fix: completionPrompt continues existing session instead of recursive answer()

[buger]

Summary

• Bug: completionPrompt called this.answer() recursively, which created a fresh TaskManager and iteration counter, wiping all task context from the main turn. The follow-up turn had no memory of what Phase 1 did — it saw already-modified files, concluded "no changes needed", and returned an empty result that overwrote the successful structured output (pr_urls, files_changed, summary).
• Fix: Replace the recursive this.answer() with a single streamText continuation — append the completion prompt as a user message to currentMessages and run one more AI pass with the same tools, TaskManager, and conversation history. Falls back to the original result if the follow-up produces nothing useful.

What changed

Before	After
this.answer() — full new agentic loop	streamText() — one more turn in same session
Fresh TaskManager — tasks wiped	Same TaskManager — tasks preserved
Fresh iteration counter (up to max_iterations)	Capped at 5 extra iterations
finalResult = completionResult unconditionally	Keeps original if follow-up is empty

Test plan

• All 2631 existing tests pass (106 suites)
• completion-prompt unit tests pass (26 tests)
• Manual test: run engineer workflow with completion_prompt configured — verify PR URLs are preserved in output

🤖 Generated with Claude Code

[probelabs]

I was unable to complete your request due to reaching the maximum number of tool iterations.

---

Powered by Visor from Probelabs

Last updated: 2026-03-04T15:05:43.336Z | Triggered by: pr_updated | Commit: 0a50d65

💡 TIP: You can chat with Visor using /visor ask <your question>

[probelabs]

Architecture Issues (1)

Severity	Location	Issue
🟠 Error	contract:0	Output schema validation failed: must have required property 'issues'

\n\n

Architecture Issues (1)

Severity	Location	Issue
🟠 Error	contract:0	Output schema validation failed: must have required property 'issues'

\n\n ### Performance Issues (5)

Severity	Location	Issue
🟡 Warning	npm/src/agent/ProbeAgent.js:3620-3622	The completion prompt follow-up mutates currentMessages by pushing the user message directly, then passes it to prepareMessagesWithImages which creates a shallow copy. This results in O(n) array copying where n is the total message count, happening twice (once in the main loop, once in completion prompt). For long conversations, this creates unnecessary memory pressure. 💡 Suggestion Consider batching the completion prompt message addition with the existing messages array construction to avoid multiple array copies. The current approach is acceptable for typical conversation lengths but could be optimized for very long sessions.
🟡 Warning	npm/src/agent/ProbeAgent.js:3626-3627	The code resets completionResult = null and completionAttempted = false before the follow-up turn. While this is correct for the new logic, it means the completion tracking state is being mutated in-place. If the follow-up streamText call fails or times out, these reset values persist, though the code correctly falls back to originalResult. 💡 Suggestion The current implementation is correct - the fallback to originalResult handles failure cases. Consider adding a telemetry event to track when the fallback is triggered for monitoring purposes.
🟡 Warning	npm/src/agent/ProbeAgent.js:3630-3648	The completionStreamOptions object is constructed with similar properties to the main streamOptions (model, messages, tools, maxTokens, temperature, onStepFinish). This creates a partial duplication of the options construction logic. While not a runtime performance issue, it increases code maintenance burden. 💡 Suggestion Consider extracting common streamText options into a helper method or spreading from a base options object to reduce duplication. This is a minor code quality improvement, not a performance concern.
🟡 Warning	npm/src/agent/ProbeAgent.js:3656-3661	The code iterates over cpResult.response?.messages and pushes each to currentMessages individually. For responses with many messages, this is O(m) where m is message count. While typically small, using spread operator would be more idiomatic: currentMessages.push(...cpMessages). 💡 Suggestion Replace the for-of loop with spread: if (cpMessages?.length) currentMessages.push(...cpMessages);
🟡 Warning	npm/src/agent/ProbeAgent.js:3631	The completion prompt follow-up uses a hardcoded max of 5 iterations (completionMaxIterations = 5). This is a reasonable limit but is not configurable. For complex follow-up tasks, this might be insufficient, while for simple validations it may be excessive. 💡 Suggestion Consider making the completion prompt iteration limit configurable via a new option like completionPromptMaxIterations, defaulting to 5. This allows tuning for different use cases without code changes.

Quality Issues (1)

Severity	Location	Issue
🟠 Error	contract:0	Output schema validation failed: must have required property 'issues'

---

Powered by Visor from Probelabs

Last updated: 2026-03-04T15:01:12.256Z | Triggered by: pr_updated | Commit: 0a50d65

💡 TIP: You can chat with Visor using /visor ask <your question>