Context

RemoteClaw is middleware that delegates LLM execution to CLI agents (Claude Code, Gemini, Codex, OpenCode). Each CLI manages its own reasoning/thinking configuration. RemoteClaw should not attempt to translate or control LLM-level thinking settings — that's platform territory and violates the middleware boundary.

Problem

The thinkingLevel config surface is user-settable but has no runtime effect. This is a broken half-implementation.

Flow analysis

1. User sets it via CLI (--thinking high), TUI (/thinking command), ACP set_config_option thinking_level, or cron job flag
2. Value is persisted in session entries (SessionEntry.thinkingLevel)
3. Value is displayed back in the TUI (src/tui/tui.ts:757-759) and surfaced to ACP clients (src/acp/translator.ts:1001)
4. No CLI subprocess invocation passes the value. The CLI agents (src/middleware/runtimes/claude.ts, codex.ts, gemini.ts, opencode.ts) consume thinking/reasoning as OUTPUT events from their subprocess streams; they do not accept a thinkingLevel input and do not forward it as a CLI flag.
5. Merge logic in src/auto-reply/thinking.shared.ts — specifically resolveThinkingDefaultForModel — has zero callers. Dead function.
6. src/agents/subagent-spawn.ts:279 reads params.thinking into _thinkingOverrideRaw — the leading underscore indicates intentionally unused; subagent spawning silently discards the thinking override.
7. Tests asserting thinking-merge behavior (src/commands/agent.test.ts:840-846 "prefers per-model thinking over global thinkingDefault") are orphan: they assert against a mock of runEmbeddedPiAgent, which is not imported anywhere in production src/ code (the embedded execution path was gutted).

What reasoningLevel does (and why it stays)

reasoningLevel is a different concept and must be KEPT. It controls Telegram channel display behavior (off | on | stream — drives suppression/streaming/block-streaming of reasoning-event rendering in Telegram drafts) via resolveTelegramReasoningLevel in src/telegram/bot-message-dispatch.ts:92-112. Downstream it flows through src/auto-reply/reply/get-reply.ts, get-reply-inline-actions.ts, commands-status.ts. This is a legitimate channel-bound middleware concern.

Do not confuse the two. Gut only thinkingLevel / thinkingDefault / thinking-related helpers. Keep reasoningLevel everywhere.

Tasks

Session/state

• Remove thinkingLevel field from SessionEntry type in src/gateway/session-utils.types.ts:36 and src/gateway/protocol/schema/sessions.ts:101
• Remove thinkingLevel persistence in gateway server methods:
  • src/gateway/server-methods/sessions.ts:498
  • src/gateway/server-methods/agent.ts:468
  • src/gateway/server-methods/chat.ts:778-784
  • src/gateway/server-node-events.ts:167
  • src/gateway/session-utils.ts:826
• Remove from src/auto-reply/reply/session.ts session merge logic (lines 368, 370, 384, 386, 434)

Config

• Remove thinkingDefault field from src/config/types.agent-defaults.ts:178 and src/config/zod-schema.agent-defaults.ts:132-141
• Remove subagents.thinking override in src/config/zod-schema.agent-defaults.ts:182
• Regenerate src/config/schema.base.generated.ts
• Add migration warning if loaded config contains thinkingDefault or subagents.thinking (same channel used for other legacy-field warnings)

CLI surface

• Remove --thinking flag from src/cli/cron-cli/register.cron-add.ts:178-179 and register.cron-edit.ts:211-212
• Remove thinking field from cron test fixtures (src/cron/normalize.test.ts:368, src/cli/cron-cli.test.ts:233,402,408,430)
• Remove thinking: opts.thinking from src/commands/agent-via-gateway.ts:140 gateway payload
• Remove thinking option handling from src/commands/agent.ts (search for opts.thinking and the CLI option declaration)

TUI

• Remove thinkingLevel from src/tui/tui-types.ts:30
• Remove thinkingLevel from src/tui/tui-session-actions.ts:165-166, 288, 292
• Remove thinkingLevel display from src/tui/tui.ts:757-759
• Remove /thinking command handler from src/tui/tui-command-handlers.ts (search for thinkingLevel and remove the whole command branch)
• Remove from src/tui/gateway-chat.ts:80

ACP

• Remove thinking_level handling in src/acp/translator.ts:
  • set_config_option write path (~line 1030)
  • Response field emit (~line 1001)
  • Session-mode param handling (~lines 93, 203, 558-562, 663)
• Remove thinking_level from src/acp/commands.ts if exposed as a config option ID there
• Remove thinking assertions from src/acp/translator.set-session-mode.test.ts:39, 50 and from src/acp/translator.session-rate-limit.test.ts

Subagent spawn + cron jobs

• Remove thinking param from src/agents/tools/sessions-spawn-tool.ts:27, 91, 123
• Remove _thinkingOverrideRaw/params.thinking discard in src/agents/subagent-spawn.ts:279
• Remove thinking field handling from src/cron/service/jobs.ts:679-680, 763
• Remove thinking field references from src/agents/tools/sessions-history-tool.ts:66

Thinking helpers + orphan tests

• Delete the following from src/auto-reply/thinking.shared.ts, leaving only ReasoningLevel, VerboseLevel, and normalizeVerboseLevel (and anything else still used):
  • ThinkLevel type
  • resolveThinkingDefaultForModel
  • listThinkingLevels, listThinkingLevelLabels, formatThinkingLevels
  • formatXHighModelHint
  • isBinaryThinkingProvider
  • ThinkingCatalogEntry type
  • Any other helpers whose only consumers are being removed in this issue
• Delete src/tui/tui-formatters.ts:181 thinking-text display (verify it's the text-output rendering, not a level — if it's level-related, delete; if it's the stream of thinking text emitted by the CLI, keep since that's the output-event rendering and still in use)
• Delete orphan tests in src/commands/agent.test.ts:
  • expectDefaultThinkLevel helper function
  • "prefers per-model thinking over global thinkingDefault"
  • "defaults thinking to adaptive for Anthropic Claude 4.6 models"
  • Any other tests asserting getLastEmbeddedCall()?.thinkLevel

Acceptance Criteria

• rg "thinkingLevel|thinkingDefault|ThinkLevel" src/ returns no hits (migration-warning message text aside)
• rg "\.thinking\b|params\.thinking" src/ returns no hits, except thinking_delta and type === \"thinking\" references in src/middleware/runtimes/*.ts (those are output-event stream parsers and must stay)
• rg "reasoningLevel" src/ still returns live hits in Telegram, channel, ACP translator paths — this is intentional, reasoningLevel is kept
• User loading a config with thinkingDefault: "high" sees a migration warning and the config still loads
• remoteclaw cron add --thinking high — the flag either errors with "unknown option" or the whole line is removed; state the behavior in the PR description
• tsc --noEmit clean
• Test suite passes with no new skips