…inate CLI subprocesses (#2344)

The subagent `kill` and `steer` actions (exposed via the `subagents`
agent tool and `/subagents kill` / `/subagents steer` chat commands)
lost their subprocess-termination semantics during the pi-embedded gut
(`b27cecc795`, #76/#77). Kill marked the registry entry as terminated
and cleared queues but the CLI subprocess kept running until natural
completion; steer queued a replacement run but left the existing run
to finish first (via `agent.wait` settlement) instead of interrupting
it.

Fix: call `killSessionRun` from `src/agents/session-run-registry.ts`
before the queue-clear / settlement logic at all four affected call
sites. The registry is already populated by `ChannelBridge` (register
at `channel-bridge.ts:160` in a `try`, unregister at `:349` in a
`finally`), and `killSessionRun` dispatches either
`abortController.abort()` or `process.kill(pid, "SIGTERM")`. It is
idempotent (returns `false` on unknown/already-aborted keys — safe
no-op) and cannot throw (internal `try/catch` around `process.kill`).

Changes:
- `src/agents/tools/subagents-tool.ts`:
  - Import `killSessionRun` from `../session-run-registry.js`.
  - `killSubagentRun` (kill path): replace `const aborted = false;`
    dead var with `const aborted = killSessionRun(childSessionKey);`.
    The outer `killed` aggregate (`marked > 0 || aborted || cleared...`)
    now correctly reflects when a kill signal was dispatched. Previously
    `aborted` was a no-op disjunct; this only ADDS truthy cases, so no
    previously-true scenario regresses.
  - Steer path (line ~624): insert
    `killSessionRun(resolved.entry.childSessionKey);` between the
    `sessionId` resolution and `clearSessionQueues`, so the existing
    run is aborted before the `agent.wait` settle-timeout polls. The
    existing `agent.wait` stays as a 5-second settlement window.
  - `cascadeKillChildren` propagates automatically — it delegates to
    `killSubagentRun`.
- `src/auto-reply/reply/commands-subagents/action-kill.ts`: import +
  `killSessionRun(childKey);` inserted between `loadSubagentSessionEntry`
  and `clearSessionQueues`.
- `src/auto-reply/reply/commands-subagents/action-send.ts`: import +
  `killSessionRun(targetResolution.entry.childSessionKey);` in the
  `steerRequested` branch between `markSubagentRunForSteerRestart` and
  `clearSessionQueues`.

Key-type correctness: `childSessionKey` (from `SubagentRunRecord`) is
the canonical ChannelBridge sessionKey that the registry keys on.
`entry.sessionId` (Pi-era UUID) is NOT used for registry lookups —
preserved for `clearSessionQueues` and `loadSubagentSessionEntry` only.

Out of scope (separate issues per #2344 body):
- `abort.ts::stopSubagentsForRequester` carries the same
  `const aborted = false;` dead-var pattern — tracked under the `/abort`
  user command + gateway `sessions.reset` sub-issue.
- Auto-reply queue mode and `abortEmbeddedPiRun` test-mock cleanup —
  separate sub-issues.
- `pi-embedded.ts` stub barrel deletion — tracked under #2089 cascade
  sweep.

Verification:
- `pnpm check` (format + tsgo + lint + project-specific lints) → exit 0
- `pnpm test` (full parallel suite) → **800 files / 7010 passed /
  3 skipped** — no regression
- Rescan: `git grep "killSessionRun"` returns registry definition +
  registry tests + the 3 files in this PR; no stale/unwired references
- Rescan: `git grep "abortEmbeddedPiRun"` returns only test-mock
  contexts (`reply.block-streaming.test.ts`, `reply/abort.test.ts`) —
  production paths have no stray references
- Adversarial validation (fresh-context subclaude): CLEAN verdict on
  6 AC + 11 adversarial checks — confirms `killSessionRun` is LIVE
  (abortController/SIGTERM), key-type match correct, no double-kill in
  `action-kill.ts → stopSubagentsForRequester` path (latter targets
  grandchildren, not the killed session), order `mark → kill → clear →
  wait` is race-free (all sync within a tick)

Closes #2344

Refs: #2089

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

…inate CLI subprocesses (#2344) (#2461)

The subagent `kill` and `steer` actions (exposed via the `subagents`
agent tool and `/subagents kill` / `/subagents steer` chat commands)
lost their subprocess-termination semantics during the pi-embedded gut
(`b27cecc795`, #76/#77). Kill marked the registry entry as terminated
and cleared queues but the CLI subprocess kept running until natural
completion; steer queued a replacement run but left the existing run
to finish first (via `agent.wait` settlement) instead of interrupting
it.

Fix: call `killSessionRun` from `src/agents/session-run-registry.ts`
before the queue-clear / settlement logic at all four affected call
sites. The registry is already populated by `ChannelBridge` (register
at `channel-bridge.ts:160` in a `try`, unregister at `:349` in a
`finally`), and `killSessionRun` dispatches either
`abortController.abort()` or `process.kill(pid, "SIGTERM")`. It is
idempotent (returns `false` on unknown/already-aborted keys — safe
no-op) and cannot throw (internal `try/catch` around `process.kill`).

Changes:
- `src/agents/tools/subagents-tool.ts`:
  - Import `killSessionRun` from `../session-run-registry.js`.
  - `killSubagentRun` (kill path): replace `const aborted = false;`
    dead var with `const aborted = killSessionRun(childSessionKey);`.
    The outer `killed` aggregate (`marked > 0 || aborted || cleared...`)
    now correctly reflects when a kill signal was dispatched. Previously
    `aborted` was a no-op disjunct; this only ADDS truthy cases, so no
    previously-true scenario regresses.
  - Steer path (line ~624): insert
    `killSessionRun(resolved.entry.childSessionKey);` between the
    `sessionId` resolution and `clearSessionQueues`, so the existing
    run is aborted before the `agent.wait` settle-timeout polls. The
    existing `agent.wait` stays as a 5-second settlement window.
  - `cascadeKillChildren` propagates automatically — it delegates to
    `killSubagentRun`.
- `src/auto-reply/reply/commands-subagents/action-kill.ts`: import +
  `killSessionRun(childKey);` inserted between `loadSubagentSessionEntry`
  and `clearSessionQueues`.
- `src/auto-reply/reply/commands-subagents/action-send.ts`: import +
  `killSessionRun(targetResolution.entry.childSessionKey);` in the
  `steerRequested` branch between `markSubagentRunForSteerRestart` and
  `clearSessionQueues`.

Key-type correctness: `childSessionKey` (from `SubagentRunRecord`) is
the canonical ChannelBridge sessionKey that the registry keys on.
`entry.sessionId` (Pi-era UUID) is NOT used for registry lookups —
preserved for `clearSessionQueues` and `loadSubagentSessionEntry` only.

Out of scope (separate issues per #2344 body):
- `abort.ts::stopSubagentsForRequester` carries the same
  `const aborted = false;` dead-var pattern — tracked under the `/abort`
  user command + gateway `sessions.reset` sub-issue.
- Auto-reply queue mode and `abortEmbeddedPiRun` test-mock cleanup —
  separate sub-issues.
- `pi-embedded.ts` stub barrel deletion — tracked under #2089 cascade
  sweep.

Verification:
- `pnpm check` (format + tsgo + lint + project-specific lints) → exit 0
- `pnpm test` (full parallel suite) → **800 files / 7010 passed /
  3 skipped** — no regression
- Rescan: `git grep "killSessionRun"` returns registry definition +
  registry tests + the 3 files in this PR; no stale/unwired references
- Rescan: `git grep "abortEmbeddedPiRun"` returns only test-mock
  contexts (`reply.block-streaming.test.ts`, `reply/abort.test.ts`) —
  production paths have no stray references
- Adversarial validation (fresh-context subclaude): CLEAN verdict on
  6 AC + 11 adversarial checks — confirms `killSessionRun` is LIVE
  (abortController/SIGTERM), key-type match correct, no double-kill in
  `action-kill.ts → stopSubagentsForRequester` path (latter targets
  grandchildren, not the killed session), order `mark → kill → clear →
  wait` is race-free (all sync within a tick)

Closes #2344

Refs: #2089

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

…ctually terminate CLI subprocesses (#2343)

Replace three dead `const aborted = false` / `const ended = true`
placeholders left behind by the Pi-embedded gut (`b27cecc795`, #76/#77)
with real calls to `killSessionRun` / `waitForSessionRunEnd` from
`src/agents/session-run-registry.ts`:

- `abort.ts::stopSubagentsForRequester` inner loop — `killSessionRun(childKey)`
  now flows into the `killed` aggregate, so the loop actually counts
  subagent runs whose subprocesses were signaled.
- `abort.ts::tryFastAbortFromMessage` — `killSessionRun(resolvedTargetKey)`
  returned in the `aborted` field (previously always `false`). User-facing
  `/abort` now terminates the running CLI subprocess for non-ACP sessions;
  ACP cancellation path remains separate.
- `sessions.ts::ensureSessionRuntimeCleanup` — replaces the `const ended = true`
  placeholder with `killSessionRun(canonicalKey)` + `waitForSessionRunEnd`
  (15s timeout). Returns `UNAVAILABLE` on timeout, restoring the pre-gut
  contract that `sessions.reset` / `sessions.delete` surface the error
  when a subprocess refuses to end.

Also removes the `if (!params.sessionId) return undefined;` early-exit
in `ensureSessionRuntimeCleanup`: `sessionId` is used only to extend the
queue-key set; termination operates on `canonicalKey`, which is always
present. Callers with no `sessionId` now correctly attempt termination
(`waitForSessionRunEnd` short-circuits to `true` when no run is
registered, so the wait has no cost when there's nothing to wait for).

Polish: extracted `15_000` to `SESSION_RUN_TERMINATION_TIMEOUT_MS`,
grouped with the sibling `ACP_RUNTIME_CLEANUP_TIMEOUT_MS` constant.

Issue body lists three regression sites; only two apply. The third
(`commands-session.ts::applyAbortTarget`) does not exist in the current
tree — `git grep applyAbortTarget` returns zero hits. The `/session
abort` pathway was consolidated into `tryFastAbortFromMessage` at some
earlier point, so no duplicate wiring is needed. AC3 marked N/A.

Verification:
- `pnpm check` → exit 0
- `pnpm test` (full parallel suite) → 800 files / 7010 passed /
  3 skipped — no regression
- Rescan: `git grep "const aborted = false"` → zero hits across the
  whole tree (combined sweep of #2344 subagent kill/steer + #2343
  abort/subagent-cascade/fastAbort)
- Rescan: `git grep "abortEmbeddedPiRun"` → only test-mock contexts;
  production has no remaining references (test-mock cleanup tracked
  under #2089 sweep)
- Adversarial validation (fresh-context subclaude): CLEAN on 7 AC +
  9 adversarial checks — confirms waitForSessionRunEnd is LIVE (50ms
  poll tick, max 15s wall-clock), `ensureSessionRuntimeCleanup` callers
  always pass a valid `canonicalKey`, and no test asserts `aborted ===
  false` on the fast-abort path

Closes #2343

Refs: #2089, #2344, #2460, #2461

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

…ctually terminate CLI subprocesses (#2343) (#2462)

Replace three dead `const aborted = false` / `const ended = true`
placeholders left behind by the Pi-embedded gut (`b27cecc795`, #76/#77)
with real calls to `killSessionRun` / `waitForSessionRunEnd` from
`src/agents/session-run-registry.ts`:

- `abort.ts::stopSubagentsForRequester` inner loop — `killSessionRun(childKey)`
  now flows into the `killed` aggregate, so the loop actually counts
  subagent runs whose subprocesses were signaled.
- `abort.ts::tryFastAbortFromMessage` — `killSessionRun(resolvedTargetKey)`
  returned in the `aborted` field (previously always `false`). User-facing
  `/abort` now terminates the running CLI subprocess for non-ACP sessions;
  ACP cancellation path remains separate.
- `sessions.ts::ensureSessionRuntimeCleanup` — replaces the `const ended = true`
  placeholder with `killSessionRun(canonicalKey)` + `waitForSessionRunEnd`
  (15s timeout). Returns `UNAVAILABLE` on timeout, restoring the pre-gut
  contract that `sessions.reset` / `sessions.delete` surface the error
  when a subprocess refuses to end.

Also removes the `if (!params.sessionId) return undefined;` early-exit
in `ensureSessionRuntimeCleanup`: `sessionId` is used only to extend the
queue-key set; termination operates on `canonicalKey`, which is always
present. Callers with no `sessionId` now correctly attempt termination
(`waitForSessionRunEnd` short-circuits to `true` when no run is
registered, so the wait has no cost when there's nothing to wait for).

Polish: extracted `15_000` to `SESSION_RUN_TERMINATION_TIMEOUT_MS`,
grouped with the sibling `ACP_RUNTIME_CLEANUP_TIMEOUT_MS` constant.

Issue body lists three regression sites; only two apply. The third
(`commands-session.ts::applyAbortTarget`) does not exist in the current
tree — `git grep applyAbortTarget` returns zero hits. The `/session
abort` pathway was consolidated into `tryFastAbortFromMessage` at some
earlier point, so no duplicate wiring is needed. AC3 marked N/A.

Verification:
- `pnpm check` → exit 0
- `pnpm test` (full parallel suite) → 800 files / 7010 passed /
  3 skipped — no regression
- Rescan: `git grep "const aborted = false"` → zero hits across the
  whole tree (combined sweep of #2344 subagent kill/steer + #2343
  abort/subagent-cascade/fastAbort)
- Rescan: `git grep "abortEmbeddedPiRun"` → only test-mock contexts;
  production has no remaining references (test-mock cleanup tracked
  under #2089 sweep)
- Adversarial validation (fresh-context subclaude): CLEAN on 7 AC +
  9 adversarial checks — confirms waitForSessionRunEnd is LIVE (50ms
  poll tick, max 15s wall-clock), `ensureSessionRuntimeCleanup` callers
  always pass a valid `canonicalKey`, and no test asserts `aborted ===
  false` on the fast-abort path

Closes #2343

Refs: #2089, #2344, #2460, #2461

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

…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>