fix(review): 28 verified bugs across gateway, pool, generator and GUI

MA
2d ago
+1768-149

fix(review): 28 verified bugs across gateway, pool, generator and GUI

A full-codebase review found 28 bugs that the test suite did not
catch. This PR fixes all of them, each with a regression test that
failed before the change and passes after.

High

  • One U+2028/U+2029/U+0085 character in a model reply killed the whole
    request: SSE frames were split with str.splitlines(), which also
    breaks on those separators that ensure_ascii=False JSON carries raw
    (upstream.py, translate_anthropic.py).
  • Missing ddddocr defeated the entire auto captcha strategy: the lazy
    OCR factory raised RuntimeError instead of returning None, so all 15
    attempts failed with a misleading log (captcha_solver.py).

Pool / accounts

  • cool()/kill_after_strikes() downgraded quota-drained (EXHAUSTED)
    accounts to COOLING, putting them back in rotation after the
    cooldown (pool.py).
  • upsert/enable/apply_credentials(reset=True) never reset
    quota_strikes, so an account that recovered from quota exhaustion
    re-retired on its first rejection with 3 banked strikes (pool.py,
    tokens.py); upsert also force-flipped live state to READY against
    its own docstring - re-add/login/import flows now revive explicitly
    via enable() (api_admin.py, api_dash.py).
  • The edge-block notification callback fired per blocked response
    instead of once per block episode (breaker.py).

API / service

  • /health returned 500 (hmac TypeError) on a non-ASCII update header;
    after any in-app update the nonce persists for the process lifetime,
    so internet scanners kept hitting it (api_health.py).
  • generator/locate left the old manager's ClaimScheduler, captcha-mint
    thread and proxy checker running; a second scheduler then fired
    duplicate daily claims (api_dash.py).
  • Worker/claim stop and proxy-checker endpoints ran subprocess kills
    (up to 5s) and multi-MB pool.json parsing directly on the event
    loop, freezing all traffic including live SSE streams
    (api_admin.py, api_dash.py); _check_update_idle also iterated
    manager dicts without the lock (500 on a rare race).
  • The request body was deep-copied on the event loop for every
    recorded request even when JSONL capture was disabled
    (service.py).
  • proxy auto-verify 500'd on a non-numeric interval_min instead of
    400 (api_admin.py); a non-ASCII comment in proxies.txt aborted the
    whole Auto Generate preflight (generator.py).

Relays

  • The relay retry re-sent through the relay that had just failed,
    hammering the same dead relay for every attempt while never trying
    the others (upstream.py).
  • An unreadable relays.json was cached as authoritative defaults, so
    the next save permanently wiped the user's relays
    (relay_manager.py); update_options(models="GLM-5.3") exploded the
    string into single characters and silently routed nothing.

Generator

  • Telemetry and exported snapshots carried one shared hardcoded
    x-device-mid and a hardcoded Asia/Tehran timezone while headers
    claimed the exit IP's zone - the cross-account correlation signals
    risk control clusters on (invite_claim.py, export_starter.py,
    plan_check.py, claim500.py).
  • One flaky account aborted the whole recheck/export batch
    (recheck_plans.py, export_starter.py); a portless ZAI_PROXY became
    socks5://host:None and crashed the Playwright launch
    (invite_claim.py); leading-zero octets (0177.0.0.1 = 127.0.0.1
    octal) bypassed the loopback SSRF guard (proxy_parse.py).

GUI / desktop

  • Clicking Stop during a pending Save & Restart was silently
    overridden and the gateway restarted itself; Stop during the startup
    import window was a no-op and the server bound anyway; wait_ready()
    could health-check a foreign process on the default port (gui.py).
  • config.yaml - which holds the per-install admin/client keys - was
    written non-atomically from four GUI paths; a crash mid-write
    truncated it (gui.py).
  • The live smoke script closed the store under a pending usage-record
    write (scripts/smoke/live.py).

Verification

  • 1364 passed / 147 skipped / 0 failed (full suite, clean venv)
  • ruff E9,F63,F7,F82,F811,F841: clean
  • captcha-mint sidecar (bun): 30/30
  • Known pre-existing flake test_overload_rotates_accounts_after_wait_
    without_cooling reproduces on the unmodified tree and passes here;
    unrelated to these changes.

Every fix has a regression test that was observed failing before and
passing after; 5 new test files + extensions to test_relays.py.


Summary by cubic

A full-codebase review found 28 bugs that the test suite did not catch. This PR fixes all of them, each with a regression test that failed before the change and passes after.

High-impact fixes

  • SSE frames now split on real line terminators, so a stray U+2028/U+2029/U+0085 in a reply no longer fails the request.
  • Missing ddddocr degrades captcha solving instead of aborting the whole auto captcha strategy.
  • Quota-drained (EXHAUSTED) accounts stay retired; fresh credentials get a clean strike budget.
  • Relay retries no longer hammer a just-failed relay, and an unreadable relays.json can no longer be overwritten by defaults.
  • Subprocess kills, pool parsing, and request-body deep copies moved off the event loop, so live SSE traffic is no longer frozen.
  • All GUI config.yaml writes are atomic; a user Stop now wins over a pending restart.
  • Generator telemetry uses per-account device IDs and the exit IP's timezone instead of shared hardcoded values.

Verification

  • 1364 passed / 147 skipped / 0 failed; ruff and the captcha-mint sidecar are clean.