Codex review: needs real behavior proof before merge. Reviewed July 15, 2026, 12:03 AM ET / 04:03 UTC.

Summary
The PR makes omitted or false message.final source replies continue a Codex app-server turn, reserves inferred release for explicit final: true, preserves tool-owned termination and source-reply completion telemetry, updates schema guidance, and adjusts regression tests.

PR surface: Source -5, Tests +8. Total +3 across 3 files.

Reproducibility: yes. at source level: current terminal classification releases a message_tool_only turn after an omitted-final confirmed source reply, matching the reported log and lost-follow-up behavior. A live current-head channel reproduction is not included.

Review metrics: 1 noteworthy metric.

• Inferred terminal paths: 2 changed. Delivered-source and receipt-confirmed source replies with omitted final stop releasing the turn, so both continuation and recovery behavior need live proof.

Root-cause cluster
Relationship: fixed_by_candidate
Canonical: #106961
Summary: The issue is canonical; this PR and #107274 are competing candidate fixes, while the merged generic-runner PR is sibling precedent rather than a covering fix.

Members:

• canonical: [Bug]: Discord/Codex runtime: message tool is terminal, so progress updates silently end the turn #106961 - It reports the exact Discord/Codex omitted-final progress message terminating the turn.
• fixed_by_candidate: fix(codex): continue after source-reply delivery instead of terminating turn #107274 - It is an overlapping open candidate fix for the same canonical issue, with different recovery and termination details.
• partial_overlap: fix(agent): continue after source message tool replies #92343 - It fixed the analogous generic embedded-runner source-reply continuation path but predates and does not replace the Codex app-server implementation.
• same_root_cause: Fix Codex message-tool-only source reply completion #95942 - This merged Codex completion work introduced the confirmed-delivery terminal behavior whose omitted-final semantics now cause the regression.

Proposal only: this assessment does not dispatch repair, suppress jobs, mutate sibling items, close, or merge anything.

Merge readiness
Overall: 🦪 silver shellfish
Proof: 🦪 silver shellfish
Patch quality: 🐚 platinum hermit
Result: blocked until real behavior proof from a real setup is added.

Overall follows the weaker of proof and patch quality, so missing proof can cap an otherwise strong patch.

Rank-up moves:

• [P1] Add redacted live Discord or Telegram evidence showing progress delivery, subsequent tool execution, and exactly one final reply.
• Include a timeout or error recovery observation showing the delivered progress message is not replayed.

Proof guidance:

• [P1] Needs real behavior proof before merge: Focused tests and CI exercise the bridge but do not show the changed behavior in a real Codex-backed channel; add a redacted transcript, recording, terminal output, or runtime log showing progress, a later tool, one final reply, and recovery without duplication, then update the PR body to trigger review or ask a maintainer for @clawsweeper re-review.

Mantis proof suggestion
Telegram Desktop proof can directly show that an omitted-final Codex progress message does not release the turn and that one final reply follows after another tool. A maintainer can ask Mantis to capture proof by posting this exact PR comment:

@openclaw-mantis telegram desktop proof: verify an omitted-final Codex progress message continues to a follow-up tool, produces exactly one final reply, and is not replayed on recovery.

Risk before merge

• [P2] A unit-tested terminal-classification change can still duplicate or suppress channel output if progress delivery, later final delivery, timeout recovery, and stranded-reply recovery interact differently in a real Codex-backed channel run.
• [P1] Two open PRs modify the same decision surface with different treatment of completion telemetry and tool-owned termination; merging either without selecting one canonical semantic contract risks follow-up drift.

Maintainer options:

1. Prove the complete message lifecycle (recommended)
   Before merge, capture a redacted live Codex-backed channel run showing progress delivery, continued tool execution, exactly one final reply, and no duplicate during timeout or error recovery.
2. Accept test-only delivery risk
   Maintainers may intentionally merge on source and unit-test evidence, owning the unresolved real-channel exactly-once delivery risk.
3. Pause for canonical branch selection
   Keep both overlapping PRs open until a Codex-area owner chooses one semantic contract or prepares a narrow replacement.

Next step before merge

• [P1] A Codex-area owner must select the canonical branch and review live exactly-once delivery proof; there is no narrow code defect for an automated repair worker.

Maintainer decision needed

• Question: Which Codex source-reply terminal contract should become canonical for the linked regression: this PR's omitted-as-progress behavior with the completed marker preserved, the competing branch's broader classifier change, or a maintainer replacement?
• Rationale: Both open PRs target the same high-risk message lifecycle, but differ on recovery telemetry and tool-owned termination; tests alone cannot establish the intended exactly-once channel contract.
• Likely owner: omarshahine — This person introduced and merged the current Codex message-tool-only completion behavior, making them the strongest available owner for its terminal and recovery semantics.
• Options:
  • Use this branch after live proof (recommended): Keep the narrower explicit-final release rule and preserved completion marker, then require a live channel continuation and recovery transcript before merge.
  • Adopt the competing branch: Select fix(codex): continue after source-reply delivery instead of terminating turn #107274 only after it preserves or atomically redesigns exactly-once recovery and supplies equivalent live proof.
  • Create a maintainer replacement: Close both contributor branches only after a narrow owner-authored replacement exists and proves the same progress, final, and recovery behavior.

Security
Cleared: The focused TypeScript runtime, schema text, and test changes add no dependency, workflow, permission, secret, install, publishing, or downloaded-code execution surface.

Live Telegram evidence

Environment: real Telegram DM to the test bot, OpenClaw Codex app-server backed by OpenAI gpt-5.6-sol, message_tool_only source delivery, Node 24.18.0, pnpm 11.2.2.

Controlled fixture used identically for before/after: the workspace instructions required the intermediate progress message call to omit the optional final field, continue with the requested tool action, and send the completed reply exactly once with final: true. No OpenClaw source was changed for the fixture.

The channel transcript and runtime excerpts below are redacted: they expose no token, phone number, personal username, chat/user ID, real name, or absolute home/workspace path.

Before (main @ 54f70c5d8c040f935b5d6aaa72ff514387bfae2c)

Telegram transcript:

User: Send me a short progress message right now saying you're starting. Then list the files in your current working directory. Then send me the final summary of what you found.
Bot: I’m starting now—I'll inspect the current working directory and report back.

No directory result or final reply arrived before the next /new.

Native Codex rollout lines 12, 16, and 21, selected in their original order; only identifiers were redacted:

{"timestamp":"2026-07-15T12:07:04.349Z","type":"response_item","payload":{"type":"custom_tool_call","id":"[redacted]","status":"completed","call_id":"[redacted]","name":"exec","input":"const r = await tools.message({action:\"send\", message:\"I’m starting now—I'll inspect the current working directory and report back.\"});\ntext(r);\n","internal_chat_message_metadata_passthrough":{"turn_id":"[redacted]"}}}
{"timestamp":"2026-07-15T12:07:28.671Z","type":"response_item","payload":{"type":"function_call","id":"[redacted]","name":"wait","arguments":"{\"cell_id\":\"1\",\"yield_time_ms\":10000,\"max_tokens\":1000}","call_id":"[redacted]","internal_chat_message_metadata_passthrough":{"turn_id":"[redacted]"}}}
{"timestamp":"2026-07-15T12:07:41.092Z","type":"event_msg","payload":{"type":"turn_aborted","turn_id":"[redacted]","reason":"interrupted","completed_at":1784117261,"duration_ms":45956}}

Gateway lines, selected in their original order; only chat/run/session IDs were redacted:

2026-07-15T17:52:39.607+05:45 [telegram] outbound send ok accountId=default chatId=[redacted-chat-id] messageId=20 operation=sendMessage deliveryKind=text chunkCount=1
2026-07-15T17:52:40.594+05:45 [agent/embedded] codex app-server turn released after terminal dynamic tool result
2026-07-15T17:52:41.327+05:45 [agent/embedded] embedded run done: runId=[redacted] sessionId=[redacted] durationMs=76314 aborted=false

Observation: main delivered the omitted-final progress message and immediately released the turn. The native rollout then ended without a shell_command call or final message call.

After (fix/codex-omitted-final-progress-release @ 2c256aae48ca22c55a4112ea39b927372f3828a2)

This local head is one commit above upstream main 69ef3ca980a7262577c6e6bcb3d8abc08a5325b1. Targeted runtime build passed. Focused Vitest result: 2 files passed, 115 tests passed.

Telegram transcript:

User: Send me a short progress message right now saying you're starting. Then list the files in your current working directory. Then send me the final summary of what you found.
Bot (progress): I’m starting now—checking the current working directory.
Bot (final): Found 9 items in `[redacted-workspace]`:

- `.git/`
- `AGENTS.md`
- `BOOTSTRAP.md`
- `HEARTBEAT.md`
- `IDENTITY.md`
- `openclaw-workspace-state.json`
- `SOUL.md`
- `TOOLS.md`
- `USER.md`

Summary: the directory is a newly initialized OpenClaw workspace with Git metadata, configuration/personality files, and the bootstrap file still present. There is no `memory/` directory yet.

Native Codex rollout lines 12, 19, 23, and 27, selected in their original order; only identifiers and the absolute workspace path were redacted:

{"timestamp":"2026-07-15T13:14:24.683Z","type":"response_item","payload":{"type":"custom_tool_call","id":"[redacted]","status":"completed","call_id":"[redacted]","name":"exec","input":"const r = await tools.message({action:\"send\", message:\"I’m starting now—checking the current working directory.\"}); text(r)\n","internal_chat_message_metadata_passthrough":{"turn_id":"[redacted]"}}}
{"timestamp":"2026-07-15T13:14:44.577Z","type":"response_item","payload":{"type":"custom_tool_call","id":"[redacted]","status":"completed","call_id":"[redacted]","name":"exec","input":"const r = await tools.shell_command({command:\"Get-ChildItem -Force | Select-Object Mode, Length, LastWriteTime, Name | Format-Table -AutoSize\",\"workdir\":\"[redacted-workspace]\"}); text(r)\n","internal_chat_message_metadata_passthrough":{"turn_id":"[redacted]"}}}
{"timestamp":"2026-07-15T13:14:53.067Z","type":"response_item","payload":{"type":"custom_tool_call","id":"[redacted]","status":"completed","call_id":"[redacted]","name":"exec","input":"const r = await tools.message({action:\"send\", final:true, message:\"Found 9 items in `[redacted-workspace]`:\\n\\n- `.git/`\\n- `AGENTS.md`\\n- `BOOTSTRAP.md`\\n- `HEARTBEAT.md`\\n- `IDENTITY.md`\\n- `openclaw-workspace-state.json`\\n- `SOUL.md`\\n- `TOOLS.md`\\n- `USER.md`\\n\\nSummary: the directory is a newly initialized OpenClaw workspace with Git metadata, configuration/personality files, and the bootstrap file still present. There is no `memory/` directory yet.\"}); text(r)\n","internal_chat_message_metadata_passthrough":{"turn_id":"[redacted]"}}}
{"timestamp":"2026-07-15T13:14:57.259Z","type":"event_msg","payload":{"type":"turn_aborted","turn_id":"[redacted]","reason":"interrupted","completed_at":1784121297,"duration_ms":40886}}

Gateway lines, selected in their original order; only chat/run/session IDs were redacted:

2026-07-15T18:59:40.792+05:45 [telegram] outbound send ok accountId=default chatId=[redacted-chat-id] messageId=34 operation=sendMessage deliveryKind=text chunkCount=1
2026-07-15T18:59:56.736+05:45 [telegram] outbound send ok accountId=default chatId=[redacted-chat-id] messageId=35 operation=sendMessage deliveryKind=text chunkCount=1
2026-07-15T18:59:56.876+05:45 [agent/embedded] codex app-server turn released after terminal dynamic tool result
2026-07-15T18:59:57.438+05:45 [agent/embedded] embedded run done: runId=[redacted] sessionId=[redacted] durationMs=54742 aborted=false

Observation: the omitted-final progress call delivered without releasing the turn; Codex then executed Get-ChildItem, sent exactly one final: true reply, and only that explicit final released the turn. The native turn_aborted event is the expected app-server interruption caused by explicit-final release; the OpenClaw gateway records the completed run as aborted=false.

Error-recovery observation (same patch, earlier byte-identical local rebase @ b67d98468a41a365e9ecf06d66577046f4a96b1c)

Telegram transcript:

User: Send me a short progress message right now saying you're starting the recovery check. Then use the directory-listing tool on this deliberately missing path so it fails: Z:\openclaw-proof-missing-7c2f. After the error, recover by listing the files in your current working directory. Then send exactly one final summary. Do not repeat the progress message.
Bot (progress): I’m starting the recovery check.
Bot (final): Recovery check completed: the missing Z: path failed as expected, then I successfully listed the current workspace files (including AGENTS.md, BOOTSTRAP.md, SOUL.md, TOOLS.md, and USER.md).

Native Codex rollout lines 12, 15, 16, 18, 19, 21, and 25, selected in their original order; only identifiers and the absolute workspace path were redacted:

{"timestamp":"2026-07-15T12:33:38.813Z","type":"response_item","payload":{"type":"custom_tool_call","id":"[redacted]","status":"completed","call_id":"[redacted]","name":"exec","input":"const r = await tools.message({action:\"send\", message:\"I’m starting the recovery check.\"});\ntext(r);\n","internal_chat_message_metadata_passthrough":{"turn_id":"[redacted]"}}}
{"timestamp":"2026-07-15T12:33:46.548Z","type":"response_item","payload":{"type":"custom_tool_call","id":"[redacted]","status":"completed","call_id":"[redacted]","name":"exec","input":"const r = await tools.shell_command({command:\"Get-ChildItem -LiteralPath 'Z:\\\\openclaw-proof-missing-7c2f'\",\"workdir\":\"[redacted-workspace]\"});\ntext(r);\n","internal_chat_message_metadata_passthrough":{"turn_id":"[redacted]"}}}
{"timestamp":"2026-07-15T12:33:47.163Z","type":"response_item","payload":{"type":"custom_tool_call_output","call_id":"[redacted]","output":[{"type":"input_text","text":"Script failed\nWall time 0.6 seconds\nOutput:\n"},{"type":"input_text","text":"Script error:\nExit code: 1\nWall time: 0.3 seconds\nOutput:\nGet-ChildItem : Cannot find drive. A drive with the name 'Z' does not exist.\r\nAt line:2 char:1\r\n+ Get-ChildItem -LiteralPath 'Z:\\openclaw-proof-missing-7c2f'\r\n+ ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~\r\n    + CategoryInfo          : ObjectNotFound: (Z:String) [Get-ChildItem], DriveNotFoundException\r\n    + FullyQualifiedErrorId : DriveNotFound,Microsoft.PowerShell.Commands.GetChildItemCommand\r\n \r\n"}],"internal_chat_message_metadata_passthrough":{"turn_id":"[redacted]"}}}
{"timestamp":"2026-07-15T12:33:50.042Z","type":"response_item","payload":{"type":"custom_tool_call","id":"[redacted]","status":"completed","call_id":"[redacted]","name":"exec","input":"const r = await tools.shell_command({command:\"Get-ChildItem -Force\",\"workdir\":\"[redacted-workspace]\"});\ntext(r);\n","internal_chat_message_metadata_passthrough":{"turn_id":"[redacted]"}}}
{"timestamp":"2026-07-15T12:33:50.483Z","type":"response_item","payload":{"type":"custom_tool_call_output","call_id":"[redacted]","output":[{"type":"input_text","text":"Script completed\nWall time 0.4 seconds\nOutput:\n"},{"type":"input_text","text":"Exit code: 0\nWall time: 0.3 seconds\nOutput:\n\r\n\r\n    Directory: [redacted-workspace]\r\n\r\n\r\nMode                 LastWriteTime         Length Name                                                                 \r\n----                 -------------         ------ ----                                                                 \r\nd--h--         7/15/2026   2:57 PM                .git                                                                 \r\n-a----         7/15/2026   2:57 PM           7196 AGENTS.md                                                            \r\n-a----         7/15/2026   2:57 PM           1669 BOOTSTRAP.md                                                         \r\n-a----         7/15/2026   2:57 PM            244 HEARTBEAT.md                                                         \r\n-a----         7/15/2026   2:57 PM           1278 IDENTITY.md                                                          \r\n-a----         7/15/2026   2:57 PM             70 openclaw-workspace-state.json                                        \r\n-a----         7/15/2026   2:57 PM           1518 SOUL.md                                                              \r\n-a----         7/15/2026   4:37 PM           1389 TOOLS.md                                                             \r\n-a----         7/15/2026   2:57 PM            537 USER.md                                                              \r\n\r\n\r\n"}],"internal_chat_message_metadata_passthrough":{"turn_id":"[redacted]"}}}
{"timestamp":"2026-07-15T12:33:53.320Z","type":"response_item","payload":{"type":"custom_tool_call","id":"[redacted]","status":"completed","call_id":"[redacted]","name":"exec","input":"const r = await tools.message({action:\"send\", final:true, message:\"Recovery check completed: the missing Z: path failed as expected, then I successfully listed the current workspace files (including AGENTS.md, BOOTSTRAP.md, SOUL.md, TOOLS.md, and USER.md).\"});\ntext(r);\n","internal_chat_message_metadata_passthrough":{"turn_id":"[redacted]"}}}
{"timestamp":"2026-07-15T12:33:55.578Z","type":"event_msg","payload":{"type":"turn_aborted","turn_id":"[redacted]","reason":"interrupted","completed_at":1784118835,"duration_ms":26240}}

Gateway recovery lines, selected in their original order; only chat/run/session IDs were redacted:

2026-07-15T18:18:43.001+05:45 [telegram] outbound send ok accountId=default chatId=[redacted-chat-id] messageId=29 operation=sendMessage deliveryKind=text chunkCount=1
2026-07-15T18:18:47.005+05:45 [agent/embedded] codex app-server stderr: {"timestamp":"2026-07-15T12:33:47.005195Z","level":"ERROR","fields":{"error":"Exit code: 1\nWall time: 0.3 seconds\nOutput:\nGet-ChildItem : Cannot find drive. A drive with the name 'Z' does not exist.\r\nAt line:2 char:1\r\n+ Get-ChildItem -LiteralPath 'Z:\\openclaw-proof-missing-7c2f'\r\n+ ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~\r\n    + CategoryInfo          : ObjectNotFound: (Z:String) [Get-ChildItem], DriveNotFoundException\r\n    + FullyQualifiedErrorId : DriveNotFound,Microsoft.PowerShell.Commands.GetChildItemCommand\r\n \r\n"},"target":"codex_core::tools::router"}
2026-07-15T18:18:55.428+05:45 [telegram] outbound send ok accountId=default chatId=[redacted-chat-id] messageId=30 operation=sendMessage deliveryKind=text chunkCount=1
2026-07-15T18:18:55.507+05:45 [agent/embedded] codex app-server turn released after terminal dynamic tool result
2026-07-15T18:18:55.947+05:45 [agent/embedded] embedded run done: runId=[redacted] sessionId=[redacted] durationMs=36721 aborted=false

Recovery observation: the omitted-final progress message was delivered once, the deliberate missing-drive tool failed, Codex continued with a successful directory listing, and exactly one final reply followed. The progress message was not replayed.

Exactly one progress message, one continued tool execution, one final reply, no duplicates.

@clawsweeper re-review

🦞🧹
ClawSweeper re-review requested.

I asked ClawSweeper to review this item again.
Action: item re-review queued (workflow sweep.yml, event repository_dispatch).
Result: the existing ClawSweeper review comment will be edited in place when the review finishes.

Re-review progress:

• State: Complete
• Detail: The item is closed; no stale verdict was published.
• Run: https://github.com/openclaw/clawsweeper/actions/runs/29478725775
• Updated: 2026-07-16T07:06:55.305Z

Closing as superseded by maintainer PR #108487, which resolves #106961 with authoritative turn-completion finalization.

Thanks @mrfandu1 for the careful contribution and especially the live Telegram before/after and recovery evidence. Your PR correctly identified that an omitted final must not immediately release the Codex turn.

We landed the maintainer replacement at #108487. The remaining gap was finality settlement: simply continuing the turn while treating the delivered reply as already complete could suppress recovery after a later failure. The landed fix carries omitted finality to the authoritative turn outcome, promotes only the latest omitted reply on success, and leaves it non-final on failure.

Closing this as superseded by the merged root-cause fix. Thank you—your investigation and proof materially helped validate the direction.