Context Management & Reliability

1. 5.1Conversation context managementKeep exact facts alive in long conversations with a protected case facts block, trimmed tool output and position-aware ordering.
2. 5.2Escalation and ambiguity resolutionEscalate on explicit requests, policy gaps and lack of progress, not on mood or self-reported confidence, and ask for identifiers on multi-match lookups.
3. 5.3Error propagation across agentsSubagents retry transient failures locally and return structured error context (type, attempt, partials, alternatives) so the coordinator can recover and the report shows gaps.
4. 5.4Context in large codebasesCounter context degradation with scratchpad files, subagent delegation, phase summaries, /compact and manifest-based crash recovery.
5. 5.5Human review and calibrationMeasure accuracy by segment, calibrate field-level confidence on labeled data, route low-confidence cases to people and keep sampling the automated rest.
6. 5.6Provenance and uncertaintyCarry claim-source mappings and dates through synthesis, annotate conflicts instead of picking one, and separate established from contested findings.

• Keep exact facts (amounts, IDs, dates) in a protected block that is never summarized away.
• Trim verbose tool output down to the fields that matter before it enters the conversation.
• Escalate when the customer asks for a person, when policy doesn't cover the case, or when you can't make progress.
• When a lookup matches several records, ask for an identifier instead of guessing.
• Let subagents retry transient failures themselves, then report what failed, what was tried and any partial results.
• In long sessions, save findings to scratchpad files and delegate noisy exploration to subagents.
• Measure accuracy per document type and field before cutting human review, and keep sampling the automated cases.
• Keep claim-to-source links and dates through synthesis, and show conflicting sources side by side.

• Escalating based on the customer's mood or the model's own confidence score.
• Extending a policy by analogy instead of escalating the gap.
• Finishing with a clean-looking report that hides a failed source.
• Fixing context problems by resending everything, or by summarizing harder.
• Choosing one of two conflicting figures without saying they conflict.

You get 4 of the 6 scenarios at random, so practice this domain in each of them. All six scenarios.

1. Run a long support conversation and check that amounts and order numbers survive to the end.
2. Write explicit escalation criteria with examples and test them against edge cases.
3. Make a subagent fail partway and check that the coordinator's report shows the gap.
4. Build a report where two sources disagree and show both values with their dates.