Skip to content

merge(fork): adopt upstream 2.100 AR rebuild as main - #47

Draft
vadymhimself wants to merge 4906 commits into
mainfrom
integrate/main-ar-upstream-2026-11-19
Draft

vadymhimself wants to merge 4906 commits into
mainfrom
integrate/main-ar-upstream-2026-11-19

Conversation

@vadymhimself

Copy link
Copy Markdown
Collaborator

Summary

Adopts the upstream-derived AgentRadar gateway line (PR #46, 6a1446797247680fa4dfb88e3f41caeeb034d746, build 2.100.0-vadymhimself.2) as the fork's main.

  • The head is a merge of the candidate with current main using the ours strategy. Its tree is byte-identical to the candidate. Old main history stays reachable, and future branches cut from the new line share ancestry with main.
  • A trial textual merge produced 87 conflicts, so this PR replaces the tree deliberately instead of hand-resolving them.
  • Main's 10 commits that are not in .27 were audited by content. No runtime behavior is lost:
    • Codex account exhaustion and per-model family quota windows are upstream-equivalent.
    • The 10-minute empty-window ceiling and dated Anthropic tool-type replay are carried.
    • The writeJournal clock fix is carried.
    • Main's 60s combo cooldown cap is intentionally superseded by the hosted 10-minute cap (FORK-DELTA.md).
  • Not carried:

Verification

Evidence for the candidate tree is in FORK-VERIFICATION.md and PR #46. In short: AR gateway gate 534 pass / 0 fail; typecheck, structure, privacy and ratchet pass; full-suite comparison against pristine upstream found no new flag-off failures.

Live trial:

  • Laptop: running since 2026-11-19 10:33 +0800.
  • Personal Mac Mini: running since 16:58 +0800. Before cutover, an isolated sibling run on the Mini passed 38 synthetic pin, SSE and count-authority regressions.
  • Both hosts run with OPENCODEX_AGENTRADAR=1 and are monitored every 5 minutes.

Do not merge before the both-host 24-hour trial completes (2026-11-20 16:58 +0800) and required checks pass.

Checklist

  • Scope stays focused and avoids unrelated cleanup.
  • Docs or release notes were updated when needed.
  • Security-sensitive changes were reviewed for secrets, auth, and unsafe defaults.

🤖 Generated with Claude Code

lidge-jun and others added 30 commits November 11, 2026 18:36
… read (carry lidge-jun#8504) (lidge-jun#8531)

* fix(anthropic-pool): lift 429 quota cooldowns from a background usage read

A fresh usage read that shows every rejected window open lifts an Anthropic
429 quota cooldown (lidge-jun#6989), but nothing issued that read on its own: it ran
only when the dashboard, `ocx account refresh` or a reset admitted through
the proxy asked for quota. A user whose window reopened elsewhere (a reset
grant spent in the Claude app, a plan upgrade) kept getting "All Anthropic
OAuth accounts are temporarily rate-limited" from a pool whose accounts
upstream had already reopened, until the old deadline or a manual
clear-cooldown.

Add an Anthropic counterpart to the Codex cooldown recovery probe worker.
On the shared state sweep it forces a fresh-dispatch usage read for every
instance account whose cooldown is claimable, through
fetchAnthropicAccountQuota so the claim, generation fence and settlement
are the existing ones. A read that keeps the cooldown backs off per account
(one sweep interval, doubling to ten minutes); a newer refusal resets the
backoff; a deadline the sweeper reaches first is never read for; 403
entitlement refusals are never claimed.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

* fix(anthropic-pool): detach the recovery worker and qualify which cooldowns it lifts

Carry of lidge-jun#8504 review findings: register the worker inside the startup transaction and detach it on a failed start and on shutdown; state that only window-backed 429 cooldowns on enabled OAuth instances recover early; test the doubled and capped backoff on one controlled clock; add the windowless Retry-After case.

Co-authored-by: Bruno <bruno@meistrari.com>

* fix(anthropic-pool): never read usage for a cooldown a read cannot lift

A reset-derived 429 that names no rejected window is claimable, but settlement always retains it, so the worker re-read usage every one to ten minutes until a deadline days away. The claim now reports whether it names a window and the worker skips one that does not, along with paused and signed-out accounts. Adds the reset-derived, disabled/API-key instance and backoff-pruning cases.

Co-authored-by: Bruno <bruno@meistrari.com>

---------

Co-authored-by: Bruno <bruno@meistrari.com>
Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>
…fig-root ACL work (lidge-jun#8532)

The integrations writer (catalog fan-out, Aside sync and profile routes, integration routes) and the client disconnect and pre-commit rotation Desktop-store writes now admit after overlapping config-root ACL work through withConfigDirHardeningSettled; an admission timeout refuses without a write.

Refs lidge-jun#7815
)

Keep queued key tombstones through subsequent sets and apply combo-prefix
clears to the fresh file under the mutation lock. Add regressions for
shorter replacements, peer rows, retry retention, and clear ordering.

Follow-up to lidge-jun#8117.

Co-authored-by: JUN <jun@lidgeai.com>
…idge-jun#7345 part 1a) (lidge-jun#7918)

* refactor(codex): read the native-main auth.json through one module

Every read of the bytes of <CODEX_HOME>/auth.json now calls
src/codex/native-main-auth-file.ts. The module still reads the file on
each call and keeps each caller's read mode (plain, bounded,
local-bounded, or the stat-guarded read of the desktop identity reader),
so results, statuses and error messages do not change.

readBounded and MAX_AUTH_BYTES move to the new module (re-exported from
native-profile-store.ts) to avoid an import cycle.

Part 1a of the split plan in lidge-jun#7345 (from the closed lidge-jun#7344).

* test(codex): guard the auth.json literals in src with a file:line allowlist

Every src line that names "auth.json" must be on the allowlist, with the
reason it is allowed. A new reader of the native-main login therefore
fails this test until it goes through src/codex/native-main-auth-file.ts
or is added on purpose.

Part 1a of the split plan in lidge-jun#7345 (from the closed lidge-jun#7344).

* refactor(codex): share the native-main auth.json stat stamp (CodeRabbit review)

* refactor(codex): write the native-main auth.json through one module (lidge-jun#7345 part 1b)

---------

Co-authored-by: JUN <jun@lidgeai.com>
… native main (lidge-jun#8536)

* feat(codex): opt-in pool-only policy keeps automatic Pool routing off native main (lidge-jun#7410)

codexPool.poolOnly (default off) makes automatic Pool resolution serve only registered, eligible non-main accounts. One decision in codexAccountUnusableReason (reason pool_only) keeps selection, preserve paths, subagent fallback and previews in agreement; resolveCodexAuthContext additionally skips request-owned main pin preservation, stored-main discovery, both caller-main fallbacks and the no-candidate main policy check, and refuses if any selector still names main. Explicit selectors, Reserve and Direct stay outside the policy.

* fix(codex): keep explicit __main__ and Reserve selectors outside the pool-only policy

The fixed-account check and dispatch revalidation consulted usability without the explicit opt-out, so an exact stored __main__ selector and every Reserve request were refused under codexPool.poolOnly. Fixed selectors now pass explicitAccountSelector, and isCodexAuthContextUsable does for a fixedAccount main-pool context. Adds exact stored-main, automatic-context reuse and pinned-main preview cases; documents the sidecar login rule.
…-jun#8537)

* docs(compaction): point compaction combos at cooldown recovery (lidge-jun#7850)

* docs(compaction): use the site-relative combos link
…pervisor is gone (lidge-jun#8534)

Windows records the Job Object supervisor pid as the update marker's installer witness before the supervisor receives its request, and inspectUpdateMarker frees an unconfirmed Windows marker once the fence, the launcher and the supervisor are gone, so an interrupted update no longer blocks every later update and sync.

Closes lidge-jun#8524
lidge-jun#8535)

Home and the sidebar polled /healthz, which a Remote Hub's management ingress answers with 404, so a healthy hub read as offline. Home and the self-update reload check now read /api/system/health; the sidebar keeps /healthz only for a connected Child's machine listener and treats 401/403 as reachable.

The Codex Auth Pool/Direct badge is now a button that opens #providers?provider=openai&focus=account-mode, focusing the Codex account mode control.

Closes lidge-jun#8506
Closes lidge-jun#8501
… native OAuth lane (lidge-jun#8507)

* fix(anthropic): let Claude Code auto mode's server review through the native OAuth lane

Claude Code in auto mode asks the Anthropic API to review risky tool calls
as part of each model request. A session without feature flags (telemetry
off, or behind a gateway) has no local classifier to fall back on, so on
the managed native Messages lane over Anthropic OAuth every reviewed call
was denied ("[server-classifier] ... server_no_result"). Three causes:

- The dangerous-tool-use-2026-09-03 beta and its top-level `safeguards`
  body field were not admitted. They are now one nativeClientOnly
  beta/body pair in beta-body-pairs.ts, the same policy as
  context_management.
- OAuth renamed every declared client tool to `custom_<name>`, so the
  server review matched none of the calls (`tool_uses: {}`). An observed
  Claude Code client already sends the tool names its own OAuth login
  uses, so it now keeps them; the reverse map is empty and the verdicts in
  message_delta.delta.safeguard_results reach the client unchanged. Other
  OAuth callers keep the prefix; the collision checks are unchanged.
- Background sessions (`claude --bg`) send `X-App: cli-bg`, but client
  identity accepted only `cli`, so they fell back to the generic shape and
  the OAuth API refused them with 400. `cli` and `cli-bg` are accepted.

Regression tests cover the pairing, the unprefixed names for an observed
client (and the prefix for a generic one), the end-to-end native OAuth
lane including relayed verdicts, and the cli-bg identity. The managed
native Messages contract in structure/data-planes/protocol-paths.md and
the protocol-paths guide are updated.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* docs(protocol-paths): state what copying Claude Code identity headers turns on

Security review follow-up for lidge-jun#8507: recognition is header-based, so a client
that copies Claude Code's identity also sends unprefixed tool names on the
OAuth account, and an operator-added dangerous-tool-use beta forwards
safeguards for unrecognized callers whose names stay prefixed.

Co-authored-by: david-wang-0 <72378768+david-wang-0@users.noreply.github.com>

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: JUN <jun@lidgeai.com>
…idge-jun#8539)

The Child Remote Link tunnel supervisor records each tunnel state transition (spawned, connected, stopped, spawn_failed, classified ssh exit, timeout) through the existing bounded LinkDiagnosticSink with fixed fields only, matching the Home.

Refs lidge-jun#7297
…Mistral Chat (lidge-jun#8432)

* fix(chat): strip replayed reasoning_content from message history for Mistral Chat

Reasoning-capable Chat clients (OpenCode and other OpenAI-compatible tools)
replay the assistant reasoning_content their responses carried back into every
follow-up request. The Mistral wire policy from lidge-jun#8278 already omits the
top-level reasoning controls, but per-message fields pass through verbatim,
and Mistral's strict Chat schema rejects the replayed field with HTTP 422 —
killing every follow-up request on Mistral-destined routes (lidge-jun#8420).

Extend the canonical-Mistral Chat destination policy, the final wire
projection shared by the native and translated builders, to drop
reasoning_content from message history at egress. Touched messages are
replaced by stripped copies, so caller-owned message objects are never
mutated; only canonical api.mistral.ai key-auth destinations are affected, and
other destinations are unchanged.

Tests: four new cases in tests/adapters/openai/openai-chat-mistral-policy.test.ts
covering the strip on the native builder, untouched passthrough for
non-Mistral destinations and auth modes, and caller-object immutability
through applyChatDestinationPolicy.

Generated by Mistral Vibe.
Co-Authored-By: Mistral Vibe <vibe@mistral.ai>

* fix(chat): keep caller history immutable while stripping Mistral reasoning

Reviewer feedback on lidge-jun#8432: replacing array slots edits the caller's
messages array, breaking the projection's caller-isolation contract —
retries and combo siblings may read the history after the Mistral
projection runs. Build a copied array holding stripped copies of only
the touched messages, mirroring the replay planner's caller-safety rule,
and swap it in only when something was stripped.

Generated by Mistral Vibe.
Co-Authored-By: Mistral Vibe <vibe@mistral.ai>

* fix(chat): keep native reasoning replay from restoring reasoning_content on Mistral

The native passthrough applies its reasoning-replay plan after the Mistral
destination policy strips message reasoning_content, and the plan rebuilds from
the pre-strip history. With preserveReasoningContentModels set, the client's
reasoning_content came back and Mistral answered 422. Canonical Mistral Chat
now gets the no-op replay plan. Docs-site states the history strip (CodeRabbit).

Co-authored-by: Masih-0x3 <105122885+Masih-0x3@users.noreply.github.com>

---------

Co-authored-by: Mistral Vibe <vibe@mistral.ai>
Co-authored-by: JUN <jun@lidgeai.com>
lidge-jun#8044)

* fix(antigravity): surface Argon thought text and stop a quoted tag from truncating answers

Gemini 4 Argon answers through CCA with thought parts only when the request names
its tier in thinkingConfig. The Argon wire ids already carry the tier as a suffix
(gemini-4-argon-high), so the adapter sent includeThoughts alone; live CCA then
returns 200 with thousands of thoughtsTokenCount and zero thought parts, and the
plan arrives as ordinary visible text. Measured 2026-10-27: the same request with
thinkingLevel repeated alongside includeThoughts returns the thought stream, while
gemini-3.8-flash does not need it and keeps its one-field shape.

Separately, the Codex client treats a literal <oai-mem-citation> as markup and
drops everything between the first opening tag and the last closing tag. A model
that quotes the tag in prose loses the rest of its answer, which reads as a turn
truncated at a backtick. The bridge now escapes mentions while leaving a real
citation block (one whose body starts with <citation_entries>) byte for byte, in
both the streaming and buffered paths.

* fix(antigravity): hold streamed citation blocks until closed, share escaper state in the adapter encoder, keep the lone Argon level

* refactor(antigravity): move the Argon tier change to its own PR; charge held citation bytes to the translator budget

* fix(bridge): one size rule for streamed and final citation blocks; escape signed progress text's visible copy

* fix(bridge): escape quoted memory-citation tags on the Responses wire only

Maintainer follow-up to the review on lidge-jun#8044:
- Chat Completions and Messages clients never parse <oai-mem-citation>, so the
  adapter-events encoder changes are reverted and both Responses bridges take
  escapeMemoryCitationTags (default true); Responses delivery and run-turn
  paths pass false for chat/anthropic inbound wires. stripCitationMarkers no
  longer escapes on its own.
- A real block needs only a complete <citation_entries> list; the client's
  parse_memory_citation accepts a block without <rollout_ids> or <thread_ids>,
  so requiring one regressed dev.
- Comments and contracts state the client hides to the FIRST closing tag (or
  EOF), that the hold limit is UTF-16 units, and the code-span trade-off; the
  progress mirror contract names the escaped form.
- Tests: Chat/Messages paths keep raw text (stream and buffered), buffered
  Responses escapes, an entries-only block stays real, and the progress item's
  visible text is asserted directly.

Co-authored-by: andrew05060414 <59988150+andrew05060414@users.noreply.github.com>

* fix(bridge): keep sidecar and fold text raw for Chat/Messages; linear candidate scans

Review follow-up on lidge-jun#8044:
- The web-search and image sidecar loops bridge every inbound wire; they now
  take escapeMemoryCitationTags and sidecar-execution passes false for chat
  and anthropic, so those clients receive the raw tag. The Chat/Messages
  completion fold also keeps the raw text its fingerprints compare.
- A block carrying only <rollout_ids> or <thread_ids> is real, matching the
  client's parse_memory_citation.
- Candidate scans look at most one hold limit ahead and share the next
  closing-tag search; outside claimed blocks one regex pass escapes. A 3 MB
  input of unclosed candidates went from ~10 s to ~0.1 s.

Co-authored-by: andrew05060414 <59988150+andrew05060414@users.noreply.github.com>

---------

Co-authored-by: JUN <jun@lidgeai.com>
…ow name (lidge-jun#8533)

* fix(anthropic): resolve omitted output caps by Claude model id, not row name

A custom-named adapter:anthropic row (or a registry row that seeds no cap)
received no defaultMaxOutputTokens/modelMaxOutputTokens, so every Codex turn
fell back to max_tokens 8192 (8192/16384/24576 by adaptive effort) where the
same model on anthropic-apikey got 64000-128000.

routedProviderConfig now fills the routed copy from the registry by Claude
model id: operator per-model caps win, any existing default (operator or
registry-seeded) suppresses the fill, api.anthropic.com also gets the 64000
Anthropic default, and unknown gateway ids keep 8192. Runtime-only.

Closes lidge-jun#8525

* docs(structure): own anthropic output caps in their own provider doc

structure/config.md sits at its 600-line budget, so the lidge-jun#8525 contract moves to
structure/providers/anthropic-output-caps.md, registered in the manifest.

* test(layout): register the output-cap test in the adapters shard

layout.json is at the 2000-line ratchet threshold.
* feat(responses): scope custom tools by model

* test(responses): move custom-tool capability registration into shard

* fix(responses): classify compact custom-tool refusals as client errors

* test(responses): cover compact retry policy and model-settings admission

* docs(responses): keep capability guidance within document budgets

* fix(responses): keep the native compact wire when projecting custom tools

Native compact projected the full route wire (adapter and GCP marker strip)
through resolveWireProtocolOverride, so an official OpenAI provider with a
modelAdapters chat override rewrote its native compaction blob into a plain
message. Compact now projects only the captured custom-tool flag through
projectCapturedCustomToolPolicy, which resolveWireProtocolOverride also uses.

Docs: registry defaults count as the provider value, and the dashboard
Restore leaves the API-only responsesCustomTools axis in place.

Co-authored-by: Platinum-Scripts <133908028+Platinum-Scripts@users.noreply.github.com>

* fix(responses): lower compact custom tools only for an explicit false

With no declaration, native compact rewrote every custom tool except
apply_patch (for example code-mode exec) into a function tool, which changed
the default compact body. The compact rewrite and final check now run only
when the resolved supportsResponsesCustomTools is false, so the default body
is unchanged. ocx models set --reset now also clears the API-only
responsesCustomTools axis, matching its "clears every override" contract.

Co-authored-by: Platinum-Scripts <133908028+Platinum-Scripts@users.noreply.github.com>

---------

Co-authored-by: JUN <243035832+lidge-jun@users.noreply.github.com>
Co-authored-by: JUN <jun@lidgeai.com>
…bs API (lidge-jun#8540)

Adds Codex -> Codex CLI: a read-only Check over the existing plan route, and one explicitly confirmed Apply of exactly the displayed planId, followed to its final job state. Idempotent resend only for busy/admission/network failures; 403 shows local pairing and the CLI command; connected dashboards send nothing. No server change.

Closes lidge-jun#2811
…usal (lidge-jun#8541)

A refusal caused by an update marker that records an unconfirmed installer with no witness (left by OpenCodex 2.100.x and earlier on Windows) now prints the marker path and a safe rename command after the unchanged update-in-flight line; decisions, exit codes and the sync JSON are unchanged. docs-site troubleshooting covers the recovery.

Refs lidge-jun#8524
…idge-jun#8544)

The CLI parity sweep scans GUI source for /api/ literals; a comment in the Codex CLI update tab named /api/codex/* and read as an uncovered endpoint, turning tests/cli/cli-headless-parity.test.ts red on dev since lidge-jun#8540.
…-jun#8542)

Every PR event used to start three pull_request_target runs. pr-gates.yml now owns the pull_request_target and status triggers and calls pr-labeler.yml, pr-hygiene.yml and enforce-pr-target.yml as local reusable workflows (workflow_call-only, logic unchanged), so one event starts one PR gates run. permissions: {} at the top with per-job ceilings equal to each gate's requests; concurrency stays on the gate jobs. CI policy unchanged.

Closes lidge-jun#7654
* feat(advisor): add opt-in question-only consultations

Allow workers to consult the advisor without automatically attaching task
history, tool results, tool catalogs, identities, or repair evidence.
Preserve existing defaults and consent gates; pause automatic consultations
in question-only mode and require an explicit non-empty question.

* fix(advisor): preserve truncation notices and CLI discovery

Retain a bounded overflow marker through tool interception so the advisor
receives an explicit truncation notice. Cover short, limit-sized, and
oversize questions at the routed expert-provider boundary.

Publish context mode and its automation exception in the canonical CLI
capability surface. Explain truncation consistently in every GUI locale.

* fix(cli): align advisor help with context mode

Include context mode and task limits in the registry's option list and
scope automatic-policy descriptions to context mode. Exercise both real
help entry points with isolated homes and HTTP disabled.

* fix(advisor): exclude invalid questions from consultation allowance

Use synchronous runtime-owned admission before the guard reserves a manual
consultation. Empty question-only calls retain paired validation results
without consuming the consultation allowance. Keep the independent worker
continuation limit unchanged and cover correction and repeated-invalid
flows through the routed provider boundary.

* fix(advisor): preserve Unicode questions and Google compatibility

Enforce question bounds at runtime without lossy schema constraints on
Google's strict tool-schema path. Share a core-owned, Unicode-aware bound
between interception and prompt construction, preserving surrogate pairs
and an idempotent truncation notice.

Cover all Google endpoint profiles and astral questions below, at, and
above the limit at the routed expert boundary.

---------

Co-authored-by: JUN <243035832+lidge-jun@users.noreply.github.com>
Co-authored-by: JUN <jun@lidgeai.com>
 (lidge-jun#8548)

The Windows supervisor tests loaded a data: URL copy of npm-install.mjs that pinned only ./mutation-guard.mjs, so the relative imports lidge-jun#8534 added could not resolve; a shared loader now pins every relative ./*.mjs import and a cross-platform case loads the copy on every host. lidge-jun#8534's win32 rule that skipped a recorded installerPid unless terminationUnconfirmed is reverted; confirmInstallerExit already drops the Windows witness.

Refs lidge-jun#8524
…nd writes (lidge-jun#8547)

On Windows, runCli waits once for the optional config-root ACL flight before an eligible command's first write, bounded by one ACL deadline and proceeding on timeout; preflight-exempt commands, ensure, internal and hidden helpers and the server are skipped. The server's synchronous writers stay synchronous, with their residual risk documented in structure/config.md.

Closes lidge-jun#7815
…#8551)

* feat(link): opt-in SSH log capture for Remote Link tunnels

Add `ocx link capture start [--minutes N] | stop | status`, a local,
operator-only window in which this machine's Remote Link tunnels (Home -R
and Child -L) run ssh with LogLevel=DEBUG1 and store a sanitized copy of
ssh stderr under <OPENCODEX_HOME>/link/ssh-capture/.

- Off by default. The only authority is an unexpired control.json the CLI
  writes locally; it is not a config key, so sync, management routes and
  remote input cannot enable it. A symlinked, foreign-owned, group/world
  writable or over-long window reads as off. Without the capture directory
  a spawn costs one lstat and the argv and runner call are unchanged.
- Window: 15 min default, 60 min hard max. Stop or expiry stores nothing
  more and the next spawn is back at LogLevel=ERROR.
- Storage: one owner-only, no-follow, append-only file per window, capped
  at 1 MiB with a capture_truncated marker; Windows uses the private ACL
  helpers. Files are deleted 72 h after their last write (spawn, CLI, and
  an unref'd writer timer); start keeps at most four. Stop keeps files.
- Redaction is an allowlist of OpenSSH line shapes rendered with fixed
  placeholders for hosts, addresses, users, ports, paths, keys,
  fingerprints, commands, env values and server messages; only digits and
  closed-set algorithm/method/errno names survive. Unknown lines are stored
  only as the category "unrecognized" with a count.
- Classification is unchanged: the runner gains an optional stderr line
  tap; debug lines and INFO-only shapes are kept out of classifySshStderr,
  and the 64 KiB cap applies only to the kept text, so verbose output
  cannot end a healthy tunnel.

Docs: Remote Link guide troubleshooting section, CLI lifecycle reference,
structure/remote-link.md.

Refs lidge-jun#7297

* fix(link): capture retention survives restarts; INFO-only ssh lines stay out of the exit class

Refs lidge-jun#7297

* fix(link): a running tunnel follows a capture window extension

Refs lidge-jun#7297

* fix(link): keep the remaining INFO/VERBOSE-only ssh lines out of the exit class

Refs lidge-jun#7297

* fix(link): a capture marker ends the file for every process; split stderr pieces keep their line decision and the last line counts toward the cap

Refs lidge-jun#7297

* docs(link): accurate level comment in the ssh capture redactor
…idge-jun#8546)

* docs(devlog): design for a Windows service stop that frees the port

Design only, for owner approval; nothing is implemented. Option B from the 2026-10-28 design comment as one checked stop shared by ocx stop and ocx service stop, with explicit stopped/survived/denied/unknown outcomes, bound per-installation identities, and a verification plan that separates hermetic tests from real Windows device checks.

Refs lidge-jun#7771

* docs(devlog): correct the Windows stop design against the real call paths

Who calls which stop (stopWindowsChecked serves ocx stop, uninstall and /api/stop; the updater uses guardedStopWindows), the post-stop checks that already exist, an explicit binding rule including the stray wrapper, a creation-time reader and handle-based terminate as named new work, and the outcome mapping onto ServiceStopOutcome, ocx-stop/1, /api/stop and Desktop.

Refs lidge-jun#7771

* docs(devlog): terminate only wrappers; map outcomes onto existing ocx-stop/1 values

The checked stop never hard-kills the proxy child: it keeps the graceful stop, its drain, the teardown-deferral receipt and the refusal rule, after the operation has ended the manager and wrappers. The outcome table uses only existing outcome/proxy values per path (respawned is a proxy value under outcome failed), plus additive fields. Also: the scheduler service host as a launcher parent, a minimum observation window replacing the respawnable re-poll, and the updater authorizer before each terminate.

Refs lidge-jun#7771

* docs(devlog): split the Windows stop into a manager/wrapper part and a caller port check

Callers stop the child after the service stop and /api/stop runs inside the proxy, so the port check moves to the caller after its graceful child stop. Adds ocx service uninstall/remove and prepareServiceInstall to the covered paths, proves the scheduler parent through the Schedule service PID, compares parent creation times, states the unbound-wrapper outcome and the deadlines, keeps WinSW updater behavior, and resolves ARM64 with a recommended bounded fallback (owner choice).

Refs lidge-jun#7771

* docs(devlog): give uninstall its port check, budget the consented stop, close the binding gaps

Part B now also replaces the uninstall respawn-window check and runs after the guarded settle (kept); non-stopped outcomes abort before destructive steps; /api/stop with a receipt skips part A; worst-case budgets per caller (28 s against Desktop 30 s); descendant of a BOUND wrapper and an explicit choice for hand-started copies; the guarded inspection gains a bound-stray state; new work and the async return type listed; ARM64 fallback residuals stated; stale deadline and caller counts fixed.

Refs lidge-jun#7771

* docs(devlog): plain ocx stop never stops the child after a failed part A; one deadline for the consented stop

* docs(devlog): absent outcome for part A; admission-checked deadline passed by each Desktop path

* docs(devlog): receipt records part A; bind orphans only under a registered task; foreign port stops install; ARM64 device

* docs(devlog): cap every post-admission step, budget WinSW and the registration probe, keep a 10 s POST without a receipt

* docs(devlog): final enumeration after the hold, counted; service stop defers teardown through a receipt

* docs(devlog): budget the POST at 10 s always and the exit ladder in sequence; Desktop exit drain 65 s

* docs(devlog): service stop sends no undeferred child stop when the receipt claim fails

* docs(devlog): correct WinSW and receipt-probe arithmetic; one place for the stray enumeration

* docs(devlog): three-cell WinSW row; Desktop exit drain 75 s with a pre-admission allowance
…hboard) (lidge-jun#8555)

* feat(link): read-only Remote Link diagnostics viewer (CLI, route, dashboard)

The bounded <configDir>/link/diagnostics.jsonl (+ .1 backup) had a writer and no reader.

- src/link/diagnostics-read.ts reads both files read-only, no-follow and non-blocking,
  refuses symlinks, FIFOs and a symlinked link directory, reads at most the newest 256 KiB
  of each, and re-applies the writer's allowlist (linkDiagnosticMetadata, now exported)
  plus timestamp/suppressed to every line. Malformed, overlong or disallowed lines are
  counted in `skipped` and never echoed. Oldest file first, optional since, limit 1-1000.
- `ocx link diagnostics [--limit <n>] [--since <iso|epoch-ms>] [--json]` reads the local
  file without the proxy, is exempt from the shim-repair preflight (and so from the
  config-root hardening wait), writes nothing, and prints the path when nothing is
  recorded yet.
- GET /api/link/diagnostics returns the same ocx.link.diagnostics/1 document without the
  path, with the same authority as GET /api/link/status, before the link lifecycle is
  resolved. A connected Child's machine listener serves the same read to its dashboard
  session only, never relayed to the Home.
- The Remote Link page gets a collapsed Diagnostics section that loads only when opened
  and re-validates records against the closed vocabulary; copy in all 11 locales.

Refs lidge-jun#7297

* fix(gui): count-neutral wording for skipped diagnostics lines

Refs lidge-jun#7297

* fix(link): diagnostics viewer survives out-of-range timestamps; Child reads its own diagnostics on the machine plane

Refs lidge-jun#7297

* test(gui): Port is the same word in French for the diagnostics port column

Refs lidge-jun#7297

* docs(link): skipped counts rejected lines; refused files appear in files

Refs lidge-jun#7297
…ry (lidge-jun#8553)

advisor-question.ts (added in lidge-jun#7526) is reached from advisor-slot.ts but was missing from RESPONSES_CORE_MODULES, failing responses-core-modules.test.ts on dev.

Co-authored-by: lidge-jun <243035832+lidge-jun@users.noreply.github.com>
…rvives a full SSH capture queue (lidge-jun#8559)

Fixes the windows 8/11 and 1/11 failures in dev batch 43045070937.
…s stop reasons (lidge-jun#7771 slice 3) (lidge-jun#8560)

* feat(desktop): hand the stop deadline to ocx stop and read the Windows stop reasons

Desktop's two ocx stop children now receive their deadline as
OCX_STOP_DEADLINE_MS (decimal milliseconds from the child's start): the
consented takeover passes the remaining startup deadline minus 5 s, the
exit/restart drain passes its deadline minus 3 s, and a value that would
not be positive is not set. An older CLI ignores the variable, so mixed
Desktop/CLI versions keep today's behavior.

The exit/restart drain deadline is 75 s on Windows (30 s elsewhere), so
a registered service's checked stop fits.

StopSummary reads the additive serviceStopReason and portHeldByForeign
fields with serde(default). They change no verdict and add no Outcome or
Proxy value; they append an actionable sentence to the stop description
(elevated terminal, retry from a terminal, wrapper started outside the
service, unreadable process state, another program on the port). Older
documents read exactly as before and an unknown reason is tolerated.

Refs lidge-jun#7771

* fix(desktop): always set the stop deadline variable; tolerate a null portHeldByForeign

Refs lidge-jun#7771
* fix(usage): yield between costly ledger rows

* fix(usage): yield inside a chunk on elapsed time only

A per-row 1,000-row check yielded several times per chunk of cheap rows and
made cold scans of large ledgers slower than dev. The row limit stays at the
chunk boundary; inside a chunk only the 8 ms budget yields. Assert that a
paused scan publishes the same checkpoint digest as an unpaused one.

Co-authored-by: seedds <seedds@gmail.com>

---------

Co-authored-by: seedds <seedds@gmail.com>
Co-authored-by: lidge-jun <243035832+lidge-jun@users.noreply.github.com>
* fix(usage): page filtered index candidates

* fix(usage): release filtered readers and fall back on page errors

- Close the candidate reader on the unchanged and generation-mismatch
  branches of refreshFilteredAggregate; only the paging paths closed it,
  so repeated unchanged reads leaked one SQLite connection each.
- Wrap index-side candidate-page failures in UsageFilterIndexPageError and
  answer from the streaming ledger scan instead of rejecting the request;
  generation changes keep the rebuild retry.
- Give the write-contention degradation test a 20s timeout: the extended
  busy-retry window takes several seconds by design.

Co-Authored-By: Epinephrine <luvs01@hanmail.net>

* test(usage): give the two-process rebuild contention test room

The extended busy-retry backoff plus a 30k-row double scan pushes this
test past the default 5s timeout on CI (observed 5002ms).

Co-Authored-By: Epinephrine <luvs01@hanmail.net>

* fix(usage): retry streaming fallback after index page failures

Keep bounded retry handling for ledger and aggregation-input changes after the index pager becomes unavailable. Cover ledger replacement, overlay changes, and real SQLite pager rollback/error wrapping in active and passive reads.

* test(usage): drop stray patch markers from filter-index tests

Leftover '+' diff-application artifacts between test blocks; they parse as
unary-plus expressions but carry no meaning.

Co-Authored-By: Epinephrine <luvs01@hanmail.net>

* fix(usage): heal corrupt index pages and split the rebuild retry schedule

- nextPage() requests an index rebuild when a page fails with SQLITE_CORRUPT/NOTADB, so a damaged projection heals instead of degrading every query to streaming\n- UsageLedgerRebuildRequiredError retries on the original short schedule; the 9-delay backoff stays for SQLITE_BUSY contention\n- document per-page transactions on close(), name the fallback retry contract in the structure doc, and fix the stale one-transaction claim\n- cover the per-page generation recheck against real SQLite with a rebuild between pages

Co-Authored-By: Epinephrine <luvs01@hanmail.net>

* test(usage): cover the mid-aggregate rebuild retry

Publish a new generation between the first and second candidate pages of a live aggregate, asserting the restarted paging matches the full-scan oracle with fresh totals.

Co-Authored-By: Epinephrine <luvs01@hanmail.net>

* docs(usage): docstring the paged read path

Document candidatePager.nextPage, readActive, accumulateCandidates and refreshFilteredAggregate.

Co-Authored-By: Epinephrine <luvs01@hanmail.net>

* fix(usage): re-check transient index-open classifications before degrading

Two racing writers can observe the projection mid-creation: classifyDatabase
sees the schema's statements commit one by one, and a journal sidecar can
vanish mid-stat in preparePrivateSqliteDatabase. Both produced an instant
foreign/unsafe degrade to streaming even though the index was healthy a
moment later; the two-process rebuild test observed exactly that as
candidates:null. Re-check both on the short rebuild schedule; only a home
change still degrades immediately.

Co-Authored-By: Epinephrine <luvs01@hanmail.net>

* test(usage): cover transient index-open classification recovery

- transient foreign file clears mid-retry and a later attempt builds the index
- transient unsafe sidecar clears mid-retry and recovers the same way
- permanent unsafe sidecar still falls back to streaming, unmodified

Co-Authored-By: Epinephrine <luvs01@hanmail.net>

* docs(usage): align retry wording with the atomic schema contract

The schema now commits in one BEGIN IMMEDIATE transaction, so the stale
partial-schema justification for the foreign/unsafe re-check is replaced;
the structure note names the classifications that share the short schedule.

Co-Authored-By: Epinephrine <luvs01@hanmail.net>

* docs(usage): align busy-retry bounds and test query shape with current contract

* perf(usage): push the page cursor into each candidate branch

The outer WHERE byteOffset > ? on the compound select could not reach
the UNION branches, so every page materialized the whole candidate set
before dropping the rows below the cursor. Each branch now carries the
cursor bound and its own LIMIT, and the wrapped UNION still merges and
deduplicates the bounded branch pages. Adds a >2,000-candidate
provider-only test and a paged plan assertion on the branch seeks.

Co-Authored-By: Epinephrine <luvs01@hanmail.net>

* perf(usage): page candidates from ordered indexes without re-sorting

The candidate pager wrapped every branch in a compound UNION and sorted the
merged range with an outer ORDER BY byteOffset. Each branch index was ordered
(generation, timestamp, byte_offset) or (generation, provider|model, ...), so
SQLite could not deliver byte order from it and materialized a temp B-tree over
the whole matching set on every page. A deep page on a 400k-row ledger cost
seconds against a sub-second streaming scan.

Each branch is now one keyset statement whose index is ordered
(generation, <equality prefix>, byte_offset, ...), and the pager merges the
branch heads in JS. Per branch it keeps a fetch cursor and the surplus rows the
last page did not need, so a candidate is fetched about once instead of being
re-read on every page.

Index changes behind schema version 5 (recreated, then rebuilt from the ledger):
- idx_usage_rows_window replaces idx_usage_rows_ts; the window is a residual
  predicate on a byte-order index rather than a sort-inducing timestamp seek.
- key, client, admission and unattributable indexes carry byte_offset before
  any other varying column.
- lone provider/model filters page from new offset-ordered attribution indexes;
  the unused idx_usage_attr_model is dropped.
- the Codex roster expands to one equality per label, since client IN (...)
  spans key values and forces a sort.

Tests: new tests/usage/usage-filter-paging.test.ts pins the no-temp-B-tree plan
for every branch shape, multi-branch cursor order and page bounds, the corrupt
projection rebuild, and the version 5 recreation. The transient-open retry tests
replace their real timers with a deterministic beforeRetryDelay hook; the
page-yield test observes the scheduled macrotask instead of racing it.

The 400k-row benchmark runs on macmini-cf over SSH, not the interactive machine.

Co-authored-by: Epinephrine <luvs01@hanmail.net>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

* fix(usage): stop a paged read at its snapshot's indexed offset

The indexed offset is where the next unindexed row starts. Each page runs in
its own transaction, so a concurrent refresh can index a row at exactly that
offset while a read is paging; with a <= bound the read returned it, the row
reader rejected it as content_changed, and the whole read retried until it
failed. Bound candidates with <, and start an append read's branch cursors at
afterOffset instead of adding a second lower bound SQLite seeks from on every
page.

Co-authored-by: Epinephrine <luvs01@hanmail.net>

---------

Co-authored-by: Epinephrine <luvs01@hanmail.net>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Co-authored-by: lidge-jun <243035832+lidge-jun@users.noreply.github.com>
luvs01 and others added 30 commits November 17, 2026 19:09
* fix(catalog): bound eager tool warning diagnostics

* docs(catalog): explain diagnostic budget in Spanish

* test: isolate excerpt masking complexity from SSE serialization
…idge-jun#8851)

Co-authored-by: luvs01 <luvs01@users.noreply.github.com>
…un#8834)

* test(link): gate post-proof timeout fixtures on data receipt

* test(link): observe socket settlement before timeout assertions
…ecord (lidge-jun#8634)

* fix(responses): keep a child's reasoning replay out of the parent's record

A child agent request carries its own thread-id plus x-codex-parent-thread-id, and reasoningReplayConversationIdFromResponsesRequest preferred the parent header. Both turns shared one replay scope, so after the child was served, commitReasoningReplayServingIdentity recorded its provider/adapter/model under the parent's key. The parent's next Claude turn then saw a route change, stripped its replayed thinking signatures and re-wrote the prompt prefix (lidge-jun#8632).

Resolve the request's own thread-id first and keep the parent header as the fallback, so callers that only carry the parent identity still share one record. Log conversation ids and pool affinity are unchanged.

* test(responses): pin the parent-child-parent serving cache and budget the cold child build

Review follow-up for lidge-jun#8634. The architecture contract now states that custom native v1 compaction uses the request-thread-first replay scope (own thread-id, parent header as fallback) and that the default-v1 previous_response_id retention scope is unchanged.

Adds the full serving-cache sequence: parent A (Claude) commits, child B (both headers) commits, parent A returns with no detected route change and an unchanged rejection policy, plus a parent-only fallback control and the unchanged request-log identity. An end-to-end combo/route variant asserts the parent's wire body keeps its Claude thinking signature after the child turn and that a real model switch still strips it.

The Cross-platform failure in tests/clients/link-ocx-child-e2e.test.ts was the fixture's cold Cargo build inside a 30 s case. The build now runs once in beforeAll with its own 240 s budget (cargo spawn 180 s) and both lifecycle cases still keep their real assertions and 30 s deadlines.

* docs(responses): qualify when a child's thread-id isolates its replay record

CodeRabbit review on lidge-jun#8634: the contract claimed child isolation unconditionally. Isolation applies only when the child supplies its own valid, distinct thread-id; a request carrying only x-codex-parent-thread-id still resolves to the parent's record and a successful custom compact response commits provenance there (src/server/responses/compact.ts).
…idge-jun#8805)

* fix(responses): fail closed when stream replay classification fails

* test(responses): cover coercible non-string replay selectors

* test(link): make post-proof data failure fixtures deterministic

Carry the validated fixture from lidge-jun#8834
(0fbc43b). Gate data delivery before
advancing the original deadlines and drive native socket processing
through checkpoint requests without relaxing timeout assertions.

* test(responses): isolate the excerpt masking performance budget

Carry the validated fixture from lidge-jun#8815, matching blob
982f0d2. Keep the same credential set,
3.9 MB input and 1,500 ms limit on the production excerpt masker; retain
the full SSE rewrite and verify its redaction matches that result. Add a
deterministic guard against a whole-text search per matching window.
…dows git config, test network guard) (lidge-jun#8870)

Lane R5 of the 2026-11-17 triage round: read zh-CN and ko-KR schtasks status words in the Windows checked stop (lidge-jun#8863), use /dev/null as the empty global git config on Windows (lidge-jun#8746), and refuse real network egress from the test process (lidge-jun#8848).

Co-authored-by: andrew05060414 <wsy1731095638@gmail.com>
lidge-jun#8872)

- endTask keeps a zh-CN or ko-KR access denial ("拒绝访问", "액세스가 거부")
  as denied before the idle-state read, so Part A still stops at a denial.
- The test network guard follows redirects one hop at a time and refuses a
  hop to a non-local host; real fetch followed them internally, past both
  hooks. The final response keeps redirected; manual and error modes are
  untouched.
- An explicit non-loopback proxy is refused even for an allowed target, for
  fetch and WebSocket.
- Fetch schemes other than http(s), ws(s), data:, blob: and file: (s3:) are
  refused.

Each case has a regression test that fails without its fix.
…-jun#8792)

Co-authored-by: zhuxt2015 <13765310+zhuxt2015@users.noreply.github.com>
…e-jun#8756)

* fix(kiro): keep the measured schema keywords on claude-opus-5.5

Claude Desktop (116 tools) and Claude Code (25 tools) declare twelve of the 29 stripped keywords. Five days of production traffic sent them unstripped to claude-opus-5.5: 10,565 HTTP 200, no ValidationException. The other seventeen keywords and every other model id keep the current policy.

* docs(kiro): state the tool budget of the production sample

---------

Co-authored-by: svc-aibot-363 <271381517+svc-aibot-363@users.noreply.github.com>
…-jun#8869)

Lane R2 of the 2026-11-17 triage round: with emptyCompletionRetry enabled, hosted-search Chat turns retry an empty completion once with the gathered results, with the retry given its own iteration (lidge-jun#8842).

Co-authored-by: Ada Urbaniec <ada.urbaniec@gmail.com>
…guard, Spanish adapters parity) (lidge-jun#8874)

* docs(es): add the Gemini tool-catalog guidance line to the Spanish adapters reference

lidge-jun#8792 added a reference/adapters.md bullet naming `429 RESOURCE_EXHAUSTED`; its PR CI predated the
Spanish localization identifier check, so dev now fails docs-spanish-localization. This adds the
matching Spanish bullet.

* fix(tests): let the network guard pass WebSocket over Unix sockets

lidge-jun#8872 refused every scheme outside http(s)/ws(s)/data/blob/file, which included ws+unix:.
Local and remote messaging reach their daemon through ws+unix:// WebSockets, so the guard
rewrote them to its refused socket and the messaging suites failed on dev with
daemon_unavailable and remote_control_unknown. ws+unix: and wss+unix: never leave the
machine, so they are allowed like data:, blob: and file:.
…hropic pool failover fixes (lidge-jun#8871)

Lane R1 of the 2026-11-17 triage round: hop on unoffered tool-call refusals (lidge-jun#8563), time a rotated attempt from its own send (lidge-jun#8749), resolve Devin context windows for combo input admission (lidge-jun#8575), and scale Anthropic pool failover within the funded send budget (lidge-jun#8857), with review follow-ups.

Co-authored-by: Omar <13569609+omar391@users.noreply.github.com>
Co-authored-by: Andrew (agent) <wsy1731095638@gmail.com>
Co-authored-by: Vadym O <bolein95@gmail.com>
…eveloper-role rejection (lidge-jun#8864) (lidge-jun#8867)

* docs(devlog): plan lidge-jun#8864 Aside/Prime developer role and DashScope rejection

* fix(clients): Aside and Prime send system instead of developer; Alibaba presets fold the role

Aside and Prime reuse Pi's models.json builder but never received the lidge-jun#5674
developer-role option, so their Pi runtime sent the system prompt as
"developer" and DashScope answered 400. Both export paths now write
compat.supportsDeveloperRole: false (session affinity stays Pi-only), and a
shape-discovered regression test covers every Pi-builder client.

alibaba-token-plan and alibaba-token-plan-intl now record
foldDeveloperRoleToSystem, so the native Chat passthrough folds developer
into system for any client, including saved configs and an overridden
DashScope base URL.

Refs lidge-jun#8864

* docs(es): Prime and Aside integrations state the developer-role compat flag

* fix(errors): name a developer-role rejection without quoting the provider

A provider that rejects the developer role (DashScope "developer is not one
of ...", Moonshot "role 'developer' is not allowed", serde "unknown variant
`developer`") used to reach the client as "The provider rejected the
request." with nothing in the request log, so the cause took a manual
bisect to find.

A fixed-phrase verdict on upstream 4xx evidence now selects a proxy-owned
actionable sentence on Chat and Messages, and provider bodies record the
existing closed pair messages.role/enum, which the dashboard already shows.
No upstream text reaches a client body, row, usage.jsonl or the console; a
body this proxy formatted keeps validation-only capture.

Refs lidge-jun#8864

* test(errors): cover the Messages boundary and keep each row's role pair separate

Final review follow-ups for lidge-jun#8864: a Messages request through startServer now
asserts the fixed developer-role sentence end to end, and each row gets a
fresh messages.role/enum array instead of a shared module constant that a
valid list would store by reference.

Refs lidge-jun#8864

* test(errors): keep nested objects visible in the boundary console capture

The boundary test asserts that no provider text reaches console output, but
String() turned an object argument into "[object Object]", so a regression
that logged { error: { message } } would have passed. Non-string arguments
are now printed with Bun.inspect at full depth. Addresses the CodeRabbit
review on lidge-jun#8867.

Refs lidge-jun#8864
…lidge-jun#8871 follow-up) (lidge-jun#8875)

* fix(combos): honor dotted devin override keys for hyphenated requests

Devin per-model overrides were matched against the request's own spellings
only, so a request for swe-1-7-medium never tried an override keyed swe-1.7
and fell through to the provider-wide contextWindow. After the exact, collapsed
and hyphenated candidates miss, record keys are now compared after folding dots
to hyphens, for both modelContextWindows and modelMaxInputTokens. Exact keys
still win.

Follow-up to the CodeRabbit thread on input-admission.ts in lidge-jun#8871.

* fix(combos): fold case in the dotted Devin override fallback

CodeRabbit on lidge-jun#8875: modelRecordValue matches record keys case-insensitively, but the dotted-key
fallback compared exact case, so an override keyed SWE-1.7 missed a swe-1-7-medium request.

* fix(combos): lower-case before collapsing Devin effort suffixes in the override fallback

CodeRabbit on lidge-jun#8875: collapseDevinModelUid matches effort tokens case-sensitively, so
SWE-1-7-MEDIUM kept its suffix and missed a swe-1.7 override.
…utcome (lidge-jun#8379) (lidge-jun#8876)

The racing-uninstall test asserted that every child that recorded its path saw that path removed,
before it checked the uninstall's status. Any first run that did not reach removal therefore failed
with "expected false, received true": a refusal for a reason other than a busy lock (on Windows, the
unlocked check can fail to open a manifest a registration is renaming over) or a partial run that
stopped in its snapshot. Neither loses a registration: the path stays listed in the untouched or
tombstoned manifest, and a rerun removes it.

settleRacingUninstall now checks exactly that, reruns such a run once the race is over, and puts the
uninstall result and child exit codes in every failure message. Four forced cases pin each outcome
in order; the refused and partial ones fail against the old assertions.

Closes lidge-jun#8379
…nclude, Command Code root schemas, JEV continuation effort (lidge-jun#8873)

Lane R3 of the 2026-11-17 triage round: rewrite the Claude Code sentences Cognition refuses for Devin (lidge-jun#8703), normalize the OpenCode Go output include (lidge-jun#8827), normalize root tool schemas for Command Code's Claude upstreams (lidge-jun#8591), and keep JEV's decided effort through tool-loop continuations (lidge-jun#8861, continuation fix only), with review follow-ups.

Co-authored-by: 袁威豪 <137183806+jhl880617@users.noreply.github.com>
Co-authored-by: Ingwannu <186453546+Ingwannu@users.noreply.github.com>
Co-authored-by: Mursyid <ciddhere@gmail.com>
Co-authored-by: tachyon6 <songjunjun62754@gmail.com>
…xture (lidge-jun#8879)

The queue fixture CLIs append one JSON line per frame from several processes while the
test and the fixtures read the same file, so a read could land mid-append and fail with
"JSON Parse error: Unterminated string" (seen on lidge-jun#8878 run 44104587671, test 1/5, "wait
requests share the four-operation bound"). Parse only newline-terminated lines.
…ction (lidge-jun#8878)

* docs(devlog): plan key-level Codex journal restore that never reverts later config edits

* test(codex): reproduce restore losing concurrent writes and adopting stale injected values

Adds a behavior-neutral seam (runBeforeCodexConfigPublishForTests) at the three
restore publications of config.toml and child-process tests under temporary
CODEX_HOME/OPENCODEX_HOME/HOME. On this commit 11 of 18 fail for the recorded
reasons: start/client journal recovery leaves routing behind, a stale journal is
retained, an in-place write landing during restore is overwritten, a rename
during restore crashes compensation, an app rewrite to an earlier injection's
URL is adopted as user-owned, and a dropped model_catalog_json is not drift.
The 7 passing cases guard stop/disconnect/untouched behavior.

* fix(codex): restore never reverts config.toml changes made after injection

Every OpenCodex publication of config.toml now goes through
publishCodexConfigIfUnchanged, which refuses retryably and writes nothing when the
file no longer holds the bytes the decision was computed from. The lock's inode
witness already refused an atomic rename (how Codex writes) but missed an
in-place writer, which restore renamed over while reporting success.

Failure compensation reverts config.toml only while it still holds this lock
section's own publication, and skips surfaces already holding their preimage, so
a foreign write is never rolled back and a refused restore names only what it
could not put back.

Start and mismatched-client journal recovery use the three-way owned-field
restore that disconnect already used. They used to refuse as "incomplete" after
any later edit and leave a dead loopback URL in config.toml.

The journal records every root routing URL injections under it wrote
(injectedRootHistory, optional, version stays 1). A Codex app write-back of an
earlier injection's URL is stripped and re-pointed instead of being snapshotted
as user-owned. Drift detection also covers a dropped model_catalog_json and such
write-backs, and the first heal in a process warns once.

structure/codex-home.md and structure/overview.md (INV-RESTORE-02) state the
invariant; tests/codex-integration/codex-restore-later-edits.test.ts enforces it.

* fix(codex): keep the drift heal notice on the info channel; repair the write-back drift test literal

* docs(codex): state compensation ownership for held and unheld callers

* fix(codex): report an unreadable preimage surface as unrestored instead of throwing

* docs(devlog): record verification and PR for the journal restore unit

* test(layout): register the journal restore tests in their own shard to keep layout.json under the ratchet

* test(codex): reproduce the CodeRabbit round on lidge-jun#8878 before fixing it

Adds setAfterCompensationDecisionForTests, a behavior-neutral seam between
compensation's ownership decision and its publish/remove, and seven cases that
fail on this commit:
- start and mismatched-client recovery crash on ConfigWriteContentChanged when
  config.toml changes while they restore (journal kept, start aborted);
- compensation overwrites (replace) or deletes (remove) a write that lands after
  it decided the file was OpenCodex's publication;
- reinjection over a journal without injectedRootHistory loses the URL it replaces;
- a multiline model_catalog_json assignment, and an unparseable config, are
  reported as a missing catalog key.

* fix(codex): fold the CodeRabbit round on lidge-jun#8878

- Compensation publishes over, or removes, config.toml only while the file still
  holds the exact bytes its ownership check hashed; it no longer re-reads before
  publishing, and removing a file this section created checks content first.
- Start and mismatched-client journal recovery treat ConfigWriteContentChanged as
  "defer to the next start" (journal kept, start continues); the failed-connect
  rollback records it as a partial rollback.
- markJournalInjectedState records the outgoing latest URLs in injectedRootHistory
  before overwriting them, so journals written before the history keep them.
- Drift decides key presence from the parsed document (multiline assignments are
  present; an unparseable file reports nothing missing).
- The publication docstring, INV-RESTORE-02 and structure/codex-home.md state the
  remaining compare-then-rename window, and the native-child digest limit is
  documented where it is recorded.
…lidge-jun#8893)

* feat(models): add GPT-6.2 Sol and Astra and upgrade subagent defaults

* test: sync GPT-6.2 catalog and CI expectations

* test(claude): verify the current generated agent roster

* feat(providers): seed GPT-6.2 Sol and Astra preemptively on GPT-6.1 Sol providers

Mirror each provider's GPT-6.1 Sol row on Kiro, CodeBuddy, Devin (dashed uids
and derived pricing), TokenLab, Modelflare and GitHub Copilot, and opt both
OpenRouter routes into Fast, following the Sonnet 5.5 and GPT-6.1 Terra
preemptive-seed precedent. No provider default changes.

Generated with [Devin](https://devin.ai)

Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>

* docs(providers): separate preemptive Copilot Responses defaults from observed rejections

GPT-6.2 Sol/Astra and GPT-6 Sol/Luna default to Responses on Copilot without an
observed /chat/completions rejection, so list them as preemptive built-in
defaults instead of in the observed-rejection list, in English and every
translation that repeats it.

Generated with [Devin](https://devin.ai)

Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>

* fix(models): give GPT-6.2 Sol and Astra their 128K output ceiling

The configured-native template moved to GPT-6.2 Sol, but neither GPT-6.2 model
had a generated metadata row, so native output admission and configured natives
resolved an unknown output ceiling. Add the verified openai, openai-codex and
openrouter rows (128K output, API pricing, native/API windows) to the metadata
snapshot and regenerate.

Generated with [Devin](https://devin.ai)

Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>

---------

Co-authored-by: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
lidge-jun#8895)

The prepublish audit blocks the 2.100.5 release on picomatch <= 4.0.7 (ReDoS),
which has no patched release. picomatch reaches the GUI only through vite's build
tooling; `bun audit --production` in gui/ is clean, so neither gui/dist nor the
published package ships it. Ignore exactly that advisory in the GUI audit only,
keep the root audit unchanged, and guard the exception's scope with a test.

Generated with [Devin](https://devin.ai)

Co-authored-by: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
…8592)

* test(gui): await the target-switch layout commit before assertions

* test(gui): wait for the new target read to reach in-flight state

The layout-commit wait still could not prove the new target's passive-effect read had started: the hook kicks it from a passive effect plus a queued microtask, so a trailing zero-delay timer could observe only two reads on a loaded runner (the still-open review finding).

Wait until the fixture's own in-flight state is painted instead: refreshing turns true only after the read's fetch was issued and React committed the new counter, so the timer becomes a yield between observations rather than an effect-ordering assumption. Releasing the old read stays in the layout phase, preserving the regression window.

Verified: bun test tests/codex-account-pool-target-switch.test.tsx -> 13 pass / 0 fail / 86 assertions (4 runs); bun run build (tsc -b + vite + vite classic) passed; git diff --check passed. Baseline and change both emit the same flaky not-wrapped-in-act noise, so none is introduced.

---------

Co-authored-by: luvs01 <luvs01@users.noreply.github.com>
Co-authored-by: JUN <jun@lidgeai.com>
…hropic pool (lidge-jun#8886)

* fix(anthropic): retain native features for embedded Claude clients

* fix(anthropic): preserve native identity for Claude project children only

* docs: describe Claude cloud project child identity

---------

Co-authored-by: JUN <jun@lidgeai.com>
…ws (lidge-jun#8812)

* fix(models): bound alias collision construction and reuse listing views

* test: isolate excerpt masking complexity from SSE serialization

---------

Co-authored-by: JUN <jun@lidgeai.com>
…tion (lidge-jun#8862)

* fix(combo): do not record synthetic replay refusal as quota failure and allow compaction hops on reset

* fix(combo): require store === false for compaction replay-safety shortcut

* fix(responses): separate combo quota attribution from replay policy

Co-authored-by: Omar <13569609+omar391@users.noreply.github.com>

* fix(usage): classify replay refusals as transport failures

Give exact transport codes precedence over generic Chat error types.
Verify the final Chat log and exercise the two-target combo over HTTP.

Co-authored-by: Omar <13569609+omar391@users.noreply.github.com>

---------

Co-authored-by: Omar <13569609+omar391@users.noreply.github.com>
Co-authored-by: JUN <jun@lidgeai.com>
* fix(cursor): bound tool envelope echo filtering work

* fix(echo): settle line endings at the pending content limit

* Handle Qoder echo budget failures at event boundaries

---------

Co-authored-by: JUN <jun@lidgeai.com>
Reauthor hosted pinning, canonical Messages fidelity, pool policy, physical accounting, analytics and provider/state contracts against upstream 6bae884. Preserve generic fixes and drop carried historical port debt. Add AR parity and trust-boundary regressions plus a required local CI gate; document paired upstream and hosted validation and release boundaries.

Origin: hosted fork v2.65.0-vadymhimself.27; per-feature attribution and validation are recorded in FORK-DELTA.md and FORK-VERIFICATION.md. Includes upstream recovery-attempt PR lidge-jun#8882 and sidecar unused-hop refund lidge-jun#8899.

Co-authored-by: Gauge <gauge@agentradar-house.staging.agentradar.com>

Co-Authored-By: Claude Code <noreply@anthropic.com>
Reject malformed strict SSE tails before relay. Preserve a read-only OAuth binding for token counts and revalidate after pacing; queued pause/removal/cancellation cannot dispatch stale authority and replacements use only the new credential. Bound Retry-After through the existing formatter. Add controlled regressions and supersede held build .1 with 2.100.0-vadymhimself.2.

Co-Authored-By: Claude Code <noreply@anthropic.com>
Adopts the upstream 2.100 rebuild (6a14467, PR #46) as the fork's main
line while keeping the old main history reachable. Main's own runtime fixes
are carried or upstream-equivalent; see FORK-DELTA.md.

Co-Authored-By: Claude Code <noreply@anthropic.com>

This branch has not been deployed

No deployments
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.