Skip to content

claude-reasoning-effort.md

Latest commit

 

History

History
53 lines (35 loc) · 6.45 KB

claude-reasoning-effort.md

File metadata and controls

53 lines (35 loc) · 6.45 KB
Raw

Governed thinking-effort qualification

This document is the reviewed capability record for LemmaComputer's first qualified route and agent adapter; it is not a requirement that other adapters use Anthropic. The product control is exposed only when both the selected agent runtime and governed model route match reviewed qualifications. Unknown or discovery-only agents, providers, models, or client versions fail closed and do not expose an effort selector. The provider-neutral cross-runtime and route-registration contract is defined in Reasoning adapter qualification.

Route qualification ID: anthropic-claude-4.6-4.8-effort-route-2026-08-13

Agent-adapter qualification ID: claude-cli-2.1.215-governed-effort-adapter-2026-08-13

Qualified model routes

Provider route Provider model Product levels Provider mechanism Status
Direct Anthropic claude-sonnet-4-6 Low, Medium, High Adaptive thinking plus output_config.effort Qualified by documented contract
Direct Anthropic claude-opus-4-8 Low, Medium, High Adaptive thinking plus output_config.effort Qualified by documented contract
Amazon Bedrock Current Sonnet 4.5 profile None Not assumed equivalent to the direct Anthropic effort contract Unsupported; fail closed
Any other provider or model Any None Unknown Unsupported; fail closed

Qualified agent adapters

Agent runtime Pinned version Product levels Status
Claude CLI 2.1.215 Auto, Low, Medium, High Qualified for signed, conversation-pinned propagation
Hermes Agent CLI 0.19.0 Low, Medium, High Qualified separately under hermes-claw-0.19.0-governed-effort-adapter-2026-08-13; route capability still intersects independently
Hermes Desktop 0.17.0 Low, Medium, High Qualified separately under hermes-desktop-0.17.0-governed-effort-adapter-2026-08-13; unsupported upstream levels are hidden on LemmaComputer routes
Codex CLI 0.144.4 None Discovery recorded; credentialed reasoning-plus-tool evidence remains incomplete; fail closed
Any other runtime or version Any None No reviewed registration; fail closed

Auto is a LemmaComputer product setting, not a provider effort value. Control resolves it to the organization maximum (low, medium, or high) before route selection. The protected maximum max is deliberately clipped to product high; xhigh and max are not user-selectable in this phase.

This is Auto thinking effort, not Auto model mode. Employee Web Chat exposes only the explicit Lite, Balanced, and Pro model tiers from #68.

The qualification is based on Anthropic's current documented contracts for adaptive thinking, effort, extended thinking, the Messages API, and the Claude Code CLI. A provider or Claude Code upgrade must add a new reviewed qualification ID rather than widening the existing match.

Enforcement path

  1. Routing administration derives route capability from the provider and model. A separate code-owned registry derives agent-adapter capability from the agent catalog ID and pinned client version. Administrator-supplied capability claims are ignored.
  2. Control intersects route capability, agent-adapter capability, the employee's Team route, and the protected organization maximum. Only that intersection reaches the UI.
  3. The conversation stores the selected product effort. Once created, the selector is locked; a different effort starts a new conversation so prompt-cache behavior remains predictable.
  4. Control signs the requested effort and maximum into the AI task binding. The workspace gateway removes client-supplied thinking, output_config, reasoning, and enabled reasoning_effort fields. The only client-side value it may preserve is the non-escalating Chat Completions opt-out reasoning_effort: none for unqualified tool-capable runtimes.
  5. After governed route resolution, the LiteLLM callback injects only the resolved low, medium, or high value. The route must advertise that exact capability or routing fails closed.
  6. Usage admission records both requested and resolved effort. Provider-reported reasoning tokens remain a separate reasoning_token usage bucket; hidden reasoning text is never stored or exposed.

Adapter extension contract

Adding an agent does not add a new Web or Control conditional. Exact runtime pins may first land as discovery records, which remain ineligible for product controls. Promotion adds a reviewed qualified registration containing the exact catalog ID, pinned runtime version, supported product levels, and completed runtime evidence. Provider/model route qualifications remain separate because an agent transport and a model route can support different subsets. A level is exposed only when both sides include it.

Auto, Low, Medium, and High are stable LemmaComputer product intents, not claims that different providers use identical token budgets or reasoning semantics. Each adapter or route may translate or narrow them. No adapter may expose hidden thinking content.

Expected behavior and evidence limits

Higher effort can increase latency and reasoning-token cost. Changing effort can also invalidate prompt-cache reuse, which is why the phase-0.5 control is stable for the life of a conversation. The Activity panel continues to show only allow-listed summaries and actions, never hidden chain-of-thought.

This record qualifies the documented direct-Anthropic request/response contract, the Claude CLI adapter, and the local gateway behavior. Hermes is qualified by its own exact adapter registrations and separately qualified organization route, not merely because its upstream runtime has a similarly named setting. Codex and any other agent remain unqualified until their own gates pass. This record also does not claim a live Anthropic account smoke test or comparative latency/cost benchmark; those require an explicitly provisioned direct Anthropic route and provider credentials. See the Agent model and reasoning adapter playbook for the Claude/Hermes implementation comparison and failure lessons.