CCDV-F Key Terms — Flashcards
Forty-plus term-and-definition cards across all eight domains — screenshot, print, or drill front-to-back before the exam.
Each card names a term on top and the exam-relevant definition below. Cover the bottom half and quiz yourself before flipping.
Applications & Integration, Model Selection
stop_reason
Why generation ended: end_turn, max_tokens, stop_sequence, tool_use, pause_turn, or refusal — each needs a different fix.
Idempotency (Messages API)
Not built in — no idempotency-key mechanism exists. A retried, side-effecting tool call can duplicate; the app must dedup it.
Spend-cap 429
A special rate-limit case: no retry-after, won't clear until the next billing cycle — distinct from an ordinary token-bucket 429.
Message Batches API
Async, high-volume processing at a higher output-token ceiling (up to 300K beta vs. 128K sync) — for work nothing is waiting on in real time.
Workload Identity Federation
Exchanges a cloud IdP JWT for short-lived Claude API credentials — no static secret, built for CI/production.
App Attest
Hourly-expiring, Messages-API-scoped tokens for iOS/macOS apps calling the API directly from-device.
Prompt caching prefix order
Must be tools → system → messages, breakpoint on the last stable block — reordering silently breaks the cache.
Minimum cacheable prefix
Varies by model (as low as 512 tokens on the newest tier, up to 4,096 on older models) — a short prompt can silently fail to cache.
Adaptive thinking
thinking: {type: "adaptive"} + output_config.effort on Fable 5/Opus 5/Sonnet 5 — replaces the older budget_tokens mechanism, which only Haiku 4.5 still uses.
Sampling-parameter rejection
temperature/top_p/top_k return a 400 on the newest model tier (Opus 5, Fable 5, Sonnet 5) — accepted on older tiers.
Agents & Workflows, Tools & MCPs
Workflow vs. agent
Workflow: your code owns a predefined control-flow path. Agent: the model dynamically directs its own tool usage — reserve for genuinely open-ended tasks.
Orchestrator-workers
A central LLM dynamically delegates unpredictable subtasks to worker LLMs — one of the five named workflow patterns.
Subagent isolation
A subagent gets a fresh, isolated context by default — it does NOT inherit the parent conversation's history unless explicitly passed in the task description.
Claude Agent SDK
Packages the full Claude Code harness (tools, agent loop, hooks, subagents, MCP) as a library — Python/TypeScript officially, other languages shell out to the CLI. Still self-hosted.
Managed Agents
Like the Agent SDK, but Anthropic also hosts the sandbox — the one tier that removes hosting responsibility entirely.
tool_result / is_error
The structured signal for a failed tool call — a business-logic failure distinct from HTTP-transport success.
strict: true (tool definition)
Guarantees the returned input is validated exactly against the tool's JSON Schema, eliminating a class of argument-parsing bugs.
MCP primitives
Tools (model-controlled actions), Resources (app-controlled read-only data), Prompts (user-controlled templates).
Streamable HTTP
Current recommended MCP transport for remote/multi-client servers — supersedes the older HTTP+SSE transport.
CIMD
Client ID Metadata Documents — current recommended MCP auth-registration flow, superseding Dynamic Client Registration (DCR still works for backward compatibility).
Prompt/Context Engineering, Security, Claude Code, Debugging
Context rot
Quality drift over a long session as irrelevant accumulated context crowds out what matters — fixed by compaction, external notes, or sub-agent isolation of noisy tool output.
Syntactic vs. semantic correctness
Valid JSON output is not the same guarantee as factually/semantically correct output — defensive parsing is required regardless of how output was requested.
Direct vs. indirect injection
Direct: the end user is the adversary (mitigate with input validation/classifiers). Indirect: trusted user, untrusted fetched content (mitigate by confining it to labeled tool_result blocks).
Prompted instruction ≠ security control
"The model was told not to X" and "the app technically prevents X" are different guarantees — deterministic enforcement must sit outside the model.
bypassPermissions
Claude Code mode intended for containers/VMs as non-root — not general workstation use.
auto permission mode
Built-in default on paid Claude Code plans as of mid-2026 — a second classifier model reviews most actions instead of asking a human each time.
Settings precedence
Managed → CLI flags → project local (settings.local.json) → shared project (settings.json) → user settings — highest wins.
Skills
`.claude/skills/<name>/SKILL.md` — custom slash commands were merged into Skills; bodies load into context on demand, unlike CLAUDE.md which is resident every turn.
Four-layer failure isolation
API-layer (status codes), tool-execution (is_error), prompt/context (model correct on bad input), or genuine model-output — the rarest, most over-diagnosed category.
max_tokens truncation vs. reasoning failure
stop_reason: "max_tokens" means the budget ran out — fix by raising the budget, not by rewriting the prompt as if the model got confused.
How to use these
Drill the cards you keep flipping wrong into the corresponding domain guide's section — a flashcard tells you what you don't know, the guide tells you why.
Independent, unofficial study material from Claude Cert Prep. Not affiliated with Anthropic.