fix(coding-agent): restore the /thinking interactive dispatch

[ndaemy]

Summary

/thinking and /thinking <level> were listed in autocomplete and /help but had no branch in the interactive submit handler, so the text fell through to session.prompt() and reached the model as an ordinary user message (queued as a steering message while a reply was streaming).

Upstream 496185f added the thinking entry to BUILTIN_SLASH_COMMANDS together with its dispatch; the sync merge 4632790 (#1119) kept the entry and dropped the dispatch. This restores it:

• setupEditorSubmitHandler: /thinking opens the level selector, /thinking <level> sets the level (case-insensitive; unknown levels report the available ones).
• createBaseAutocompleteProvider: argument completions for thinking from session.getAvailableThinkingLevels().
• handleThinkingCommand / selectThinkingLevel / showThinkingSelector from upstream, with upstream's setThinkingLevel(level, { persist }) mapped onto the fork's split setters: /thinking <level> and Enter in the selector use setSessionThinkingLevel (session scope), Ctrl+S uses setThinkingLevel, which records the per-model remembered level (modelThinkingLevels) exactly like the /settings thinking row. Ctrl+S does not write the global defaultThinkingLevel; per-model memory is what startup consults first (feat(coding-agent): per-model reasoning memory, /reasoning + /efforts commands, persistent /fast #894).
• getAvailableThinkingLevels() is awaited at every new call site because InteractiveSession widens it for the shared-host proxy.

docs/settings.md and docs/usage.md still described Ctrl+S as saving the global startup default (wording carried over from upstream); they now describe the per-model memory and the defaultThinkingLevel fallback.

/reasoning, /efforts, and Shift+Tab are untouched.

Fixes #1437

Verification

• test/suite/regressions/1437-thinking-command-dispatch.test.ts: submit-handler routing (/thinking high → handler, never session.prompt; bare /thinking → selector; /thinking-… not matched) and handler behavior (session-scoped apply, case-insensitive match against a Promise-returning level list, unknown-level error text, selector path, Ctrl+S persist path).
• bun run --cwd packages/coding-agent test test/suite/regressions/1437-thinking-command-dispatch.test.ts test/interactive-mode-exit-command.test.ts test/interactive-mode-status.test.ts test/suite/reasoning-commands.test.ts — 4 files, 94 tests passed.
• bun run check — passed.
• senpi-qa TUI run (tmux, isolated agent dir, local fake OpenAI-compatible model so the request body is visible), before → after:
  • idle /thinking high: user message sent to the model → no request, status Thinking level: high, footer mock-model:high
  • /thinking low while a reply streams: appended to the next request → applied immediately, no extra request
  • /thinking turbo: sent to the model → Unknown thinking level "turbo". Available levels: off, minimal, low, medium, high.
  • bare /thinking: sent to the model → thinking-level selector (Enter to select · Ctrl+S to set as default · Esc to cancel)
  • /efforts high, /reasoning, Shift+Tab: unchanged

Tracker

• packages/coding-agent/src/modes/interactive/changes.md: entry "Restore the /thinking interactive dispatch (/thinking is listed in autocomplete and /help but is sent to the model as a user message #1437)" with the split-setter mapping and the expected conflict zones (upstream carries the same methods with the { persist } option).
• packages/coding-agent/CHANGELOG.md: [Unreleased] > Fixed entry.

---

Summary by cubic

Restores interactive /thinking dispatch for #1437: /thinking now opens the selector and /thinking <level> changes the session level instead of sending the command to the model as user text. Ctrl+S in the selector remembers the level for the current model, while defaultThinkingLevel remains the fallback.

• Matches levels case-insensitively and reports available levels for invalid input.
• Adds autocomplete from the session’s supported thinking levels, including shared-host sessions.
• Adds regression coverage and updates the usage and settings documentation.

^{Written for commit fbd4406. Summary will update on new commits.}