Domain 5 — Context Management & Reliability (15%)

Smallest weight, but heavily mixed with the scenarios — appears in nearly every question.

Topics

5.1 The CALM framework

• Cache stable prefixes (system, tools, reference docs).
• Align prefix structure across requests — don't reshuffle.
• Limit history (sliding window, summarization, retrieval).
• Monitor cache hits and token budgets.

5.2 Prompt caching mechanics

• cache_control: {type: "ephemeral"} markers.
• 5-min TTL default (1h tier where available).
• Up to 4 cache breakpoints per request.
• Cache write ≈ 125% input cost; cache read ≈ 10%.
• Pays off after ~2 reuses.
• Any change to cached prefix invalidates it.

5.3 Conversation compaction strategies

Strategy	Best for
Progressive summarization	Long chat assistants
State object	Transactional / workflow agents
Sliding window	Recency-dominant assistants
Retrieval-on-demand	Very long-lived assistants
Subagent delegation	Multi-task agents

5.4 Token estimation & budgets

• ~4 chars / token, ~0.75 words / token.
• Use count_tokens for precision.
• Budget categories: stable prefix, history, working memory, current turn.

5.5 Stale data hazards

• Tool results expire (balance, status, etc.).
• Mark facts with capture timestamps.
• Refresh-on-need beats remember-forever for volatile facts.

5.6 Long-conversation context preservation

• Keep thinking blocks (with signatures) for tool-use threads.
• Keep matching tool_use + tool_result pairs together.
• Summarize old turns into a single assistant message rather than dropping.

5.7 Reliability primitives

• Capped iteration counts in agent loops.
• Structured error responses (is_error: true + type + message).
• Defensive reads after uncertain state.
• Confirmation gates before destructive ops.
• Graceful degradation messages when tools are unavailable.

5.8 Escalation

• "I don't know" / "I can't" are valid agent states.
• Route to human when:
  • Confidence is low.
  • Required tool is unavailable.
  • Iteration cap hit.
  • Policy rule (refund > threshold, etc.) triggers.

5.9 Observability

• Log: request_id, conversation_id, tool call name + args + result, cache hits, token usage.
• Tag every fact in extractions with provenance.
• Audit-friendly: a regulator should be able to trace any decision back to source data.

Common question shapes

• "Conversation costs ballooned in week 3. Likely cause?" → Cached prefix changed; cache misses; or no compaction.
• "Agent retried a payment after timeout and double-charged. Fix?" → Uncertain-state pattern: read-before-retry on non-idempotent ops.
• "Pick the compaction strategy for [long support flow, research project, transactional checkout]." → Match strategy to flow type.
• "Where to put a 50KB reference doc?" → System prompt content block with cache_control.

Cheatsheet

• CALM: Cache, Align, Limit, Monitor.
• Stable prefix; volatile suffix.
• Idempotent retries; defensive reads otherwise.
• Escalate > hallucinate.

See also

• context-management
• multi-turn-conversations
• prompt-caching
• error-handling-in-tools