Claude Code clean-room documentation snapshot
This snapshot is strong on product contracts and weak on implementation visibility. Every claim comes from official documentation observed on 2026-08-10. Public distribution tag v2.1.226 aligns the time boundary; it is not product-core source.
1. Loop — unknown
Official material establishes an agentic terminal workflow, but this snapshot does not assert an internal sampling loop, tool-result continuation mechanism, termination enum, or retry scheduler. Future work can describe external behavior only through official structured output or authorized black-box experiments.
2. Context — partial
Subagent documentation establishes separate child context windows and describes auto-compaction, resume, and startup loading. It does not publish the main context reducer algorithm, summary prompt, or persistence format; those remain unknown.
3. Tools — partial
Documentation publishes permission-rule syntax, subagent tool allow/deny controls, and MCP scope. That supports a capability contract, not a claim about internal registry, validator, or dispatcher code.
4. Safety — documented product contract
Permissions decide whether tools, files, and domains may be accessed; sandboxing provides OS-level filesystem/network enforcement for Bash and child processes. Official guidance recommends both rather than treating one as a substitute. Deny, ask, managed precedence, and sandbox details remain bound to the documentation date. Claim: claude-code-doc-execution-policy.
5. Reliability — unknown
Official error and troubleshooting pages exist, but this project has no version-pinned experiments for retry, partial tool calls, crash recovery, or resume. Internal policy is not asserted.
6. Extensibility — documented product contract
Hooks cover lifecycle points including PreToolUse, PostToolUse, PermissionRequest, SessionStart/End, and Stop. A PreToolUse deny can block before permission-mode checks, while allow cannot override stricter deny rules. Concurrent hooks and merge precedence are experiment-worthy failure surfaces. Claim: claude-code-doc-hooks.
7. Orchestration — documented product contract
Subagents have separate context, prompts, tools, models, permission modes, hooks, skills, turn/depth limits, and optional worktree isolation. Documentation also distinguishes foreground/background execution and resume. Claim: claude-code-doc-subagents.
Agent Teams is a separate experimental contract: a lead coordinates teammates with independent context windows, a shared task list, and direct messages between teammates. Unlike isolated subagent delegation, team members can coordinate laterally; unlike an explicit worktree mode, teammates do not receive automatic workspace isolation. The documentation version observed here also describes direct teammate spawning without the earlier setup-tool lifecycle. This snapshot records the public contract without inferring an internal scheduler. Claim: claude-code-doc-agent-teams.
8. Interfaces — partial
Known surfaces include CLI/TUI, IDE integration, and Agent SDK. No version-pinned protocol claim exists yet, so UI behavior is not presented as a stable wire contract.
9. Observability — unknown
Transcripts, hook JSON, and programmatic output expose observable events, but this project has not captured a native trace and does not know the authoritative internal event store. The page shows an evidence ledger rather than simulating an internal event bus.
Completing Tier B well
Closed-source research does not wait for source; it uses stricter boundaries. Every claim binds to an official page and capture date, dynamic documentation changes trigger staleness, experiments record only public output, hidden prompts/chain-of-thought/private telemetry are excluded, and explanations of internal causes are labeled as inference.