inventory whitelist support fails16. Both restore to smoke32/32 and linked18/18. Verification units pass682 with2 skips; independent integration and harness reviews are clean. State checks pass180, with reviewed desktop source checkpoint fe7551adb; unrelated/native checkpoints are retained.

Local receipts: /tmp/893-f4-qualification-all-dev.json, /tmp/893-f4-qualification.json, /tmp/893-f4-harness-probes.json, /tmp/893-f4-collected-results.json. Committed lifecycle and negative-control details are in the evidence record.

The receipts below are historical through 8d9daab94, and retain their own source boundaries.


Desktop verification evidence for #893. This comment tracks qualification; the PR remains draft.

Explicit Human sign-in

Immutable gateway invitation

At behavioral checkpoint 41cec9e2, the disposable native production fixture held in every combination:

ok    choice-human-recovery-native [chromium columns]
ok    choice-human-recovery-native [chromium sidebar]
ok    choice-human-recovery-native [webkit columns]
ok    choice-human-recovery-native [webkit sidebar]

The old-UI control failed new invitation inherited a previous permanent activation stop; restoration passed. The nowrap control failed the constrained command-overflow assertion; restoration passed all four cases. Native audit write failure/restoration, permanent receiver-write conflicts, explicit Human refusal/acceptance, revoke and later false enrollment are detailed in the committed evidence. The reducer relocation preserves the tested body hash; subsequent product checks passed.

Final remote head: 8d9daab946b9ff9649b7e1166ec237840dc8d6da; clean local tree and remote SHA verified. Exact-head CI42803280390 passed all 13 jobs, including frontend 455 files / 5,802 tests, 44 webhook tests, all native platforms and the final required-check gate. Its stress job records:

gateway-pairing: 4/4 held
ok    016 gatewayPairing [chromium columns]
ok    037 devices [chromium columns]
conversation-stress: 71/71 held
ok    001 devices [webkit columns]

The rebuilt local production scripted aggregate passes, including native gateway 4/4. Selected Gateway stress passes in both engines and both layouts: default1 10/10, raised3 14/14 per layout (parallel6, turns4, followers4, rounds40, seed7 for raised). These selected raised runs are distinct from the full default CI cohort.

The broad local functional gate is red: 985/1,102 attempted cases held; sixteen whole checks are dev-server-only. Targeted controls on main reproduced every failing test name. The full paired WebKit-sidebar focus profiles held main 10/23, 10/23, 10/23 versus branch 10/23, 10/23, 11/23; the burst failure occurred main3/3 versus branch2/3. Main also reproduces linked devices 2/18 with mcpServers.list, message delivery 9/14, and the visual/layout failures. The PR description records the full control table, profiles and limitations. Unsupported/setup-invalid control attempts are excluded; no assertion/deadline was relaxed. This remains a draft, not an overall browser-pass or merge-readiness claim.

Both earlier CI harness defects are fixed and independently reviewed. The committed evidence preserves the behavioral control/source cuts; the PR description and this comment are the final qualification record. Real devices, live accounts, native WKWebView execution and deployment were not exercised.

Keep #893 cancellation, focus and consent evidence tied to source ea66af4. Record the passing production and raised stress checks beside the unresolved Windows and broad UI gates; retain the historical model receipt.
Use the row frame as the pointer owner while its trailing action is shown. Keep the control-specific pressed and selected rules unchanged.

Fixes #1325.
Record the unchanged browser regression and its revert probe for #1325.

The control-only hover selector came from 45458da (Desktop: one component per repeated pattern). Keep this repair separate from sidebar parking and pairing.
Keep the failed Linux measurements and add a manual measurement on the
standard Mac matrix leg without changing the budget or required jobs.

Part of #1325.
Add a focus-only measurement plan without changing full/default scope or time bounds.
Keep test frame handles in the native allocator, and guard moved-row setup and identity.
Retain default and raised gateway-pairing coverage in the composed stress plan.
Correct the ADR's stale frame-count description to the existing shared time budget.

Scoped independent reviews are clean. Actual stock focus remains red:82/93 held,
8 failed and3 could not run; the helper repair is not a claim to fix those failures.
No production focus clock, deadline, auth/grant capability or application state changes.
No state checkpoint advances for these verification/tooling-only changes.
Varuas37 added a commit that referenced this pull request Nov 17, 2026
**Before:** Native CI could lose an assertion when GitHub truncated its
annotation or later successful output displaced it. The external-event
HTTP fixture also waited for socket EOF, and ordinary
preparation-fixture teardown cut its transport before strong retirement.

**After:** Failed children retain the existing bounded tail in
correlated annotation parts. HTTP tests consume the actual framed
response under the existing 20-second bound, and ordinary preparation
teardown keeps its transport available through service retirement.
Execution results, admission rules and deadlines retain their existing
owners.

Follow-up verification work from #1325. Its hover repair already merged
in #1337 and #1325 is closed. This PR's twelve-file delta contains
CI/test tooling, fixture repairs and their owning documentation.

## Background

The native matrix runs real Cargo executables. Its original single
oversized annotation did not establish that GitHub received the
assertion. Receiver message/count limits differ from the local bounded
tail.

The preceding published dd7eb81 head passed eleven jobs in
[CI43660899364](https://github.com/nessalabs/nessa-agent/actions/runs/43660899364),
but Windows server-other failed four external-event requests at their
original 20-second read bound. Multipart annotations reconstructed the
complete failure report with verified SHA256
f4d549a204ea62266b468702cc6a44183cf18167c204377e3f55671822ea3d9d. That
proves retained evidence, not the cause of those timeouts.

## The problem

Reporting and test-fixture lifecycle must not invent successful
execution, authority or release. A complete HTTP response does not
require socket EOF; an End request count does not establish
acknowledgement or persistence. An ordinary teardown should let its
existing strong retirement finish before deliberately cutting the mock
transport.

## The solution

```mermaid
flowchart LR
  F["Failed child: existing bounded tail"] --> P["Indexed, correlated annotation parts"] --> R["Reader checks completeness and digest"]
  F --> E["Original child result stays authoritative"]
```

- Existing native reporting retains default libtest capture, bounded
multipart output and original child results even when formatting or
reporting fails. Lease-timeout observations remain bounded and
nonblocking after the original deadline; save entry is an observation,
not a write receipt. Optional Mac measurement retains its original
parameters.
- The external-event fixture uses the existing HTTP client's framing,
explicit local TLS selection and a total 20-second send/body-receive
bound. Exact signed body bytes and physical duplicate headers reach the
actual router. Valid-MAC duplicate-header tests reject before
authority/store/provider effects. Synchronous JSON CPU decoding is not
claimed bounded.
- The child-preparation fixture keeps every earlier owner/capacity,
first actor/cause, Closed, once-End and capacity-reuse assertion. It
awaits unchanged strong service/storage retirement, observes three End
requests, then closes the mock connection. SDK stale-generation
rejection and deliberate disconnect cases remain unchanged.

The original full pre-push attempt at `8f68b1f` failed after 1,883
conversation passes with AttachmentAuthorizationStale in strong
retirement. Sixteen isolated original controls passed, so its exact
intermittent cause remains unproven. The new ordering has its own
executable regression; neither selected passes nor source inspection
establish that original cause.

## How we checked it

Published and remotely verified source:
[3ec743a](3ec743a).
It includes main `a8ee3b188b7ea6eab537c8ffe017ce8bbcdf5895` with all 46
incoming paths byte-exact. The full normal-host mandatory pre-push
passed in 1890 seconds. No hook, assertion, timeout, concurrency setting
or required aggregate was bypassed. Local native desktop host checks
explicitly skip where WebKitGTK is absent.

An intervening mandatory build attempt failed linking the server test
executable with SIGBUS amid severe disk pressure, before native tests
ran. Generated workspace-package artifacts were then cleared after all
checks ended, restoring about 11 GB free; source, raw verification
evidence and the pinned browser CLI were retained. The successful full
retry above uses unchanged build, thread, feature and deadline settings.

New [exact-head
checks](https://github.com/nessalabs/nessa-agent/commit/3ec743acb7bb433671ce8c51e8f5493fd02fad77/checks)
retain their own current status; this publication statement does not
claim a hosted all-platform pass. Prior dd7 failures remain retained.

- HTTP repair: 20 native tests and focused Node/CI contracts passed,
CI21 strict all-target Clippy/test compilation and isolated crate checks
passed, and four actual fault/restoration pairs exercised EOF
completion, missing timeout and both duplicate-header guards. Fresh full
CODE R2 is clean; independent 20 native, 13 Node and 38 adversarial
checks passed.
- Preparation repair: both child variants and adjacent preparation paths
passed 9/9. Restoring cut-before-retire while retaining the new count
executed one failing test (1 versus 3 End requests); restoring the
repair passed 9/9. The full default-parallel conversation library passed
1,884 with zero failures and seven existing ignores. CI21 and isolated
strict all-target Clippy, formatting and docs validation (179 documents,
zero issues) passed. Fresh assembled CODE R1 is clean; independent 19
native and 13 Node passed.
- Prior runner/lease integration reviews and actual Cargo
partition/receiver controls remain source-bound. Counts from overlapping
review suites are not presented as unique tests. The earlier
ENOSPC/linker failure executed no tests and is not negative regression
evidence.

## Tradeoffs and follow-ups

Masking, cancellation, receiver capacity, a later hang and output beyond
the retained tail can still leave incomplete evidence. Digests correlate
content, not unique executions. Default capture can delay an assertion
if a later case hangs. The original Windows timeout and intermittent
stale-ticket causes remain unproven.

This residual changes no product UI, admission, credential class,
same-person attribution, share grant or owner-phone consent. Browser
gate 17 does not apply to this CI/test-only residual; #1296's failed and
incomplete browser/stress/quiet-host performance/native WK/manual
qualification remains required independently. No merge or issue #893
completion is claimed.

The fixture ordering is recorded in existing ADR329 cycle 8; HTTP
boundaries remain in the owning external-events design. No production
lifecycle/model change or unrelated state-page review occurred, so
unrelated sourceRevision/modelCheck checkpoints remain unchanged.

## Rollout

No migration or product setting. Revert the scoped CI/reporting and
fixture changes to restore the earlier behavior.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant