Summary

• Failure-mode-aware image handling: Replace unconditional image stripping on fallback retry with resolveRetryImages — only strip on format errors (model can't handle images) or cross-provider retries (privacy boundary)
• Cross-provider image privacy: Strip user-supplied images when fallback targets a different provider, preventing unintended third-party disclosure
• Prompt-detected image bypass fix: Add suppressPromptImageDetection to the embedded runner so detectAndLoadPromptImages is skipped on cross-provider retries — closes the gap where prompt-referenced local images bypassed the image strip
• Partial tool execution detection: Thread previousFailureReason and previousPartialExecution through ModelFallbackRunOptions so the retry callback knows what the prior attempt did; warn fallback models about already-executed tools to prevent replaying non-idempotent operations

Builds on openclaw#55632 (prompt preservation for new sessions). Supersedes openclaw#44188. Closes openclaw#43481, openclaw#43492.

Design boundary

All changes stay at the runAgentAttempt / runWithModelFallback boundary. The pi-runner internals are not modified beyond accepting the suppressPromptImageDetection flag. This is intentional — the runner's internal loops (profile rotation, cooldown skips) are treated as opaque. See openclaw#44188 for the full design discussion.

Changes

failover-error.ts

• New PartialExecution type with toolNames and didSendViaMessagingTool
• FailoverError carries optional partialExecution field
• sanitizeToolNames static method (caps at 20 entries, strips non-alphanumeric, truncates to 100 chars)
• describeFailoverError returns partialExecution when present

model-fallback.ts

• ModelFallbackRunOptions extended with previousFailureReason and previousPartialExecution
• Fallback loop accumulates failure reason and merges partial execution across attempts (union tool names, OR messaging flag)
• Threaded into runOptions for next candidate

attempt-execution.ts

• resolveRetryImages: failure-mode-aware image handling (format → strip, cross-provider → strip, same-provider non-format → preserve)
• buildPartialExecutionSystemContext: system prompt warning about already-executed tools (same-provider only, CWE-200)
• runAgentAttempt extended with primaryProvider, previousFailureReason, previousPartialExecution
• Both CLI and embedded paths use resolveRetryImages and effectiveExtraSystemPrompt

Embedded runner (params.ts, run.ts, attempt.ts)

• New suppressPromptImageDetection param gates detectAndLoadPromptImages

agent-command.ts

• Passes primaryProvider, previousFailureReason, previousPartialExecution from fallback runOptions into runAgentAttempt

Test plan

• resolveRetryImages — 6 tests: first attempt, format error, rate_limit, cross-provider, undefined provider, undefined images
• buildPartialExecutionSystemContext — 3 tests: tool listing, messaging warning, empty tools
• FailoverError.sanitizeToolNames — 3 tests: cap at 20, truncate to 100, filter empty
• FailoverError.partialExecution — 2 tests: carries when provided, undefined when absent
• describeFailoverError — 2 tests: includes/omits partialExecution
• runWithModelFallback — 4 tests: threads previousFailureReason, threads previousPartialExecution, merges across attempts, omits when absent
• pnpm check — clean (0 warnings, 0 errors)
• pnpm build — clean (no type errors, no INEFFECTIVE_DYNAMIC_IMPORT)

🤖 Generated with Claude Code