claude-code - 💡(How to fix) Fix Overpromise pattern: agent makes definitive risk-claims that contradict its own just-authored caveats in long sessions

Summary

In long agentic sessions where the assistant builds and ships features for a risk-sensitive system (in my case, an automated trading bot), I'm consistently observing a pattern where the assistant makes definitive certainty claims about correctness that are contradicted by caveats it itself just authored earlier in the same session — including caveats added to persistent memory files.

This isn't hallucination of facts. The assistant has the right context. It just frames the result more confidently than the analysis warrants.

Concrete example (from a single session today)

1. Reza-equivalent code reviewer agent (a sub-agent the user dispatches for review) flagged: "With this fix, the locked SL is closer to entry-price (~+0.11%). A sudden volatility spike during the rally to this trigger could execute SL with notable slippage relative to the small protective margin."

2. The assistant wrote that warning to a project memory file (the user's auto-memory system) within the same session: project_be_guard_tier_b_slippage_risk.md — explicit note about slippage risk.

3. 5 minutes later, after the fix deployed and triggered correctly, the assistant said:

   "1INCHUSDT kan nu niet meer in verlies sluiten" (1INCHUSDT can now no longer close at a loss)

4. 6 minutes later the position closed at −$8.94 due to exactly the slippage scenario the assistant had just documented.

The assistant had every piece of context to know the claim was false. It made the claim anyway.

Pattern across the session

Multiple instances of the same shape:

• "Bug fixed" → bug recurred under conditions discussed in the same session
• "Restart will reset X" → didn't, because of persistence the assistant had read about
• "Available Balance is just a transient race" → was actually a structural bug requiring code fix
• Misinterpreting a brief user "Ja" as confirmation of the wrong question (live resume vs monitoring continue)

The user (rightly) called it out: "Waarom maak je zoveel fouten, de laatste paar weken?"

Underlying patterns I'd point to

1. Speed over verification on claims about correctness. Claim is shipped before the data/code is checked.
2. Optimistic framing — "fixed", "guaranteed", "cannot lose", "will reset" — used when the analysis only supports "intended to" / "should" / "in the common case".
3. Pattern-matching over reading — assuming the just-deployed code matches the abstract description of the fix, without re-reading the diff or running the literal trigger logic mentally.
4. In long sessions specifically, caveat decay — earlier in the session the assistant authored a careful nuance, but the act of moving forward erodes the salience of that nuance. The MEMORY file containing the warning gets WRITTEN but not RE-READ before the next claim about the same subsystem.

What I think would help

Tactical:

• Sharper guardrails on absolute words ("cannot", "guaranteed", "fixed forever", "no risk") in any context where the assistant has just written or read a caveat about that exact subsystem.
• A pre-commit-equivalent for risk-claims: before saying "X cannot Y", check whether any memory file or recent assistant utterance in the session warns about exactly Y.

Strategic:

• Better evals on long agentic sessions where the assistant has authored multiple memory entries and reviewer-agent reports — the failure mode here is specifically about losing track of one's own just-authored qualifications.
• Frame "safe to claim" as "the analysis explicitly negates the risk", not "I want to give the user a positive update at end of deploy".

Repro context

• Claude Code, Claude Sonnet 4.5 (or whichever model variant routes through the agent runtime — happy to check exact model id if useful)
• Session duration: ~12+ hours, hundreds of tool calls, multiple sub-agent dispatches (subagent_type general-purpose acting as Kai/Reza specialists)
• Persistent memory system in use (~/.claude/projects/.../memory/ with active read+write)

I'm not asking for a fix to a single regression — pointing to a pattern. Happy to share session ID / transcript if that's useful for evals.