Fixes #106961

Summary

Codex-backed agent turns prematurely terminated after an intermediate source-reply message send, preventing subsequent tool execution; this fix removes the implicit terminate-on-delivery classification so the turn continues until the normal Codex turn/completed lifecycle.

• Problem: In message_tool_only mode, a successful message-tool send to the source channel set terminate=true on the dynamic tool response, causing scheduleTurnReleaseAfterTerminalDynamicTool to interrupt the turn before subsequent tools could execute.
• Solution: Remove deliveredSourceReply and receiptConfirmedSourceReply from the withDynamicToolTermination conditions; only an explicit final=true argument now terminates the turn, matching the generic embedded-runner behavior from PR fix(agent): continue after source message tool replies #92343.
• What changed: extensions/codex/src/app-server/dynamic-tools.ts (terminate classification + sourceReplyFinal telemetry), extensions/codex/src/app-server/dynamic-tools.test.ts (updated assertions + new regression test)
• What did NOT change: shouldReleaseTurnAfterTerminalDynamicTool, resolveTerminalDynamicToolBatchAction, shouldBlockTerminalReleaseForNonTerminalDynamicToolResult, generic embedded-runner suppression logic, finalize path, any config/env/schema surface

Change Type (select all)

• Bug fix
• Feature
• Refactor required for the fix
• Docs
• Security hardening
• Chore/infra

Scope (select all)

• Gateway / orchestration
• Skills / tool execution
• Auth / tokens
• Memory / storage
• Integrations
• API / contracts
• UI / DX
• CI/CD / infra

Linked Issue/PR

• Related [Bug]: Discord/Codex runtime: message tool is terminal, so progress updates silently end the turn #106961
• Related fix(agent): continue after source message tool replies #92343 (prior generic fix establishing the desired behavior)
• This PR fixes a bug or regression

Motivation

Issue #106961 reports a Codex-specific recurrence of premature message-tool turn termination on v2026.7.1. The prior generic fix (PR #92343, commit 4620929) corrected the embedded runner to continue after source-reply delivery, but PR #95942 (commit 9b9a124c) re-introduced the same termination behavior in the Codex app-server path. Commit 4dc6e182c3 further hardened the incorrect "degrade to terminate-on-delivery when final is omitted" logic. This regression breaks Discord/Codex workflows where an agent sends an acknowledgment ("One moment") and then needs to execute further tools.

Real behavior proof (required for external PRs)

Behavior addressed: Codex dynamic tool bridge terminates the turn immediately after a successful source-reply message send, preventing subsequent tool calls in the same turn.

Real environment tested: Linux, Node v24.13.1, branch fix/codex-source-reply-terminate-106961

Exact steps or command run after this patch:

# ── Proof A: Real runtime proof via node --import tsx ──
node --import tsx -e "
import { createCodexDynamicToolBridge } from './extensions/codex/src/app-server/dynamic-tools.ts';
// ... creates a real bridge with message_tool_only mode,
// sends an intermediate message (final omitted), then calls exec
"

# ── Proof B: Full vitest test suites (175 tests) ──
node --experimental-vm-modules node_modules/.bin/vitest run \
  extensions/codex/src/app-server/dynamic-tools.test.ts

node --experimental-vm-modules node_modules/.bin/vitest run \
  extensions/codex/src/app-server/dynamic-tool-execution.test.ts \
  extensions/codex/src/app-server/run-attempt.dynamic-tools.test.ts

node --experimental-vm-modules node_modules/.bin/vitest run \
  src/agents/embedded-agent-runner/run/message-tool-terminal.test.ts

Proof A runs the actual patched createCodexDynamicToolBridge directly via node --import tsx — the same function the real OpenClaw runtime uses, not a copied boolean classifier. Proof B runs the standard vitest suites.

Evidence after fix:

# ── Proof A: Node --import tsx runtime proof ──
$ node --import tsx -e "..."

=== Full Runtime Proof: #106961 Fix ===

Test 1 — Intermediate message (final omitted):        # Regression scenario
  terminate: undefined ✅ OK                           # Turn NOT released
  telemetry.didDeliverSourceReplyViaMessageTool: true ✅
  exec success: true ✅                                # Subsequent tool executes

Test 2 — Final reply (final=true):
  terminate: true ✅ OK                                # Explicit final still terminates

Test 3 — Progress update (final=false):
  terminate: undefined ✅ OK                           # Explicit progress keeps alive

Test 4 — Non-message_tool_only mode (automatic):
  terminate: undefined ✅ OK                           # Other modes unaffected

=== OVERALL: ALL PASSED ✅ ===

# ── Proof B: Vitest test suites ──
$ node --experimental-vm-modules node_modules/.bin/vitest run \
  extensions/codex/src/app-server/dynamic-tools.test.ts

 Test Files  1 passed (1)
      Tests  106 passed (106)

$ node --experimental-vm-modules node_modules/.bin/vitest run \
  extensions/codex/src/app-server/dynamic-tool-execution.test.ts \
  extensions/codex/src/app-server/run-attempt.dynamic-tools.test.ts

 Test Files  2 passed (2)
      Tests  35 passed (35)

$ node --experimental-vm-modules node_modules/.bin/vitest run \
  src/agents/embedded-agent-runner/run/message-tool-terminal.test.ts

 Test Files  2 passed (2)
      Tests  34 passed (34)

──────────────────────────────────────────────────────────────
 Total: 5 test files, 175 tests, all passed.

The key regression test "continues after an intermediate message-tool-only source reply without final" validates the exact #106961 scenario:

1. Intermediate message sent via Codex message tool with sourceReplyDeliveryMode=message_tool_only
2. result.terminate === undefined — turn is NOT released
3. Subsequent tool call (exec) executes normally — success === true
4. telemetry.didDeliverSourceReplyViaMessageTool === true — delivery recorded
5. No terminal-release log appears after the intermediate message — the turn stays alive
6. sourceReplyFinal=true is preserved for all source-reply sends (stranded-reply recovery safety)

Observed result after fix:

1. Intermediate message-tool-only source replies (final omitted) no longer set terminate=true
2. sourceReplyFinal telemetry remains true when final is omitted for delivered/receipt source replies — preserves completed marker for stranded-reply recovery (no duplicate replay on error/timeout)
3. Receipt-confirmed source replies (final omitted) no longer set terminate=true
4. Explicit final=true + actual source reply delivery still terminates the turn
5. Cross-channel message-tool sends with final=true do NOT terminate (not a source reply)
6. Tool-returned terminate:true (toolConfirmedSourceReply) still terminates unless overridden by final=false
7. Yield results still terminate correctly
8. didDeliverSourceReplyViaMessageTool and messagingToolSourceReplyPayloads are still correctly collected

What was not tested: Live Discord/Codex run with a real Codex-backed agent; requires Crabbox/Testbox + Discord credential. stranding-recovery duplicate protection relies on sourceReplyFinal=true remained consistent — the Completed marker is preserved so recovery will not re-deliver an already-successful message.

Root Cause

PR #95942 (commit 9b9a124c) added Codex app-server source-reply support by including deliveredSourceReply and receiptConfirmedSourceReply in the withDynamicToolTermination conditions. This re-introduced the premature termination that PR #92343 (commit 4620929) had just removed from the generic embedded runner. Commit 4dc6e182c3 further hardened the incorrect "final omitted → degrade to terminate-on-delivery" fallback, making the regression harder to detect.

Regression Test Plan

• pnpm test extensions/codex/src/app-server/dynamic-tools.test.ts — covers all source-reply terminate scenarios including the new "continues after an intermediate message-tool-only source reply without final" test
• pnpm test extensions/codex/src/app-server/dynamic-tool-execution.test.ts — covers shouldReleaseTurnAfterTerminalDynamicTool (unchanged by this fix)
• pnpm test extensions/codex/src/app-server/run-attempt.dynamic-tools.test.ts — covers terminal diagnostics (unchanged)
• pnpm test src/agents/embedded-agent-runner/run/message-tool-terminal.test.ts — covers generic hook behavior (unchanged)
• pnpm test src/agents/embedded-agent-subscribe*.test.ts — covers duplicate-output suppression (unchanged)

User-visible / Behavior Changes

Codex-backed Discord agents will no longer stop executing after sending an intermediate acknowledgment message. The agent will continue the turn, executing subsequent tools and delivering the final result, matching the behavior of the generic embedded runner.

Security Impact

• New permissions/capabilities? No
• Secrets/tokens handling changed? No
• New/changed network calls? No
• Command/tool execution surface changed? No
• Data access scope changed? No

Best-fix Verdict

• Best fix: Yes — correcting the Codex terminal classification to match the established generic source-reply contract is the narrowest fix that addresses the regression without introducing new surfaces.
• Refactor needed: No
• Alternative considered: Adding a new config option to control terminate-on-delivery was rejected per AGENTS.md "Config/env surface bar is high" and the Codex review's explicit requirement "without adding a new config option or parallel message schema."

AI Assistance

• AI-assisted: Yes
• AI model: glm-5.1
• Human confirmed understanding of code changes: Yes
• AI prompts / session excerpts: Issue analysis, source code tracing across dynamic-tools.ts, embedded-agent-subscribe.ts, agent-loop.ts, and upstream Codex protocol (codex-rs/protocol/src/dynamic_tools.rs)

Risks and Mitigations

Highest-risk area: Duplicate channel messages after fix — if the generic-layer suppression (suppressMessageToolOnlySourceReplyOutput) does not correctly receive the didDeliverSourceReplyViaMessageTool flag, the agent could send both the intermediate message and the final output to the source channel.

Mitigation: The didDeliverSourceReplyViaMessageTool and messagingToolSourceReplyPayloads collection occurs before withDynamicToolTermination in the code path and is not gated by the terminate flag. The generic suppression layer checks messagingToolSourceReplyPayloads.length > 0 to activate suppression, which remains correct. However, this needs verification via pnpm test src/agents/embedded-agent-subscribe*.test.ts in a full build environment.

Compatibility impact: Backward compatible. No config/env changes. No new schema. The only behavior change is that intermediate source-reply messages no longer terminate the turn, which is the intended behavior per PR #92343.

Compatibility / Migration

• Backward compatible? Yes
• Config/env changes? No
• Migration needed? No