Implement #7: Advanced Agent Patterns

[yagop]

Closes #7

Chapter 7 adds the advanced patterns on top of the Chapter 5/6 loop: a budgeted autonomous loop, extended thinking, self-critique, structured output, error recovery, and subagent orchestration. Authored single-pass; the cutoff-sensitive thinking shape was verified against the installed SDK and a live probe.

Definition of done

• Chapter at chapters/07-advanced-patterns.md, <=120 lines, <=4 main-line H2s plus a closer
• Every sample runnable with bun run, imported via <<< @/examples/07-advanced-patterns/file.ts
• One-home rule held (the reflection-as-user-turn gotcha and the thinking shape live in prose, not code comments); reader addressed as "you"; intro + sections open warm
• Samples use only real @anthropic-ai/sdk surface; ASCII punctuation only
• Optional material (token guardrails, planning tool) lives in Going-deeper :::details asides, not main-line H2s
• Linked from README.md and the .vitepress/config.ts sidebar; bun x vitepress build passes
• thinking parameter shape verified - see below
• Sample line budget: samples are 24-66 lines, above the issue's stated <=35 (skill budget is <=100)

thinking shape verification (cutoff-sensitive)

Checked against the installed @anthropic-ai/sdk .d.ts and a live probe on the gateway:

• ThinkingConfigParam = { type: 'adaptive', display? } | { type: 'enabled', budget_tokens, display? } | { type: 'disabled' }
• ThinkingBlock = { signature: string; thinking: string; type: 'thinking' }
• Live probe: thinking: { type: 'adaptive' } -> returned [thinking, text] blocks (2803 thinking chars, signature present); { type: 'enabled', budget_tokens: 1024 } also works.
• The sample uses adaptive (current form); an :::info callout documents the enabled/budget_tokens form for models 4.6 and earlier.

Verification (run live)

Chapter budget: wc -l -> 69; grep -c '^## ' -> 4. tsc --noEmit clean; vitepress build passes (all six snippets resolve). Each sample run:

autonomous-loop.ts    46   extended-thinking.ts  26   reflection.ts    24
structured-output.ts  38   error-recovery.ts     66   orchestrator.ts  53

• autonomous-loop -> 2 steps (tool_use -> end_turn), prints running token total, answers correctly.
• extended-thinking -> a real [thinking] block (the bat-and-ball riddle reasoning) printed separately from [answer].
• reflection -> draft -> plain-user critique turn -> revised tagline.
• structured-output -> forced emit_result returned label: positive, confidence: 0.95 (the gateway honored forcing for this tool-eliciting prompt).
• error-recovery -> read notes.txt (ok) and config.txt (is_error -> missing); the model recovered and reported both.
• orchestrator -> coordinator's summarize_text call dispatched to the summarize() helper subagent (its own create), returning the one-sentence summary.

Notes

• Only extended-thinking.ts enables thinking, keeping it clear of the forced-tool_choice and thinking-block-preservation constraints that apply when thinking is combined with tools.
• Sample sizes (24-66 lines) follow the skill's 100-line budget rather than the issue's older <=35.

Review flags

None.

🤖 Generated with Claude Code