Skip to content

feat(chatgpt-unblock): PAC-fallback mode so traffic survives opencodex stopping - #5947

Closed
lcxhh521 wants to merge 59 commits into
lidge-jun:devfrom
lcxhh521:feat/chatgpt-desktop-pac-fallback
Closed

lcxhh521 wants to merge 59 commits into
lidge-jun:devfrom
lcxhh521:feat/chatgpt-desktop-pac-fallback

Conversation

@lcxhh521

@lcxhh521 lcxhh521 commented Sep 26, 2026 •

Copy link
Copy Markdown
Contributor

Maintainer review follow-up — 2026-10-15

The maintainer requested changes in review5530867941. The PAC preflight-before-quit bug is fixed in cc989575; the launch/watch path prepares a readable, nonempty, successfully encoded PAC snapshot before quitting, and reuses that snapshot when launching. Missing/unreadable/directory/empty PAC inputs and encoding failures leave the running app untouched. Broader PAC routing, stop/crash recovery and certificate-trust/interception design acceptance remain unresolved. Passing focused checks and the earlier CodeRabbit review do not settle these requirements. This PR is not ready for merge.

Validation of exact tree 06d4b89203092ed5d97686d10daeb540289787fc: regression-first checks reproduced the premature quit and invalid-PAC launches; the final focused wrapper passed 227 tests, 6 skipped, 0 failed. Four skips are macOS-only; two unreadable-file tests skip when run as root and were separately executed successfully as a non-root user (the 14 targeted cases all passed without skips). Typecheck, structure checks, privacy scan, whitespace checks and the docs build passed (577 pages, 81,320 links). Independent review found no blocking issue in this narrow patch. This does not verify real-app stop/crash recovery or resolve the broader design hold.

Full-review follow-up — 2026-10-15

Head 61575aa0 addresses all six findings from the full CodeRabbit review of 45ed3fdc. Restore now queries only launchd's loaded state, retaining the loaded-watcher refusal while avoiding installed-file reads. Diagnostics warn non-fatally if PAC fallback is configured without send unblock and leave the saved flags unchanged. The public guide and structure contracts now state this dependency and the 65535/65534 entry-port boundary; caller inventory and moved test references are corrected.

Exact published tree 52376dcd52d0eaae0c80c2d1c9eb80c9b100339a matches the locally tested tree. Regression-first restore and config checks failed before the fixes and passed after. Official focused wrapper: 152 passed, 4 macOS-only skips, 0 failed across six files. Typecheck, structure SSOT, privacy and whitespace checks passed. Public docs build passed (577 pages and 81,320 internal links; existing chunk/i18n/404 warnings remain). Independent review found no blocking issue. This scoped run is the local full-suite exception; actual macOS runtime and the broader changed-suite limitations below still apply.

All 39 inline threads are resolved. CodeRabbit completed an incremental confirmation on this exact head on 2026-10-15, processing all 13 changed files and reporting no actionable comments; its status is success. Automatic reviews remain paused. Its three functional Actions runs are already terminal action_required, requiring upstream approval before execution.

Review fixes — 2026-10-15

Head 45ed3fdc preserves the existing PAC fallback feature and addresses the remaining valid review findings: watcher plist freshness now matches both installed watch paths; HTTP CONNECT handles password-only credentials and rejects malformed encoding before writing; hand-built SOCKS username-only credentials match the existing empty-password URL behavior; the missing structure mapping and duplicate test/import are corrected. All 33 inline review threads are resolved (six fixes in this increment; the seventh remaining documentation item was already fixed by the preceding conflict merge). Independent review found no blocking issue in this increment.

Validation of exact tree 066d7dcf84946d69736561cae972864d9ceb1203: 203 passed, 0 failed across nine focused files (83 transport/credential checks plus 120 watcher/bash/structure checks). Typecheck, structure SSOT, privacy and whitespace checks passed. Regression-first checks demonstrated the watcher and credential failures before the fixes. These are synthetic/loopback tests, not new real macOS runtime evidence.

The automatic broader changed-suite was blocked by the cloud execution review before its test selection completed; it was not retried or counted as passed. Read-only fixture inspection found the examined provider.invalid names used as request targets with loopback dials, but cannot identify the blocked suite's full selection. The explicit scoped checks above are the documented local validation exception; full-suite and actual macOS PAC recovery remain unverified for this head. The preceding conflict-merge validation below applies to its own tree, not a full rerun of this increment.

All three functional GitHub Actions runs for this head are terminal action_required pending upstream approval. CodeRabbit completed a full review of head45ed3fdc on 2026-10-15; its six findings were subsequently addressed in61575aa0 as described above. Ready for review is not CI-passed or merge-approved.

Conflict refresh — 2026-10-15

Merged current dev 7640a942 into the original branch, preserving the PAC fallback implementation and existing feature scope. This is not a withdrawal or a replacement design. The current conflict set had grown from the earlier two-file maintainer note to eight files.

Resolution preserves upstream discovery/exit failure handling and SOCKS transport fixes, retains the PAC watcher/status/restore gates and both sides' tests, reconciles test layout entries, and regenerates the structure index. The src/chatgpt/desktop-unblock/ implementation, schema and listener lifecycle are unchanged from the previous PR head.

Validation on the exact published tree b594abe7509c507af2930e1d0d42a528f22c46cb: typecheck, structure check, privacy scan and generated skill surface check passed. The official focused test wrapper passed 1715 tests, 10 skipped, 0 failed across 108 files, with 14229 assertions. The skips include five macOS PAC/intercept runtime cases and five Windows cases; this cloud run does not establish real macOS PAC stop/recovery behavior. Full repository validation was not run; this focused run is the documented local exception.

The existing inline review findings were subsequently addressed as documented above. The maintainer's conflict comment did not reject this PR or change its review state.

Summary

Current state (5d6a9685a, 2026-10-05). The app-server shim reached dev through #6361, and the send-unblock intercept is #6365, with the ready-marker watcher stacked on it in #6381. This branch now builds on #6381's head (336044304, dev merged, 0 behind). The only thing it adds is PAC fallback, re-applied onto that structure:

  • entry-proxy.ts and pac.ts are unchanged in substance. runtime.ts binds the CONNECT entry, rewrites chatgpt-unblock.pac at every start before the chatgpt-unblock.ready marker, and releases both listeners on failure.
  • The launch watcher gains the PAC switch, the entry probe (curl exit 0 or 56) and PAC-aware restore, on top of fix(codex): honor requested restart after config changes #6381's script with its readiness wake and five-minute guard. ocx chatgpt launch, install-watcher and status pass and report the entry port.
  • chatgptDesktop.pacFallback joins the zod-only schema in src/config/schema/chatgpt-desktop.ts.
  • Dropped from this branch: its own app-server shim copy (dev's hardened src/chatgpt/app-server-shim/ is used) and the pre-feat(chatgpt): local-CA send-unblock intercept, stacked on #6361 (split from #5947) #6365 intercept code. The bullets below describe the original branch; read them with that in mind.

Opt-in PAC-fallback mode for the ChatGPT desktop send-unblock intercept, so the app keeps working when opencodex stops.

Merge order: this PR now carries the send-unblock work directly. #5733 was closed unmerged as superseded by this branch (all four of its commits are patch-equivalent here), and the branch has been rebased onto current dev, so every commit in this PR is only this feature's work:

  • 50e6c804-equivalent: the send-unblock intercept (opt-in), its launch watcher, and the relay fixes CodeRabbit found while reviewing this PR (a missing error listener on the WebSocket tunnel between the handshake and attach(), and duplicate test-layout keys);
  • the remaining commits are the PAC work and the RFC 1929 authenticated-SOCKS fix for the WebSocket dial (review by Ingwannu).

What the mode does (chatgptDesktop.pacFallback, default off; only takes effect together with unblockSend):

  • The app is launched with an inline --proxy-pac-url=data:application/x-ns-proxy-autoconfig;base64,... switch instead of the --host-resolver-rules switch (a file:// PAC is ignored by the app, see Verification). The PAC is rewritten at every opencodex start; the watcher script rebuilds the switch from that file at run time.
  • chatgpt.com goes to a new loopback CONNECT entry listener (listener port + 1), which accepts only CONNECT chatgpt.com:443 and splices the bytes onto the existing TLS listener. The splice is backpressure-safe (Bun sockets are unbuffered, so unwritten bytes are queued and the producer paused until drain), and a client that never finishes its head is closed at the deadline.
  • Every other host — and chatgpt.com while opencodex is stopped — follows the system route captured at start, never a hard-coded DIRECT: the scutil --proxy proxies (HTTPS, HTTP, SOCKS5, then DIRECT) in system-proxy mode; the system PAC script itself, embedded in the generated file, when a PAC is configured (the PAC mode of VPN clients such as ShadowsocksX-NG); DIRECT in TUN mode or without a proxy. A system PAC that cannot be read, or that is too large to travel in one launch argument (the encoded switch is capped at 512 KiB, half of macOS ARG_MAX), degrades to the scutil proxies, then DIRECT, with a startup warning.
  • When opencodex stops, the entry listener dies with it; Chromium's CONNECT is refused and it falls through to the next entry of the PAC answer, so the app keeps working without a restart. Only the send unblock pauses until opencodex is back.
  • ocx chatgpt launch, restore, install-watcher and status follow the configured mode. restore undoes either switch, so it also works after pacFallback was toggled. The watcher refuses to route the app while the entry listener is down.
  • The English guide documents the mode. dev carries no localized guide for this page, so the earlier localized copies are not part of this branch.
  • App-server gate ([Bug] Codex App (macOS): send button disabled at quota exhaustion - the #5947 intercept never sees the gate reads (they come from the bundled app-server) #6196), opt-in chatgptDesktop.appServerShim. On some builds the composer's send gate follows what the bundled codex app-server reports to the app over stdio JSON-RPC, and the app-server fetches it with its own HTTP client, so no Chromium switch reaches it. The desktop app picks its server binary from CODEX_CLI_PATH; with the flag, ocx chatgpt launch / the watcher start it with open --env CODEX_CLI_PATH=<launcher>. The launcher execs the real binary, so the server stays the process the app started (same pid, parent and code-signing identity; the app checks that identity before it lets a peer onto its app-tools pipe), and redirects only the server's stdout, through process substitution, into a line filter: a line that mentions no rate-limit field is written back as the exact bytes it arrived in, and a plain-quota rateLimitReachedType / ordinaryUsageAllowed is opened (also when a quota window reads 100%) while workspace, credit and spend-control reasons are kept; with one of them present the web-side rate_limit flags also stay as sent. Only the account/rateLimits/updated notification and a result whose top level holds the rate-limit snapshot are rewritten; tool results and other messages that nest similar fields pass through. Stdin, stderr and signals go straight between the app and the server. No environment variable, address or config key changes, so the server's child processes and other Codex clients are unaffected, and the launcher first runs the filter once on empty input and runs the real binary with its stdout untouched if that fails, so a filter that is missing or does not load never leaves the server writing into a dead pipe. The resolver and PAC modes are unchanged; the flag is off by default. The exported app-server protocol schema lists account/rateLimits/read and account/rateLimits/updated as the messages that carry this state; the usage rewrite also drops a plain-quota rate_limit_reached_type on the web-page surface.
  • Watcher robustness (reported by TooSpace while reproducing). Every lookup of the app uses pgrep -a -x ChatGPT: macOS pgrep skips its own ancestors, so ocx chatgpt launch, restore and status run from a terminal inside the desktop app could not see the app. install-watcher now runs bash -n on the generated script and refuses, cleaning up, to load one that does not parse, instead of reporting success for a watcher that would die on every launch.
  • Review follow-ups (CodeRabbit, on this PR and on fix(codex): honor requested restart after config changes #6381, which carries this diff): the launcher probe above; the PAC size cap; SOCKS5 credentials need both a username and a password (RFC 1929 lengths are 1-255, so user@ or :pass@ is refused instead of sending a zero length); the translated guides now describe the app-server shim and its troubleshooting entry; ocx chatgpt help lists every subcommand.
  • Write boundary. A malformed chatgptDesktop block (for example { unblockSend: true, port: 65536 }) is now rejected by validateConfigCandidate with schema_invalid: chatgptDesktop.port instead of being silently dropped; a hand-edited file still degrades to off on load.

Verification

Runtime-owned PAC integration — 2026-10-16

Exact head 36a6319ad4bb87402ee71e1b9cb4a49fcecb0dfb, tree 9a1898c6e97e4d3dcba2cd3fe594a5e18f6b897d, includes current dev 2a3fd009. This increment implements the approved runtime-owned policy directly on the current branch. The earlier unpublished prototype was unavailable in the refreshed workspace, so this is a fresh reconstruction with fresh review and tests, not reuse of its prior test evidence.

Both launch modes capture one system PAC/static-chain snapshot for HTTP and WebSocket upstreams. PAC runs in an isolated QuickJS WASM worker (quickjs-wasi 3.6.2), with Acorn 8.17.0 syntax validation; both production dependencies are exactly pinned with lock integrity. The whole result chain is validated before dialing. Explicit DIRECT is preserved; invalid/unsupported rules, lookup/evaluation errors and exhausted chains never select an implicit DIRECT or environment proxy. Without a system policy, existing OpenCodex outbound selection remains unchanged. HTTP reuses the exact selected, TLS-verified socket; only pre-connect TCP/DNS failures may advance a listed chain. CONNECT/SOCKS rejection, TLS errors and application failures never replay a POST.

This is deliberately a restricted PAC subset, not full browser compatibility: date/time-range and myIpAddress helpers, IPv6 extensions, extended glob patterns, dynamic/async code and catch/finally recovery are unsupported. The parser distinguishes actual syntax from comments/strings. Catch/finally is rejected because the evaluator cannot reliably report a guest-caught allocation failure. Unsupported operations, aborts and resource limits close the policy until restart. Native getaddrinfo cannot be physically cancelled; the owner ignores late completion and cannot issue another lookup after cancellation. The public guide lists limits and compatibility tradeoffs.

  • Official isolated wrapper, bun scripts/test.ts --parallel=1: 1,084 passed, 11 skipped, 0 failed across 66 files (7,528 assertions), covering all desktop/shim suites, new PAC integration tests, relevant CLI contracts, worker/build packaging and structure/layout/import/core-boundary guards. Skips: 5 macOS-only and 6 Windows-only. After final fixture/resource-test strengthening, the changed evaluator file passed 25/25 (2,203 assertions), and compiled packaging passed 1/1 (9 outer assertions). Additional shared opaque-tunnel tests passed 61/61. All unchanged production code in the aggregate is identical to this head.
  • Focused transport matrix plus existing WS regressions: 57/57, including proxy/origin TLS rejection, no implicit fallback, pre-connect failover, single-socket POST/no replay, streaming/decompression, cancellation and absolute deadlines. Runtime ownership fixtures: 13/13, both launch modes, startup failure and shutdown cleanup.
  • Compiled helper runs from an empty directory after deleting its build source, with embedded worker and WASM, no source/node_modules lookup, strict-rule rejection and natural exit. Missing-worker and missing-WASM ablations fail as intended. Independent security/correctness review found and we fixed helper-return handle leakage; removing the scope fix fails the strengthened endurance test after 1,075 requests, while the fixed policy completes 2,000 requests under a 1 MiB guest heap. No remaining code blocker was reported for the documented subset.
  • Typecheck, structure SSOT, privacy, skill-surface and changed-file whitespace checks pass. Docs build: 577 pages, 81,564 internal links. bun audit --audit-level high: no vulnerabilities reported across 146 packages; this is registry-advisory evidence, not proof of absence. Hygiene checker tests pass 31/31. Actual complete-PR patch hygiene has no empty-catch or other code-hygiene finding.
  • The full repository/automatic changed suite was not rerun; this scoped aggregate and targeted final reruns are the documented local exception. The preceding dev-refresh's seven broad-test failures and clean-dev reproductions remain recorded below, not counted as new passing evidence.
  • Exact-head PR hygiene is terminal failure: unsponsored_surface, solely for package.json/bun.lock. A maintainer must review the dependency surface and apply maintainer-sponsored; no sponsorship or exception label was self-applied. Cross-platform, React Doctor, Catalog and Service lifecycle are terminal action_required, awaiting approval rather than passing.
  • Still held: runtime policy protects only traffic reaching the active listener. Chromium's embedded PAC after exit retains non-mandatory behavior; real-app cached-PAC stop/crash recovery and certificate/interception acceptance remain unverified. Isolated macOS tests are pending separately and are not claimed here. This remains draft and does not settle the maintainer's changes-requested design review.

Prior watcher safety and dev-refresh verification — 2026-10-16

Exact head d596984145d08affca423e7c2d8aa2346b2788ca, tree cbdae2765e7fe1ed61c88722c3aba4fb9b0493d7, matches the locally tested tree. The branch includes current dev at 2a3fd00985c4f825af9ac41547ee7ac42fce4570. The latest refresh resolved only CLI help and its generated lifecycle reference: dev's new codex-provider command and all existing watcher/intercept commands are retained. Configuration validation, migration, startup ownership guards and both test-map sets preserve both sides. Desktop-unblock production code and its lifecycle are unchanged from the preceding f0026eb9 head; all prior safety fixes remain. No runtime PAC evaluator or new dependency was added.

The bounded proxy-policy follow-up rejects unreadable/malformed system proxy configuration and unreadable/oversized configured PACs, releases both listeners on startup failure, validates canonical proxy addresses and generated JavaScript strings, and removes synthetic DIRECT from configured static proxy chains in both PAC and resolver-rule launch modes. No-proxy/TUN direct routing and native restore remain available. Existing system PAC support and its explicit DIRECT decisions are preserved; embedded PAC execution errors/invalid results can still fall back directly under Chromium's non-mandatory semantics, so complete PAC fail-closed behavior is not established.

The preceding increment gates identity, HTTP and WebSocket admission until PAC/startup completes and closes admission before shutdown. CONNECT teardown terminates owned outgoing sockets and late connection completions. CA status structurally inspects exact-fingerprint user-domain SSL settings and login-keychain presence, rejecting deny/restricted/malformed settings as positive trust evidence; it checks current self-signed CA validity and safely quotes the printed manual trust command. This status is configuration evidence, not proof of actual app leaf acceptance.

The preceding loader hardening adds an absolute 3-second HTTP(S) PAC-download deadline, closes rejected/oversized response sockets, and reads local PACs through one regular-file descriptor with a 4 MiB + 1-byte cap. Growth or pathname replacement cannot trigger an unbounded read; short reads, read failures and FIFO rejection are covered. Existing system PAC routing behavior is unchanged.

The independent SSE fix (6b9f39b197d20ea251b86ee0e9b8839b4b8a0b86) bounds rewrite buffering at 8 Mi UTF-16 code units per line. Oversized lines stream unchanged until their newline, then normal rewriting resumes; decoder state is flushed at end-of-stream. Regression coverage includes early output before a newline, an oversized unterminated line and subsequent-line rewriting. No runtime PAC evaluator or new dependency is included in this increment.

The preceding test-only increment replaces all four duplicated port-pair allocators, including embedded crash/startup scripts, with one bounded helper. It retries only bind-time EADDRINUSE, reserves the origin while probing its neighbor, and always attempts both cleanups; unexpected bind or cleanup failures propagate. Six deterministic regressions cover collision retries, port-range overflow, both unexpected-bind positions, exhaustion and cleanup errors. No suppression or empty-catch workaround was added.

  • Current merged-tree validation: official bun scripts/test.ts --parallel=1 aggregate passed 1,681 tests, 5 macOS-only skips, 7 failures across 93 files (13,790 assertions). Coverage includes all desktop-unblock and config suites, desktop/shim command contracts, CLI help/capabilities/registry/usage comparison, skill-generation/structure/layout/import-cycle guards, shared tunnels, core Lab isolation and Codex-disabled startup. All seven failures reproduce on clean dev 2a3fd009: three CLI/help fixtures (two encounter unavailable systemd) and four settings-injection cases blocked by this cloud's non-root-owned temporary-directory guard. Baseline commands used the same wrapper on tests/cli/cli-help.test.ts --test-name-pattern 'status prints diagnostics|recover-history requires exact confirmation|invalid service and codex-shim' (3 failures) and tests/config/settings-stream-mode.test.ts --test-name-pattern 'rewrites the live Codex config|fresh Codex home|paginated Codex home' (4 failures). No environment permissions or safety checks were changed. Typecheck, structure SSOT, privacy, skill-surface and complete-PR hygiene passed; hygiene checker tests passed 31/31. Docs build passed: 577 pages, 81,540 internal links. Independent review found no merge-induced defect.
  • Prior hygiene head f0026eb9ff1ff4ab5c0b1887e873d84a426b0615, the official wrapper bun scripts/test.ts --parallel=1 tests/clients/desktop-unblock-*.test.ts tests/ci-workflows/file-size-ratchet.test.ts tests/ci-workflows/structure-ssot.test.ts tests/test-layout.test.ts tests/test-layout-tooling.test.ts passed 401 tests, 5 macOS-only skips, 0 failures across 19 files (1,701 assertions). node --test .github/scripts/pr-hygiene.test.cjs passed 31 tests. The actual hygiene checker reproduced the two reported paths before the fix and returns no failures on the complete current PR diff. Typecheck, structure SSOT, privacy, skill-surface and whitespace checks passed. Independent review found no blocker. Hosted PR hygiene passed on that head; the readiness gate now reports only the intentionally open checklist, rather than empty_catch.
  • Prior merged head f4edc36a45d99199efa75494ca9fac1029f43407, official focused wrapper bun scripts/test.ts --parallel=1: 1,257 passed, 5 macOS-only skipped, 4 failed across 67 files, with 6,405 assertions. Coverage includes all desktop-unblock suites, shared opaque/desktop/SOCKS tunnels, configuration suites, desktop command contracts, shim launcher/auto-restore, core boundary and structure/layout guards. All four failures are settings-injection cases refusing this cloud environment's non-root-owned /tmp; the same four cases fail with the same CodexUserIdentityRefusal on clean dev c03eafd46. No system permissions or safety checks were changed to bypass this environment limitation. Before the dev merge, the SSE increment separately passed 492/5 skipped/0 failed across 23 files. Real loopback fixtures cover half-open and late-connect teardown, authenticated WebSocket shutdown, child-process SIGKILL port release, stale readiness rejection and same-port restart. These are cloud fixtures, not macOS app recovery evidence.
  • Typecheck, structure SSOT, privacy scan, generated skill-surface check and whitespace checks passed. The preceding merged-head docs build passed with writable temporary HOME and telemetry disabled: 577 pages, 81,462 internal links. Independent review found no blocker in the SSE patch or mechanical integration. The earlier Bun 1.4.3 compiled-helper smoke verified standalone preflight argv; neither a full CLI compile nor real macOS execution is claimed.
  • Focused-validation exception: the full repository/automatic changed suite was not rerun or claimed passed; the scoped aggregate and clean-dev comparison above are the local evidence. Exact-head Cross-platform CI, React Doctor and Catalog async contracts report completed/action_required, each with zero jobs, awaiting maintainer approval. This is a platform gate, not passing CI.
  • Still held: complete embedded-PAC fail-closed behavior, certificate-trust/interception design acceptance, and real macOS stop/crash recovery. No system PAC support was disabled, no OS trust/network settings were changed, and no user computer was used. This follow-up does not mark the PR ready or settle the maintainer's changes-requested review.

Earlier verification (historical heads)

Head 41008f850 merges dev fe09557cb. The only conflicts were in the CLI capability registry, which dev split into per-area files: the ocx chatgpt entry moved to src/cli/capabilities-base.ts with this branch's text, help.ts keeps dev's lab line and this branch's chatgpt line, and bun run skill:surface regenerated the ocx skill reference (skill:surface:check passes). On this head, bun run typecheck, structure:check and privacy:scan pass. tests/clients/desktop-*, tests/config/, the CLI chatgpt/capabilities/help/registry tests, skill-ocx, test-layout-tooling, socks5-handshake and core-lab-boundary give 1383 pass, 5 fail across 88 files. The 5 failures are in cli-help-navigation, cli-help-recovery and cli-help-paths, and the same 5 fail on unmodified dev fe09557cb on this machine.

  • 5d6a9685a: bun run typecheck, bun run structure:check and bun run privacy:scan pass. tests/clients/desktop-* (with the new desktop-unblock-pac, -pac-runtime and -entry-proxy files), the ChatGPT and config suites, socks5-handshake, core-lab-boundary and test-layout-tooling (72 files): 1109 pass, 2 skip, 0 fail. The older bullets below refer to the pre-refresh heads.
  • bun run typecheck, bun run privacy:scan, bun run structure:check, bun run skill:surface:check — pass (re-run on the rebased branch).
  • bun scripts/test-layout/verify.ts --domain chatgpt-unblock — pass (182 tests).
  • bun test ./tests/chatgpt-unblock/ ./tests/lib/ ./tests/lab/core-lab-boundary.test.ts ./tests/test-layout.test.ts ./tests/test-layout-tooling.test.ts ./tests/ci-workflows/ — 2405 pass, 7 skip, 0 fail.
  • bun test ./tests/cli/ — 1402 pass, 4 skip; sibling-home-client-sync.test.ts reads the live sibling homes on this machine and fails the same two cases on clean dev, so it is left to CI. One cli-headless-parity.test.ts case failed once in the directory run; the file passes 3/3 on its own, on this head and on clean dev.
  • All of the above re-run on this exact head (248f966ee), on current dev (a6114b62e, 0 behind). The rebase replayed the early chatgpt_base_url commits over upstream's new structured-line parsing in src/codex/; the commit that removes that route now restores those files to the current upstream versions, so src/codex/ is identical to dev.
  • Full suite: left to CI.
  • RFC 1929: bun test ./tests/lib/socks5-handshake.test.ts (13 byte-level tests) and the credentialed-SOCKS relay end-to-end tests in tests/chatgpt-unblock/unblock-ws-relay.test.ts.
  • Regression tests were checked against the old code: the 8 MiB slow-reader tunnel test fails on the old splice (which passes at 64 KiB), the head-deadline test fails without a timeout handler, the listener-release test fails without the cleanup, and the two WebSocket late-error tests throw without the listener.
  • PAC generation was checked against scutil --proxy output on a Mac with a system proxy. Generated PACs are evaluated in a VM, including three ways a system PAC can declare FindProxyForURL.
  • App-server shim, checked against the real bundled codex app-server (26.924): with an exhausted-account usage payload it reports rateLimitReachedType: "rate_limit_reached" directly and null through the shim; workspace_owner_usage_limit_reached is left as sent; on the real account (not exhausted) the RPC output is identical with and without the shim. The same 16 read-only RPCs (account/read, account/usage/read, model/list, plugin/list, skills/list, experimentalFeature/list, ...) run against the real account and live config give identical responses directly and through the shim except two: config/read differs only in key order, app/list only in the random id of a 403 page returned both times. The reading reported in [Bug] Codex App (macOS): send button disabled at quota exhaustion - the #5947 intercept never sees the gate reads (they come from the bundled app-server) #6196 (Plus, 5-hour window 100 %, weekly 32 %) is opened with both windows kept as sent. A filter that throws while loading leaves the real server's output intact (this test fails without the empty-input probe). A system PAC whose switch would exceed the cap falls back with the system-pac-too-large route, and the switch stays within the cap. A launch, restore and status run with the app as the caller's ancestor find it, every generated script variant passes bash -n, and a script that does not parse is refused before launchd loads it (these fail on the previous launch-watcher.ts). Byte-preservation across chunk boundaries, a throwing rewrite passing its line through, launcher fail-open, the real binary keeping the launcher's pid with stdin, stderr and exit code untouched, the open --env launch/watch/restore paths and the write-boundary check are covered in tests/chatgpt-unblock/.
  • App-server shim enabled on the live ChatGPT desktop app (26.924, real account, not exhausted): the app-server is the process the app started (its parent is the app; the filter is a child of the server), the app-tools pipe accepts it (no dynamic_app_tools_peer_rejected), codex_app and codex_apps start, account/read, getAuthStatus and model/list succeed, and the account menu and model picker open. A first version that ran the binary as a child of the filter process was rejected on that pipe with untrusted-code-signing-identity, and codex_app failed with "Codex app tools pipe closed"; that is why the launcher now execs the binary.
  • macOS PAC fallback, measured on the real ChatGPT app (Chromium 154) in an isolated instance (own --user-data-dir, temp CODEX_HOME; the running app and proxy were not touched), counting the app's established TCP connections per route: file:// PAC: 0 via the entry, 0 via the system proxy, 3 direct :443 (the PAC is ignored); http:// PAC with opencodex up: 2 via the entry, 14 via the system proxy; http:// PAC with the listeners stopped: 0 via the entry, 15 via the system proxy. That is why the switch is now an inline data: PAC: with opencodex up 2 via the entry and 14 via the system proxy, with the listeners stopped 0 via the entry and 23 via the system proxy, with no restart and no dependency on opencodex to fetch the script. Headless Chrome 154 shows the same file:// behaviour.
  • Reporter-observed natural-exhaustion evidence (from [Bug] Codex App (macOS): send button disabled at quota exhaustion - the #5947 intercept never sees the gate reads (they come from the bundled app-server) #6196, not run by the author): on a Plus account with the 5-hour window exhausted and the weekly one not (the state a Pro account cannot reach), opencodex 2.73.0 with a local port of this PR, Desktop 26.928.22156, bundled CLI 0.159.2, the combined configuration (unblockSend: true, appServerShim: true) kept Send usable in the exhausted state — greyed out before launch, enabled after, account/plugin/thread-sync features intact, no peer_rejected. This is a combined-configuration observation, not a claim of shim-only sufficiency. The same reporter's single-variable A/B (appServerShim: false) also kept Send usable on that build: the gate read travelled the Chromium network stack (12+ ESTABLISHED to the listener, 0 from the app-server, whose rate-limit payload was all-null), so on current builds the network intercept is the load-bearing part and the shim is the fallback.
  • Still not run: the reverse control (unblockSend: false, appServerShim: true alone) and a live observation of a non-quota block passing through untouched on an exhausted account. It needs an account whose quota is really exhausted, which cannot be produced on demand.

Checklist

  • Scope stays focused and avoids unrelated cleanup.
  • Docs or release notes were updated when needed. (docs-site guide in all eight locales.)
  • Security-sensitive changes were reviewed for secrets, auth, and unsafe defaults. (The entry listener binds 127.0.0.1 only and refuses every request except CONNECT chatgpt.com:443, so it is never a general forward proxy. The embedded system PAC is the script Chromium would run anyway. The PAC file is written 0644 inside the config dir. The feature is off by default; no secrets are logged or stored.)

Review readiness checklist

This PR stays in draft until every box below is ticked. Tick all four boxes once the requirements are met:

  • Required local validation passed; commands, results, and any full-suite exception are documented.

  • I pushed my PR to a recent dev commit (at most 10 behind; a maintainer may still ask for the exact tip before merge).

  • I resolved all correct Codex and CodeRabbit findings.

  • My PR is ready for review.

Summary by CodeRabbit

  • New Features
    • Added an opt-in experimental macOS ChatGPT Desktop intercept that can rewrite recognized quota-related send restrictions in conversation and usage responses while preserving other restrictions.
    • Added WebSocket relay support and optional PAC-based routing, with system proxy fallback when PAC settings cannot be used.
    • Added launch-watcher installation, removal, and status commands. Watchers can restore app routing when required listeners are unavailable.
    • Added configuration for the intercept, PAC fallback, and listener port. The app-server shim remains independently configurable.
  • Documentation
    • Expanded setup guidance with certificate trust, credential-handling risks, routing behavior, and restoration details.

Builds on the maintainer's app-server shim (#6361, #6412) and send-unblock intercept (#6365), which this branch includes through #6381.

Co-authored-by: JUN jun@lidgeai.com

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

Labels

enhancement New feature or request intake: hygiene-blocked Deterministic PR hygiene checks failed

Projects

None yet

Development

Successfully merging this pull request may close these issues.

4 participants