…as 1+2+4 of #2336

Partial execution of #2336's full sweep. Covers Areas 1 (`/think` directive
removal), 2 (`thinking.ts` helper deletion), and 4 (`thinkLevel` passthrough
plumbing) plus an adjacent cleanup of `onModelSelected` callback and its
`ModelSelectedContext` type. Areas 3, 5, 6, 7, 8 remain for follow-ups (see
the end of this commit message).

**Rationale**: thinking level and model selection are CLI-runtime concerns
per the Middleware Boundary Principle — CLI runtimes (Claude, Gemini, Codex,
OpenCode) own those decisions. Anything in the fork that resolves, gates,
or stores thinking level / selected model is vestigial from upstream
OpenClaw and was already no-op'd after PR #2335. `AgentRuntime.execute`
(src/middleware/types.ts) accepts none of these fields — confirmed.

## Area 1: `/think` directive end-to-end removal

- `src/auto-reply/reply/directive-handling.parse.ts`: removed
  `hasThinkDirective` and `thinkLevel` fields from `InlineDirectives`.
- `src/tui/tui-command-handlers.ts`: deleted the `/think` command handler;
  removed `formatThinkingLevels` import; removed `opts.thinking` passthrough
  from `sendChat`.
- `src/gateway/server-methods/chat.ts`: deleted the `injectThinking`-based
  `/think ${p.thinking} ${parsedMessage}` body rewrite and the `p.thinking`
  param.
- `src/auto-reply/reply.triggers.trigger-handling.targets-active-session-native-stop.e2e.test.ts`:
  deleted the `thinkCases` block that asserted `/think:high` stripping
  (behavior replaced by no-op since directive parser is gone).
- `src/cli/tui-cli.ts`: removed `--thinking <level>` CLI option.
- `src/tui/gateway-chat.ts`, `src/tui/tui-types.ts`: removed `thinking`
  option / `ChatSendOptions.thinking` field.

## Area 2: `thinking.ts` helper deletion

- `src/auto-reply/thinking.ts`: deleted `normalizeThinkLevel`,
  `supportsXHighThinking`, `formatXHighModelHint`, `listThinkingLevelLabels`,
  `formatThinkingLevels`, `resolveThinkingDefaultForModel`,
  `isBinaryThinkingProvider`, `XHIGH_MODEL_*` constants, `CLAUDE_46_MODEL_RE`,
  `ThinkingCatalogEntry` and `ThinkLevel` types. Kept and simplified
  `listThinkingLevels` (returns the static level list). Kept
  `normalizeVerboseLevel`, `normalizeNoticeLevel`, `normalizeUsageDisplay`,
  `normalizeElevatedLevel`, `resolveElevatedMode`, `normalizeReasoningLevel`,
  `normalizeFastMode` — these are RemoteClaw's own directive-normalization
  helpers (not thinking-specific) and are load-bearing for display/security
  layers.
- `src/commands/agent/session.ts`: removed `normalizeThinkLevel` import,
  `persistedThinking` field, and its `ThinkLevel`-typed call on
  `sessionEntry.thinkingLevel`.

## Area 4: `thinkLevel` dead passthrough + adjacent cleanup

- `src/auto-reply/reply/queue/types.ts`: removed `thinkLevel?: unknown`
  from `FollowupRun.run`.
- `src/auto-reply/reply/agent-runner-utils.ts`: dropped `thinkLevel`
  from `buildEmbeddedRunBaseParams`' returned object (function itself
  is dead Pi-embedded plumbing — tracked for follow-up).
- `src/auto-reply/reply/agent-runner-execution.ts`: deleted the
  `onModelSelected({provider, model, thinkLevel: undefined})` callback
  call and the now-unused `const model = params.followupRun.run.model`
  declaration.
- `src/auto-reply/reply/get-reply-run.ts`: removed `resolvedThinkLevel`
  field from `RunPreparedReplyParams` and the `thinkLevel:
  resolvedThinkLevel` assignment in the `followupRun.run` builder.
- `src/auto-reply/reply/get-reply.ts`, `get-reply-inline-actions.ts`,
  `get-reply-directives.ts`, `commands-types.ts`: dropped
  `resolvedThinkLevel` fields + passthrough destructures.
- `src/auto-reply/types.ts`: deleted the whole `ModelSelectedContext`
  type and `GetReplyOptions.onModelSelected` callback (these echoed
  configured-value `provider`/`model`/`thinkLevel` — not actual
  CLI-reported selection; meaningful only when the Pi-embedded runtime
  owned dynamic fallback, which was gutted in #76/#77).
- `src/channels/reply-prefix.ts`: replaced the mutation-based
  `onModelSelected` callback with a static read from
  `resolveAgentEffectiveModelPrimary(cfg, agentId)` at context-creation
  time. Response-prefix template interpolation (`{model}`,
  `{provider}`, `{modelFull}`) still works with configured values.
  `ResponsePrefixContext.thinkingLevel` dropped.
- Removed `{ onModelSelected, ...rest } = createReplyPrefixOptions(...)`
  destructures and `onModelSelected` property spreads across ~16
  channel integrations: bluebubbles, feishu, googlechat, matrix,
  mattermost (monitor + slash-http), msteams, twitch, zalo, zalouser,
  discord (agent-components + message-handler + native-command),
  gateway/chat, imessage, line, plugin-sdk/inbound-reply-dispatch,
  signal, slack (monitor/dispatch + slash), telegram (bot-message +
  native-commands), tui (command-handlers + gateway-chat), web
  (auto-reply/process-message). Plus 3 test fixtures
  (`get-reply-run.media-only.test.ts` — 2 stale think-hint tests
  deleted; `get-reply-inline-actions.skip-when-config-empty.test.ts`,
  `get-reply.reset-hooks-fallback.test.ts`; `agent-runner-utils.test.ts`
  — stale `toMatchObject` keys; `test-helpers.ts`, `agent-runner.*.test.ts`
  fixtures; `slack/monitor/slash.test-harness.ts` mock return value).

## Areas deferred to follow-up issues

- **Area 3** (`SessionEntry.thinkingLevel` field removal + gateway
  schema + session.ts `persistedThinking`): still live in the session
  store, surfaced by TUI / status / Telegram / ACP / gateway session
  APIs. Mechanical but touches wire-format / session migration.
- **Area 5** (directive handling model-state internals — `aliasIndex`,
  `allowedModelKeys`, `resetModelOverride`, `defaultProvider` /
  `defaultModel`, `initialModelLabel`, `formatModelSwitchEvent`): 4
  files, significant refactoring.
- **Area 6** (docs): `/think` references in slash-commands.md,
  context.md, tui.md, group-messages.md.
- **Area 7** (display-layer audit): Android, iOS, macOS, web UI
  display references to `thinkingLevel`. Per-file judgment required
  (pass-through display = KEEP, state driver = REMOVE). Issue #2336
  explicitly flagged: "Do NOT touch native apps without confirming".
- **Area 8** (test-layer hardening): delete the global `runPreparedReply`
  shield mock at `get-reply.test-mocks.ts:39-40`; add regression test
  exercising the real resolveReplyDirectives → runPreparedReply chain.

## Additional finding (filed as follow-up)

The `buildEmbeddedRunBaseParams` / `buildEmbeddedContextFromTemplate` /
`buildEmbeddedRunContexts` / `buildTemplateSenderContext` /
`resolveRunAuthProfile` / `resolveEnforceFinalTag` /
`resolveProviderScopedAuthProfile` / `resolveModelFallbackOptions` /
`buildThreadingToolContext` cluster in
`src/auto-reply/reply/agent-runner-utils.ts` has zero production
callers — dead Pi-embedded plumbing that threads verboseLevel /
reasoningLevel / elevatedLevel / bashElevated / execOverrides for
nothing. Tracked as a separate follow-up.

## Verification

- `pnpm check` (format + tsgo + oxlint + project-specific lints) → exit 0.
- `pnpm test` (full parallel suite) → **800 files / 7010 passed /
  3 skipped** — no regression.
- Rescan: `git grep "/think\|hasThinkDirective\|thinkLevel\|supportsXHighThinking\|normalizeThinkLevel\|formatXHighModelHint"`
  in `src/` / `extensions/` returns only: `thinkingLevel` (stored in
  session, Area 3 deferred), display-layer native-app references
  (Area 7 deferred), and docs mentions (Area 6 deferred) — all scoped
  to follow-ups.

Refs: #2336, #2334, #2335, #2089

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…as 1+2+4 of #2336 (#2463)

Partial execution of #2336's full sweep. Covers Areas 1 (`/think` directive
removal), 2 (`thinking.ts` helper deletion), and 4 (`thinkLevel` passthrough
plumbing) plus an adjacent cleanup of `onModelSelected` callback and its
`ModelSelectedContext` type. Areas 3, 5, 6, 7, 8 remain for follow-ups (see
the end of this commit message).

**Rationale**: thinking level and model selection are CLI-runtime concerns
per the Middleware Boundary Principle — CLI runtimes (Claude, Gemini, Codex,
OpenCode) own those decisions. Anything in the fork that resolves, gates,
or stores thinking level / selected model is vestigial from upstream
OpenClaw and was already no-op'd after PR #2335. `AgentRuntime.execute`
(src/middleware/types.ts) accepts none of these fields — confirmed.

## Area 1: `/think` directive end-to-end removal

- `src/auto-reply/reply/directive-handling.parse.ts`: removed
  `hasThinkDirective` and `thinkLevel` fields from `InlineDirectives`.
- `src/tui/tui-command-handlers.ts`: deleted the `/think` command handler;
  removed `formatThinkingLevels` import; removed `opts.thinking` passthrough
  from `sendChat`.
- `src/gateway/server-methods/chat.ts`: deleted the `injectThinking`-based
  `/think ${p.thinking} ${parsedMessage}` body rewrite and the `p.thinking`
  param.
- `src/auto-reply/reply.triggers.trigger-handling.targets-active-session-native-stop.e2e.test.ts`:
  deleted the `thinkCases` block that asserted `/think:high` stripping
  (behavior replaced by no-op since directive parser is gone).
- `src/cli/tui-cli.ts`: removed `--thinking <level>` CLI option.
- `src/tui/gateway-chat.ts`, `src/tui/tui-types.ts`: removed `thinking`
  option / `ChatSendOptions.thinking` field.

## Area 2: `thinking.ts` helper deletion

- `src/auto-reply/thinking.ts`: deleted `normalizeThinkLevel`,
  `supportsXHighThinking`, `formatXHighModelHint`, `listThinkingLevelLabels`,
  `formatThinkingLevels`, `resolveThinkingDefaultForModel`,
  `isBinaryThinkingProvider`, `XHIGH_MODEL_*` constants, `CLAUDE_46_MODEL_RE`,
  `ThinkingCatalogEntry` and `ThinkLevel` types. Kept and simplified
  `listThinkingLevels` (returns the static level list). Kept
  `normalizeVerboseLevel`, `normalizeNoticeLevel`, `normalizeUsageDisplay`,
  `normalizeElevatedLevel`, `resolveElevatedMode`, `normalizeReasoningLevel`,
  `normalizeFastMode` — these are RemoteClaw's own directive-normalization
  helpers (not thinking-specific) and are load-bearing for display/security
  layers.
- `src/commands/agent/session.ts`: removed `normalizeThinkLevel` import,
  `persistedThinking` field, and its `ThinkLevel`-typed call on
  `sessionEntry.thinkingLevel`.

## Area 4: `thinkLevel` dead passthrough + adjacent cleanup

- `src/auto-reply/reply/queue/types.ts`: removed `thinkLevel?: unknown`
  from `FollowupRun.run`.
- `src/auto-reply/reply/agent-runner-utils.ts`: dropped `thinkLevel`
  from `buildEmbeddedRunBaseParams`' returned object (function itself
  is dead Pi-embedded plumbing — tracked for follow-up).
- `src/auto-reply/reply/agent-runner-execution.ts`: deleted the
  `onModelSelected({provider, model, thinkLevel: undefined})` callback
  call and the now-unused `const model = params.followupRun.run.model`
  declaration.
- `src/auto-reply/reply/get-reply-run.ts`: removed `resolvedThinkLevel`
  field from `RunPreparedReplyParams` and the `thinkLevel:
  resolvedThinkLevel` assignment in the `followupRun.run` builder.
- `src/auto-reply/reply/get-reply.ts`, `get-reply-inline-actions.ts`,
  `get-reply-directives.ts`, `commands-types.ts`: dropped
  `resolvedThinkLevel` fields + passthrough destructures.
- `src/auto-reply/types.ts`: deleted the whole `ModelSelectedContext`
  type and `GetReplyOptions.onModelSelected` callback (these echoed
  configured-value `provider`/`model`/`thinkLevel` — not actual
  CLI-reported selection; meaningful only when the Pi-embedded runtime
  owned dynamic fallback, which was gutted in #76/#77).
- `src/channels/reply-prefix.ts`: replaced the mutation-based
  `onModelSelected` callback with a static read from
  `resolveAgentEffectiveModelPrimary(cfg, agentId)` at context-creation
  time. Response-prefix template interpolation (`{model}`,
  `{provider}`, `{modelFull}`) still works with configured values.
  `ResponsePrefixContext.thinkingLevel` dropped.
- Removed `{ onModelSelected, ...rest } = createReplyPrefixOptions(...)`
  destructures and `onModelSelected` property spreads across ~16
  channel integrations: bluebubbles, feishu, googlechat, matrix,
  mattermost (monitor + slash-http), msteams, twitch, zalo, zalouser,
  discord (agent-components + message-handler + native-command),
  gateway/chat, imessage, line, plugin-sdk/inbound-reply-dispatch,
  signal, slack (monitor/dispatch + slash), telegram (bot-message +
  native-commands), tui (command-handlers + gateway-chat), web
  (auto-reply/process-message). Plus 3 test fixtures
  (`get-reply-run.media-only.test.ts` — 2 stale think-hint tests
  deleted; `get-reply-inline-actions.skip-when-config-empty.test.ts`,
  `get-reply.reset-hooks-fallback.test.ts`; `agent-runner-utils.test.ts`
  — stale `toMatchObject` keys; `test-helpers.ts`, `agent-runner.*.test.ts`
  fixtures; `slack/monitor/slash.test-harness.ts` mock return value).

## Areas deferred to follow-up issues

- **Area 3** (`SessionEntry.thinkingLevel` field removal + gateway
  schema + session.ts `persistedThinking`): still live in the session
  store, surfaced by TUI / status / Telegram / ACP / gateway session
  APIs. Mechanical but touches wire-format / session migration.
- **Area 5** (directive handling model-state internals — `aliasIndex`,
  `allowedModelKeys`, `resetModelOverride`, `defaultProvider` /
  `defaultModel`, `initialModelLabel`, `formatModelSwitchEvent`): 4
  files, significant refactoring.
- **Area 6** (docs): `/think` references in slash-commands.md,
  context.md, tui.md, group-messages.md.
- **Area 7** (display-layer audit): Android, iOS, macOS, web UI
  display references to `thinkingLevel`. Per-file judgment required
  (pass-through display = KEEP, state driver = REMOVE). Issue #2336
  explicitly flagged: "Do NOT touch native apps without confirming".
- **Area 8** (test-layer hardening): delete the global `runPreparedReply`
  shield mock at `get-reply.test-mocks.ts:39-40`; add regression test
  exercising the real resolveReplyDirectives → runPreparedReply chain.

## Additional finding (filed as follow-up)

The `buildEmbeddedRunBaseParams` / `buildEmbeddedContextFromTemplate` /
`buildEmbeddedRunContexts` / `buildTemplateSenderContext` /
`resolveRunAuthProfile` / `resolveEnforceFinalTag` /
`resolveProviderScopedAuthProfile` / `resolveModelFallbackOptions` /
`buildThreadingToolContext` cluster in
`src/auto-reply/reply/agent-runner-utils.ts` has zero production
callers — dead Pi-embedded plumbing that threads verboseLevel /
reasoningLevel / elevatedLevel / bashElevated / execOverrides for
nothing. Tracked as a separate follow-up.

## Verification

- `pnpm check` (format + tsgo + oxlint + project-specific lints) → exit 0.
- `pnpm test` (full parallel suite) → **800 files / 7010 passed /
  3 skipped** — no regression.
- Rescan: `git grep "/think\|hasThinkDirective\|thinkLevel\|supportsXHighThinking\|normalizeThinkLevel\|formatXHighModelHint"`
  in `src/` / `extensions/` returns only: `thinkingLevel` (stored in
  session, Area 3 deferred), display-layer native-app references
  (Area 7 deferred), and docs mentions (Area 6 deferred) — all scoped
  to follow-ups.

Refs: #2336, #2334, #2335, #2089

Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…ea 5 of #2336

Follow-up to PR #2463 (Areas 1+2+4 of #2336). Completes the directive-handling
portion of the sweep by removing model-state plumbing that was already empty/
trivial after #2335's `createModelSelectionState` removal.

Per the Middleware Boundary Principle: RemoteClaw does NOT control model
selection — CLI runtimes (Claude, Gemini, Codex, OpenCode) own those decisions.
`handleDirectiveOnly` and `applyInlineDirectivesFastLane` carried 10 fields
(`aliasIndex`, `allowedModelKeys`, `allowedModelCatalog`, `resetModelOverride`,
`defaultProvider`, `defaultModel`, `initialModelLabel`, `formatModelSwitchEvent`,
`provider`, `model`) that are now fully dead.

## Changes

- `src/auto-reply/reply/directive-handling.params.ts`: delete all 10 model-state
  fields from `HandleDirectiveOnlyCoreParams`; delete unused `surface?: string`
  from `HandleDirectiveOnlyParams`.
- `src/auto-reply/reply/directive-handling.impl.ts`: queue-validation `channel`
  semantic fix — switch from misused `params.provider` (AI provider, e.g.
  "anthropic") to `params.messageProviderKey ?? ""` (actual message channel,
  e.g. "telegram"). Aligns with `get-reply-run.ts:382` which already passes
  `sessionCtx.Provider` correctly. User-visible effect is limited to
  channel-specific queue config display when running `/queue` without args.
- `src/auto-reply/reply/directive-handling.fast-lane.ts`: drop model-state
  destructures; drop dead `sessionEntry.providerOverride`/`modelOverride`
  reads; simplify return type `{directiveAck?; provider; model}` →
  `{directiveAck?}`. *Note*: `providerOverride`/`modelOverride` ARE written
  by `src/gateway/sessions-patch.ts:213-214` and
  `src/agents/tools/session-status-tool.ts:254-255`, but their downstream
  consumption via this fast-lane path reaches only `isReasoningTagProvider`
  at `get-reply-run.ts:453` (`enforceFinalTag` for google/google-gemini-cli/
  minimax) — vestigial Pi-era plumbing per #2336's gut thesis.
- `src/auto-reply/reply/get-reply-directives-apply.ts`: delete inline empty-
  default `directiveModelState` compat shim (introduced in #2335); delete
  `aliasIndex`/`initialModelLabel`/`formatModelSwitchEvent` threading; inline
  single-call `createDirectiveHandlingBase` factory; drop `provider`/`model`
  from `ApplyDirectiveResult.continue` variant (pure pass-through of inputs
  post-gut).
- `src/auto-reply/reply/get-reply-directives.ts` (required cascade): delete
  dead `initialModelLabel`/`formatModelSwitchEvent` constructions and their
  two param passes (orphaned by the scope reduction in apply.ts); collapse
  `let provider`/`let model` to `const` (no longer reassigned).

## Deferred

- `resolveDefaultModel` stub in `directive-handling.ts` still returns
  `allowedModelKeys`/`allowedModelCatalog`/`resetModelOverride` — never read
  by callers. Separate cleanup wave.
- `aliasIndex` still threaded through `get-reply.ts` → `resolveReplyDirectives`
  (dead throughout). Separate cleanup wave.
- Adjacent Areas 3 (`SessionEntry.thinkingLevel`), 6 (docs), 7 (display audit),
  8 (test-layer hardening) of #2336 remain as separate follow-ups.

## Verification

- `pnpm check`: exit 0 (format + tsgo + oxlint + project-specific lints)
- `pnpm test`: 800 files / 7010 passed / 3 skipped — matches #2463 baseline
- Grep across the 4 target files: zero hits for all 10 gutted symbols

Refs: #2465, #2336, #2335, #2463

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…ea 5 of #2336 (#2475)

Follow-up to PR #2463 (Areas 1+2+4 of #2336). Completes the directive-handling
portion of the sweep by removing model-state plumbing that was already empty/
trivial after #2335's `createModelSelectionState` removal.

Per the Middleware Boundary Principle: RemoteClaw does NOT control model
selection — CLI runtimes (Claude, Gemini, Codex, OpenCode) own those decisions.
`handleDirectiveOnly` and `applyInlineDirectivesFastLane` carried 10 fields
(`aliasIndex`, `allowedModelKeys`, `allowedModelCatalog`, `resetModelOverride`,
`defaultProvider`, `defaultModel`, `initialModelLabel`, `formatModelSwitchEvent`,
`provider`, `model`) that are now fully dead.

## Changes

- `src/auto-reply/reply/directive-handling.params.ts`: delete all 10 model-state
  fields from `HandleDirectiveOnlyCoreParams`; delete unused `surface?: string`
  from `HandleDirectiveOnlyParams`.
- `src/auto-reply/reply/directive-handling.impl.ts`: queue-validation `channel`
  semantic fix — switch from misused `params.provider` (AI provider, e.g.
  "anthropic") to `params.messageProviderKey ?? ""` (actual message channel,
  e.g. "telegram"). Aligns with `get-reply-run.ts:382` which already passes
  `sessionCtx.Provider` correctly. User-visible effect is limited to
  channel-specific queue config display when running `/queue` without args.
- `src/auto-reply/reply/directive-handling.fast-lane.ts`: drop model-state
  destructures; drop dead `sessionEntry.providerOverride`/`modelOverride`
  reads; simplify return type `{directiveAck?; provider; model}` →
  `{directiveAck?}`. *Note*: `providerOverride`/`modelOverride` ARE written
  by `src/gateway/sessions-patch.ts:213-214` and
  `src/agents/tools/session-status-tool.ts:254-255`, but their downstream
  consumption via this fast-lane path reaches only `isReasoningTagProvider`
  at `get-reply-run.ts:453` (`enforceFinalTag` for google/google-gemini-cli/
  minimax) — vestigial Pi-era plumbing per #2336's gut thesis.
- `src/auto-reply/reply/get-reply-directives-apply.ts`: delete inline empty-
  default `directiveModelState` compat shim (introduced in #2335); delete
  `aliasIndex`/`initialModelLabel`/`formatModelSwitchEvent` threading; inline
  single-call `createDirectiveHandlingBase` factory; drop `provider`/`model`
  from `ApplyDirectiveResult.continue` variant (pure pass-through of inputs
  post-gut).
- `src/auto-reply/reply/get-reply-directives.ts` (required cascade): delete
  dead `initialModelLabel`/`formatModelSwitchEvent` constructions and their
  two param passes (orphaned by the scope reduction in apply.ts); collapse
  `let provider`/`let model` to `const` (no longer reassigned).

## Deferred

- `resolveDefaultModel` stub in `directive-handling.ts` still returns
  `allowedModelKeys`/`allowedModelCatalog`/`resetModelOverride` — never read
  by callers. Separate cleanup wave.
- `aliasIndex` still threaded through `get-reply.ts` → `resolveReplyDirectives`
  (dead throughout). Separate cleanup wave.
- Adjacent Areas 3 (`SessionEntry.thinkingLevel`), 6 (docs), 7 (display audit),
  8 (test-layer hardening) of #2336 remain as separate follow-ups.

## Verification

- `pnpm check`: exit 0 (format + tsgo + oxlint + project-specific lints)
- `pnpm test`: 800 files / 7010 passed / 3 skipped — matches #2463 baseline
- Grep across the 4 target files: zero hits for all 10 gutted symbols

Refs: #2465, #2336, #2335, #2463

Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>