claude-code - 💡(How to fix) Fix Agent behavior regressions during multi-step debugging sessions

During a long debugging session the agent exhibited several recurring behavioral patterns that caused wasted user time, incorrect reverts, and at one point a broken system state. Writing these up as a single report since they reinforce each other.

1. Acted on "latest change = most likely cause" without timeline check

When the user reported an intermittent bug, the agent began reverting the most recent edit instead of reconstructing the timeline of when the bug first appeared. One revert was of a setting the user had explicitly requested minutes earlier and that had no plausible causal link to the reported symptom. The user had to point this out directly before the agent corrected course.

2. Did not consult memory/docs that were already loaded in context

Project memory files surfaced at SessionStart contained an explicit prior description of the exact symptom being debugged — including the root cause identified in an earlier session. The agent proceeded to invent fresh hypotheses rather than searching what was already in front of it.

3. Conflated backup-file timestamps with "known-good state"

The agent treated a file labeled with a date months old as equivalent to "the system when it was healthy" and restored it without verifying:

• whether the host OS / toolchain had changed since that date in ways that break old binaries
• whether "pristine" for that component still means "runnable" today
• whether the restored state would even boot

Restoring it caused an immediate runtime failure on launch. The user had to ask "how did you decide which backup to restore?" for the agent to acknowledge it had no real criterion.

4. Confabulated plausible-sounding causal chains

When a revert didn't fix the bug, the agent generated fresh hypothetical mechanisms ("setting X likely adds a pipeline stage that delays input events") with no evidence — just to have the next suggestion ready. Presented unverified reasoning in the same confident tone as verified facts.

5. Technical jargon without checking user context

Used terms like "binary search" with a non-technical user as if shared vocabulary. Only after the user said "I don't understand what you're talking about" did the agent reformulate in plain language. This should have been the default, not a fallback.

What would have helped

• Explicit timeline reconstruction before forming any causal hypothesis — ask when the symptom first appeared, narrow candidate changes to that window
• Check loaded memory/docs for matching symptoms before guessing
• Treat any restore/revert as destructive until proven otherwise: verify the restored artifact at least starts up before telling the user to run a real workload on it
• Flag low-confidence suggestions as such rather than present them identically to verified conclusions
• Default to plain-language explanation with non-expert users; surface the jargon only if they introduce it

These aren't isolated slips — they compounded because each shortcut let the next one happen (skipping memory → needing to guess → confabulating → destructive action based on bad guess). The failure mode that ties them together is skipping verification steps the agent has the tools to do.