Code Verification Procedure

When writing or auditing documentation, verify every code example against the actual SDK source. Plausible-but-wrong code erodes developer trust faster than missing documentation.

The docs site lives under site/ in this repo. All paths in this guide — including those in skill procedures that reference it — are written relative to the repo root, so they include the site/ prefix where applicable. Run shell commands like npm run sdk:clone from inside site/ (where package.json lives).

Use the first tier available in your environment. Fall back only when the preferred tier is unavailable.

Tier 1: Local SDK Clones (preferred)

If you have the local clones populated by npm run sdk:clone, read directly from disk — it's faster than network calls and avoids API rate limits:

site/.build/sdk-python/src/strands/[path]
site/.build/sdk-typescript/strands-ts/src/[path]

If the clones aren't present, run npm run sdk:clone from site/, or fall back to Tier 2.

Key Python modules (under src/strands/):

• agent/agent.py — Agent class, constructor signature, __call__/stream_async
• agent/agent_result.py — AgentResult shape (final response, metrics, stop reason)
• agent/conversation_manager/ — ConversationManager, SlidingWindowConversationManager, SummarizingConversationManager
• agent/state.py — AgentState (per-agent key/value store)
• tools/decorator.py — @tool decorator and signature inference
• tools/tool_provider.py — ToolProvider base for grouped tools
• tools/mcp/ — MCP client/server tool integration
• tools/structured_output/ — Structured-output tool helpers
• hooks/registry.py — HookRegistry, hook callback registration
• hooks/events.py — Hook event types (BeforeInvocationEvent, AfterInvocationEvent, etc.)
• models/ — Model provider implementations (Bedrock, Anthropic, OpenAI, etc.)
• multiagent/graph.py, multiagent/swarm.py, multiagent/a2a/ — Multi-agent orchestration
• session/ — SessionManager, FileSessionManager, S3SessionManager, RepositorySessionManager
• event_loop/ — Core event loop and streaming
• telemetry/ — Tracing, metrics, OTEL integration
• interrupt.py, types/interrupt.py — Interrupt/resume primitives
• types/exceptions.py — Error classes
• types/content.py, types/tools.py, types/streaming.py — Public message/tool/stream shapes

Key TypeScript modules (under strands-ts/src/):

• agent/ — Agent class, agent result, configuration
• tools/ — tool() factory, tool registration, MCP tool support
• conversation-manager/ — ConversationManager implementations
• hooks/ — Hook registry and event types
• models/ — Model provider implementations
• multiagent/ — Graph, swarm, agent-to-agent orchestration
• session/ — Session managers and persistence
• plugins/, vended-plugins/ — Plugin system and built-in plugins
• interventions/, vended-interventions/ — Intervention hooks
• retry/ — Retry policies
• telemetry/ — Tracing and metrics
• errors.ts — Error classes
• types/ — Public type definitions (content, tools, streaming, etc.)
• index.ts — Public exports (canonical entry point for what's exported)

TypeScript supporting paths:

• strands-ts/test/ and strands-ts/src/__tests__/ — Unit tests (good for usage patterns)
• strands-ts/examples/ — Runnable examples

What to verify:

• Import paths resolve to real modules
• Parameter names, types, and defaults match class definitions
• Referenced methods exist on the classes shown
• For pages with both languages: verify each independently

Tier 2: GitHub API

If the local clones aren't available, fetch source through the GitHub API:

# Python SDK (now local at strands-py/src/strands/)
gh api repos/strands-agents/harness-sdk/contents/strands-py/src/strands/[path]

# TypeScript SDK (now local at strands-ts/src/)
gh api repos/strands-agents/harness-sdk/contents/strands-ts/src/[path]

Tier 3: Installed Package Introspection

# Python
python -c "from strands.agent.agent import Agent; help(Agent.__init__)"

# TypeScript (check exports)
node -e "const sdk = require('@strands-agents/sdk'); console.log(Object.keys(sdk))"

Useful for spot-checking parameter names and types, but doesn't show internal module structure.

Tier 4: Stop and Surface

If none of the tiers above are available, do not ship unverified code. Reviewers are not the verification step, and unverified blocks must not reach a PR description as a TODO list for someone else to resolve.

Instead, surface the gap in the output of the calling skill:

• From docs-audit: list each unverifiable item under ### Accuracy Issues (e.g., "Agent.stream_async signature — no SDK source available to verify") and add a corresponding ### Recommended Actions entry to acquire a verification source before re-auditing.
• From docs-writer: omit the affected code block from the draft and report the unverifiable item alongside the finished page — naming the specific class/method and the tier that failed. Do not emit a guessed example.

The expectation is that the next iteration of the skill resolves the gap, not that a reviewer takes it on.