PR #65778: fix(agent): preserve original prompt in fallback retry instead of replacing it

• Repository: openclaw/openclaw
• Author: zhouhe-xydt
• State: open | merged: False
• Link: https://github.com/openclaw/openclaw/pull/65778

Description (problem / solution / changelog)

Summary

When a model call fails and OpenClaw triggers a fallback retry, resolveFallbackRetryPrompt() was replacing the entire original user prompt with the fixed string "Continue where you left off. The previous model attempt failed or timed out.". This caused the agent to lose the original task context and potentially produce incorrect or unintended results.

This change prepends the retry notice to the original prompt so the agent retains full context of what it was asked to do.

Change Type

• Bug fix

Related Issue

Fixes #65760

Testing

• pnpm test src/agents/command/attempt-execution.test.ts — 17/17 passed
• pnpm tsgo — passed
• pnpm lint — 0 warnings, 0 errors

Changed files

• src/agents/command/attempt-execution.helpers.ts (modified, +1/-1)
• src/agents/command/attempt-execution.test.ts (modified, +3/-1)

PR #65805: fix(fallback): preserve original prompt in resolveFallbackRetryPrompt

• Repository: openclaw/openclaw
• Author: Bartok9
• State: open | merged: False
• Link: https://github.com/openclaw/openclaw/pull/65805

Description (problem / solution / changelog)

Summary

When a model call fails and triggers a fallback retry on a session with existing history, resolveFallbackRetryPrompt() replaces the original user prompt with the generic string "Continue where you left off. The previous model attempt failed or timed out.". This causes the fallback model to lose the original task instruction entirely, forcing it to infer the task from session history alone — which can produce incorrect or unintended results.

Root Cause

// Before: original body is discarded
if (params.sessionHasHistory) {
  return "Continue where you left off. The previous model attempt failed or timed out.";
}

When isFallbackRetry=true and sessionHasHistory=true, the function returned only the generic retry message, discarding params.body.

Fix

Prepend the retry notice to the original body instead of replacing it:

const retryNotice =
  "[System: The previous model attempt failed or timed out. Continue where you left off.]";
return \`\${retryNotice}\\n\\n\${params.body}\`;

The fallback model now receives both:

1. Context that a retry occurred (so it can leverage session history)
2. The full original task instruction (so it knows what to do)

Test Changes

Updated the existing test to verify:

• The original body is preserved in the output
• The retry notice appears as a prefix
• The original body follows after the notice

All 17 existing tests pass.

Impact

• All users with model fallback/retry configured benefit
• Agents no longer silently lose their task when a model times out
• Subagent spawns are most affected since their task instruction is their only context

Fixes #65760

Changed files

• src/agents/command/attempt-execution.helpers.ts (modified, +6/-1)
• src/agents/command/attempt-execution.test.ts (modified, +12/-8)

PR #65885: fix: preserve original prompt in resolveFallbackRetryPrompt

• Repository: openclaw/openclaw
• Author: mushuiyu886
• State: closed | merged: False
• Link: https://github.com/openclaw/openclaw/pull/65885

Description (problem / solution / changelog)

Summary

Fixes #65760

Changes

• Fix resolveFallbackRetryPrompt() to preserve the original user prompt instead of discarding it on model fallback retry
• Prepend a retry notice [Retry after previous model attempt failed or timed out.] to the original prompt body
• Update corresponding test expectations

Details

Previously, when a model call failed and triggered a fallback retry, resolveFallbackRetryPrompt() replaced the entire original prompt with the generic message "Continue where you left off. The previous model attempt failed or timed out.". This caused the agent to lose all task context.

Now the function prepends a retry notice to the original prompt, preserving full task context while informing the model that a retry occurred.

Before

return "Continue where you left off. The previous model attempt failed or timed out.";

After

return `[Retry after previous model attempt failed or timed out.]\n\n${params.body}`;

Test plan

• Existing unit test updated to verify new behavior
• All pre-commit checks pass (lint, type-check, import cycle checks)
• Verify that fallback retry preserves original prompt in a live session

Changed files

• src/agents/command/attempt-execution.helpers.ts (modified, +1/-1)
• src/agents/command/attempt-execution.test.ts (modified, +2/-2)

PR #66029: fix(agents): preserve original prompt on model fallback retry (#65760)

• Repository: openclaw/openclaw
• Author: WuKongAI-CMU
• State: open | merged: False
• Link: https://github.com/openclaw/openclaw/pull/66029

Description (problem / solution / changelog)

Summary

When a model call fails and OpenClaw triggers a fallback retry, `resolveFallbackRetryPrompt()` replaced the entire original user body with the fixed string `"Continue where you left off. The previous model attempt failed or timed out."` This discarded the original task instruction entirely, forcing the fallback model to reconstruct the task from session history alone — sometimes impossible (e.g. freshly spawned subagent where only the last message survives), sometimes just wrong.

Prepend the retry context to the original body instead so the fallback model has both the recovery signal AND the task:

``` [Retry after the previous model attempt failed or timed out]

The existing subagent-spawn branch (no persisted session history) is unaffected — that path still returns `params.body` verbatim.

Credit to #65760 for a clean root-cause trace (incl. pointer to `attempt-execution.helpers.ts:resolveFallbackRetryPrompt`).

Closes #65760.

Changes

• `src/agents/command/attempt-execution.helpers.ts` — prepend retry prefix to `params.body` instead of replacing it
• `src/agents/command/attempt-execution.test.ts` — update the matching test to assert the new prefixed form; leave other tests (first attempt, no-history subagent) untouched

Test plan

• Updated test exercises the fallback-with-history path directly
• No-history subagent test still asserts `originalBody` return
• First-attempt test still asserts `originalBody` return
• Minimal diff, no surrounding refactor

🤖 Generated with Claude Code

Changed files

• src/agents/command/attempt-execution.helpers.ts (modified, +7/-1)
• src/agents/command/attempt-execution.test.ts (modified, +2/-2)