align default models & additional small cleanup by tibo-openai · Pull Request #8 · openai/codex

tibo-openai · 2025-04-16T17:29:57Z

No description provided.

Three critical fixes for Android/Termux compatibility: 1. Disable sandbox on Android - Android/Termux does not support landlock/seccomp - Modified get_platform_sandbox() to return None on Android - File: core/src/safety.rs 2. Preserve LD_* environment variables on Android - Android needs LD_LIBRARY_PATH to find libraries in custom paths - Skip LD_* removal in process-hardening on Android - File: process-hardening/src/lib.rs 3. Use $SHELL instead of getpwuid() on Android - getpwuid() returns 'login' instead of actual shell on Termux - Use $SHELL environment variable which is correct - File: core/src/shell.rs Fixes: bash execution 'Permission denied' errors in Agent mode Platform: Android/Termux only (Linux/Mac/Windows unchanged)

- Document Patch #8: Fix bash execution in Agent mode on Android/Termux - Three critical fixes: sandbox disable, LD_* preservation, shell detection - Update npm package.json to 0.55.4-termux - Add missing codex.js wrapper for npm binary Tested successfully: Agent mode bash commands now work without Permission denied errors

Upstream changes (40+ commits): - New: codex-shell-tool-mcp MCP server - New: execpolicycheck CLI command - New: TUI animations toggle feature - New: Shell timeout increased to 1 hour - Refactor: execpolicy migration (execpolicy2 -> execpolicy) - Removed: tiktoken-rs, shell_command feature - Improved: FreeBSD portability, fuzzy search (8->20 results) Termux patches verified compatible: - #1 Browser login (termux-open-url) ✅ - #2 RAM optimizations (lto=false, codegen-units=16) ✅ - #8 Bash execution (sandbox, LD_*, shell detection) ✅ - #9 Auto-update execution ✅ Stats: 195 files changed, +5915 insertions, -2293 deletions 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>

…8+ requirement - Upstream: OpenAI Codex rust-v0.79.0 - Termux patches validated: #1-#6, #8, #9 - Testing: 47/49 passed (2 skipped: WebSearch, Git) - Documentation: - Added GLM-4.7 quickstart guide (docs/GLM4.7-quickstart.md) - Updated README with GLM-4.7 Path 2 in Quickstart - Node.js requirement updated to >=18 (recommended >=22) - Web search safety warning added - Tagline updated to "Built from upstream" - Package: npm-package synced with main README - Binaries: codex (59M), codex-exec (34M), codex-tui (41M) verified

- Updated workspace version to 0.80.0-termux - Termux patch #2 (compilation) preserved: lto=false, codegen-units=16 - Termux patch #9 (auto-update execution) preserved in main.rs - Process hardening removal upstream improves bash execution (patch #8) Upstream changes (77 commits from rust-v0.79.0): - Metrics capabilities (otel/metrics modules) - Requirements/list API - Thread fork endpoints - Elevated sandbox onboarding - Skills explicit invocation via V2 API - Updated dependencies: lru 0.16.3, zstd added - Removed codex-process-hardening from Codex CLI - Updated models.json

- Updated workspace version to 0.80.0-termux - All Termux patches verified functional (#1-#6, #9 active) - Patch #8 no longer required (resolved upstream by PR openai#8951) - Process hardening removed from Codex CLI improves bash execution - Test suite: 49/49 passed (CODEX_TEST_REPORT_v0.80.0.md) - Upstream changes: * Thread fork endpoints (conversation branching) * Requirements/list API * Elevated sandbox onboarding NUX * Skills explicit invocation via V2 API * Metrics capabilities (otel/metrics module) * Removed process hardening from Codex CLI - Binaries compiled: codex (60M), codex-tui (42M), codex-exec (35M), codex-app-server (38M) - Updated patches/README.md for v0.80.0

- Remove old test report (CODEX_TEST_REPORT_v0.79.0.md) - Update version references from 0.79.0 to 0.80.0 - Update upstream reference from rust-v0.79.0 to rust-v0.80.0 - Update test date from 2026-01-08 to 2026-01-10 - Update test results to 49/49 passed (0 skipped) - Update patch description: 8 patches (#1-#6, #9) - Patch #8 resolved upstream - Sync npm-package/README.md with root README.md

- CHANGELOG.md: Added 0.80.0-termux entry with: * Upstream rust-v0.80.0 release details * Process hardening removal (PR openai#8951) * New features: thread fork, requirements API, elevated sandbox NUX, skills V2, metrics * Patch #8 no longer required (resolved upstream) * Test results: 49/49 passed (0 skipped) * Documentation updates - README.md: Fixed remaining reference to old test report * Updated CODEX_TEST_REPORT_v0.79.0.md → CODEX_TEST_REPORT_v0.80.0.md * Updated test results: 47 passed → 49 passed, 2 skipped → 0 skipped - npm-package/README.md: Synchronized with root README.md * Same version and test result updates as README.md All linked files verified and aligned to v0.80.0-termux.

Discovered during live validation (2026-04-27): - openai#7: codex_models_manager refresh uses env-only auth (cosmetic, opaque error) - openai#8: rollout items not persisted for MiniMax sessions (blocks resume/fork) - openai#9: wire_probe example permanently available for opaque-error debugging Closed items section seeded with the 4 fixes shipped during validation (developer-role remap / system-merge / bridge lifecycle / wire dump).

Generalizes the consecutive-system coalesce shipped in c1579ac. While investigating TODO openai#8 (rollout persistence), discovered MiniMax actually enforces TWO related constraints: 1. No two adjacent system messages -> "invalid chat setting (2013)" 2. No system message after position 0 -> "chat content has invalid message role: system (2013)" Both surface from api.minimax.io/v1/chat/completions with the same opaque error code 2013. Constraint openai#1 was hit on first turns; openai#2 was hit only on resumed sessions where Codex emits a turn_context refresh as a mid-conversation system message. The probe matrix in wire_probe.rs confirms both behaviors live (probes 05a/05c/05d). Refactor: - coalesce_consecutive_system_messages -> consolidate_system_messages_to_leading - New behavior: collect every system body anywhere in the array, drop empty ones, emit ONE leading system message joined by blank lines, preserve insertion order of system bodies, preserve all non-system messages in their original relative order. - Content-preserving: the model sees every instruction body it would have seen, just hoisted to position 0. Tests: - New: translate_hoists_non_leading_system_to_position_zero (regression for the resumed-session shape). - New: translate_drops_empty_system_messages_during_consolidation. - Replaces translate_does_not_merge_non_adjacent_system_messages (whose old behavior is now incorrect). - 17 minimax_adapter tests passing. Validated: $ codrex exec resume <minimax-session-id> -m minimax/MiniMax-M2.7 "respond just OK" ... codex OK While investigating: confirmed `failed to record rollout items: thread X not found` is a PRE-EXISTING upstream bug (reproduces verbatim in `codex-cli 0.125.0`). The error is cosmetic — the rollout file is written to disk BEFORE the shutdown race that triggers the log line. Resume works correctly despite it. TODO openai#8 downgraded to cosmetic; the real blocker for MiniMax resume was this hoist fix, not the rollout race.

…+ add mid-conv system fix After investigation: - TODO openai#8 ("failed to record rollout items") is a PRE-EXISTING upstream bug reproducible in codex-cli 0.125.0. The rollout file IS persisted; the ERROR is racing log noise. resume works correctly. - The actual MiniMax-resume blocker was a distinct bug: mid-conversation system messages rejected by MiniMax. Fixed in 7f33ffe and added to Closed section. Net: Phase 3 unblocked. No deviation from upstream needed for openai#8.

… menu, shortcuts - openai#1 Agent tool execution now gates through sandbox.RequestApproval() before every tool call. Safe ops auto-approve; dangerous ops trigger WebSocket → frontend modal flow. - openai#5 Auto-compress context when msg count > 40 to keep last 8 pairs. Prevents token limit errors in long conversations. - openai#8 Chat '+' button now opens a dropdown with file picker and slash command shortcuts (/spec, /steer, /execute). - openai#9 Global keyboard shortcuts wired: Ctrl+Alt+B → toggle right panel, Ctrl+Alt+S → sidetasks, Ctrl+Shift+G → review, Ctrl+T → browser, Ctrl+P → files. - openai#6 API key empty check: provider returns clear error when key missing.

Covers the 4 deferred items from subproject 4's final review: - #1: TopBar cwd + model from ConversationState (reducer captures thread.cwd + model/rerouted; sendRequest-backed turn/start + thread/resume feed authoritative ThreadStartResponse/ThreadResumeResponse back via a synthetic responseMeta action) - openai#2: AgentPanel real sub-agent data (extract from collabAgentToolCall.agentsStates into state.subAgents) - openai#3: full styles.css var-ification (~50 hardcoded colors → theme variables) Key insight during research: ThreadStartResponse + ThreadResumeResponse both carry authoritative model + cwd via the Task 4.1 sendRequest channel, so #1b is clean (not partial as originally thought). Plan written across 4 Write/Edit calls to avoid the single-Write-too-long interruption that hit the first attempt (documented in plan §实现备注 openai#8).

…persistence (P1-5) Four hardening pieces on top of the P1-4 orchestrator (judge hooks already landed there): - circuit breaker (openai#6): consecutive same-signature tool failures trip at a threshold (default 5) and kill the task tree via the shared interrupt path; any success resets the streak. Signatures are deliberately coarse — shell-wrapper-stripped head token — because the observed 328-failure loop was one tool with varying arguments. Breaker-tripped tasks finalize as Failed (never Interrupted). - southbound auto-retry (openai#7): transient-failure relaunch of the same LaunchPlan, triple-gated — failure text matches known transient patterns (stream disconnected / connection refused / timeouts, from the Phase 0 window postmortems), zero tool activity observed (no side-effect risk), and a retry budget (2). Each attempt streams to its own events file (<id>.aN.jsonl); retries recorded on the task. - preflight (openai#7/openai#16): bare-IP endpoint probe before any resource allocation, with triage in the error — refused (host up, service down; hints at dspark_proxy), timeout (host likely down; suggest ping triage), or HTTP status anomaly. - diff persistence (openai#8): finalize now auto-commits the dirty worktree onto the task branch before reclaim (models edit files but never commit — previously KeepBranch preserved a ref pointing at base and the actual output evaporated with the worktree). diff --stat lands on the record and in status; full diff reviewable via the branch. Tests: 30 lib (signature/breaker/transient units, worktree commit-and-retrieve) + 10 e2e (breaker trip <10s, streak reset, retry-then-recover with per-attempt event files, non-transient no-retry with stderr surfaced, diff preserved through reclaim, preflight dead-vs-live endpoints). clippy clean. Authored by Claude Fable 5 (Claude Code).

完善对SVN的支持 PR-URL: openai#8

Preserve the pending trigger while messages remain, and restore Storybook composer focus when the queue drains during exit. Cover real exit animation continuity and connection changes across browser engines. Implements CNB openai#9; specification CNB openai#8.

External review (44 findings) triage — these 7 confirmed items fixed: - openai#22 codex-config and codex-core TEST targets had 11 compile errors (ModelProviderInfo literals missing experimental_bridge; a WireApi match missing Anthropic). All patched; test targets now compile under no/single/both bridge features. - openai#9 MCP flat-name resolution was DEAD in the production dispatch path: router.rs wraps the bridge's namespace-less flat name with with_default_namespace() BEFORE registry lookup, so the fallback's namespace.is_none() guard never fired — MCP tools were unresolvable via the chat bridges in real usage. The original unit test called registry.tool() directly with a plain name, bypassing the router (a false green). Fix: the fallback now resolves through the flat index regardless of the incoming namespace (safe: the index only contains namespaced tools, which the exact-key lookup already missed); new regression test drives the router-wrapped form. - openai#8 Both bridges synthesized fixed txt_0/rsn_0 item IDs every turn; retained assistant history dedupes by ID, so each new message REPLACED the previous one. IDs now carry a per-turn unique suffix (nanos + counter). - openai#31 The nextest live-test config never applied: it sat at the repository root (cargo workspace — and thus nextest config lookup — is rooted at codex-rs/) AND used invalid syntax ([[test-group]], group=). Merged into codex-rs/.config/nextest.toml with the official [test-groups.<name>] tables and test-group = override keys, per-test slow-timeout/retries with exponential backoff, and the ci profile the workflow references; config parse verified. Corollary: the earlier "per-vendor groups fixed rate limiting" claim was unfounded — the observed retries came from upstream's default retries=1. - openai#29 Cassette replay fell back to LIVE on missing/corrupt fixtures, silently consuming vendor quota in declared-offline runs; now fails fast with the record hint. - openai#30 The nightly workflow passed only API keys — vendors() skips entries without URLs/models, so every scenario would silently skip. The workflow now passes complete per-vendor config and fails when no vendor tests are listed. - openai#35 The marker assertion matched the raw JSON line (the command string contains the marker too), so broken stdout capture could still pass; now parses the completed event and asserts the aggregated_output field specifically. Verified: full live matrix 64/64 (mimo+glm+step), offline replay 43/43 with credentials removed, rig bridge units 18/18, config/core test targets compile.

…t [skip ci] Nextest resolves slow-timeout to the first matching override and parses filters lazily, so a mac window appended below the packaged rules would have looked applied while doing nothing - the four measurements that establish both are recorded next to the widening. The openai#8 leg's local non-reproduction and the ACP scratch that landed on the system disk are logged with their counts.

w

094f6e0

tibo-openai merged commit 1b40e19 into main Apr 16, 2025

tibo-openai deleted the tibo/codex/3 branch April 16, 2025 17:32

rumple pushed a commit to rumple/codex that referenced this pull request Apr 24, 2025

w (openai#8)

dd53ab1

zcg added a commit to byebye-code/codex that referenced this pull request Dec 3, 2025

[Claude Code] After prompt openai#8

12c7f2d

shawn-dsz mentioned this pull request Mar 20, 2026

macOS startup panic when SystemConfiguration DNS state is unavailable #15297

Closed

aresxu8866-hue mentioned this pull request Aug 16, 2026

[app][macOS][26.810.52044] Unbounded SkyComputerUseService respawn loop crashes the app in ~2 min — onServiceConnectionLost has no failure counter, backoff, or circuit breaker #38841

Open

sekfung pushed a commit to sekfung/codex that referenced this pull request Aug 20, 2026

合并来自 feat/svn-support 的合并请求 openai#8

b7cd6a8

完善对SVN的支持 PR-URL: openai#8

kratosboom mentioned this pull request Aug 22, 2026

feat(apex): establish authenticated runtime trust boundary GPRO-Master/ApexCode#3

Draft

weidian0039 mentioned this pull request Sep 12, 2026

App-created heartbeat automation keeps firing 9 times after the usage limit is reached, draining the whole monthly Codex quota #45070

Open

7596ff mentioned this pull request Sep 17, 2026

Linux/KDE: desktop app SIGSEGV after automatic Look and Feel/XSettings refresh #46113

Closed

zPrometheus98 mentioned this pull request Sep 28, 2026

ChatGPT Desktop 26.924.22138-1 crashed #48959

Closed

Navigation Menu

Repository navigation

align default models & additional small cleanup - #8

align default models & additional small cleanup#8
tibo-openai merged 1 commit into
mainfrom
tibo/codex/3

tibo-openai commented Apr 16, 2025

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant