feat: runtime model switching via /model command

[Talla]

Summary

Adds a /model command that lets users switch between Claude models (Opus, Sonnet, Haiku) and set effort levels at runtime via an inline keyboard UI — no bot restart or config changes needed.

• Inline keyboard flow: /model → pick model → pick effort level (if supported) → done
• Model-aware effort levels: Haiku has no effort support, Sonnet supports low/medium/high, Opus adds "max"
• Session handling: Switching models forces a new session (CLI doesn't support model changes on resumed sessions). Users are warned upfront before selecting.
• Per-user override: Stored in context.user_data (in-memory). Resets on bot restart — server default (CLAUDE_MODEL env var) is used until the user overrides.

Implements the feature requested in #138.

Changes

File	What
command.py	/model command, model_callback, _handle_model_selection shared logic, model/effort constants
callback.py	Route model: and effort: callbacks in classic mode
orchestrator.py	Register command + callback in both modes, add to bot menu, thread overrides to run_command
message.py	Thread model_override + effort_override through all 4 classic run_command call sites
facade.py	Accept and forward model_override + effort_override params
sdk_integration.py	Apply model_override and effort to ClaudeAgentOptions
test_model_command.py	17 new tests (keyboard, selection, labels, effort config)
test_orchestrator.py	Updated handler count assertions

Test plan

• All 529 tests pass (512 existing + 17 new)
• Manual: /model shows keyboard with Opus/Sonnet/Haiku + Reset
• Manual: Selecting Haiku skips effort selection
• Manual: Selecting Sonnet shows Low/Medium/High (no Max)
• Manual: Selecting Opus shows Low/Medium/High/Max
• Manual: After switching, next message uses the new model
• Manual: "Reset to default" clears override

Closes #138

🤖 Generated with Claude Code

[RichardAtCT]

Good implementation overall — the UX flow is clean and the test coverage is solid. Six issues to address before merging:

---

🔴 Bug: action closure in _model_effort_handler (callback.py)

This is the most critical issue. _model_effort_handler references action from the outer scope, but action is a loop variable (or however handlers is constructed). Both "model" and "effort" entries point to the same closure, and action will be whatever value it held last when the dict was defined — likely "effort" for both handlers. This means a model:opus callback will be reconstructed as effort:opus and silently fail.

Fix by capturing explicitly:

async def _make_model_effort_handler(action: str):
    async def handler(query, param, context) -> None:
        await _handle_model_selection(query, f"{action}:{param}", context)
    return handler

Or just inline two named functions — cleaner and avoids the factory pattern for async.

---

🔴 Missing type annotations (mypy strict)

_handle_model_selection is missing return type and parameter annotations. With disallow_untyped_defs=true this will fail CI. Should be:

async def _handle_model_selection(
    query: CallbackQuery,
    data: str,
    context: ContextTypes.DEFAULT_TYPE,
) -> None:

Same applies to model_command — update and context need annotations.

---

🟡 effort_override not visibly applied in sdk_integration.py

The diff shows model_override wired into ClaudeAgentOptions but effort_override is elided. If it's genuinely there in the full diff, fine — but if not, this is a silent no-op: users select an effort level, get confirmation text, but the SDK ignores it. Needs explicit confirmation.

---

🟡 Hardcoded model IDs will go stale

"haiku": "claude-haiku-4-5-20251001" is already date-versioned and will need updating as Anthropic releases new versions. Consider moving _MODELS to config/env so it's overridable without a deploy, or at minimum add a comment linking to the Anthropic model docs.

---

🟡 user_data vs chat_data — potential conflict with PR #165

PR #165 migrates session state to chat_data. This PR stores model/effort overrides in user_data. In a group chat, one user's model choice would only affect their user_data while the session (in chat_data) is shared — meaning the model applied depends on who sent the last message. These two PRs need to be reconciled before either merges.

---

🟢 Test quality — mostly good

17 tests covering happy path, reset, Haiku (no effort), and effort selection is solid. Two suggestions:

• Add a test for the action closure bug (would have caught it)
• Add a test asserting force_new_session is set on model change

---

Summary: The action closure bug is a functional regression that will silently swallow model selections. The mypy annotations are a CI blocker under strict mode. Everything else is addressable before or alongside merge.

— Friday, AI assistant to @RichardAtCT (posted as @RichardAtCT — FridayOpenClawBot access pending)

[RichardAtCT]

This is the preferred /model implementation — closing #139 in its favor. The inline keyboard UI and per-model effort levels are a nice touch.

Action needed: The branch is currently in CONFLICTING state. Please rebase on main (we're merging several PRs that touch orchestrator.py). Once rebased with passing tests, this should be ready to merge.

Note: #139 has been closed as superseded by this PR.

[Talla]

Thanks for the detailed review! All points addressed in this force-push. Here's what was done:

---

🔴 Bug: action closure in _model_effort_handler
Fixed. Replaced the shared closure with two explicit lambdas, each hardcoding its own prefix:

"model": lambda q, p, ctx: _handle_model_selection(q, f"model:{p}", ctx),
"effort": lambda q, p, ctx: _handle_model_selection(q, f"effort:{p}", ctx),

Also moved the import to module level. Added a regression test that would have caught this.

🔴 Missing type annotations
Fixed. _handle_model_selection now has query: CallbackQuery, data: str, context: ContextTypes.DEFAULT_TYPE.

🟡 effort_override not visibly applied in sdk_integration.py
It was wired (effort=effort_override in ClaudeAgentOptions) — just not visible in the diff due to context. Added an explicit log line to make it obvious:

if model_override or effort_override:
    logger.info("Runtime model/effort override active", model_override=..., effort_override=...)

🟡 Hardcoded model IDs will go stale
Went further than a comment. Dropped _MODELS entirely and replaced with short CLI aliases ("opus", "sonnet", "haiku"), which the Claude CLI resolves to the current latest of each family automatically — no version numbers to maintain anywhere. Confirmed these are accepted by the CLI (--model flag docs) and the SDK types (Literal["sonnet", "opus", "haiku"]).

🟡 user_data vs chat_data — conflict with PR #165
PR #165 is not merged yet and our PR should go in independently. model_override and effort_override intentionally stay in user_data (per-user model preference makes sense semantically). force_new_session also stays in user_data to be consistent with current main. Added a comment flagging the dependency:

# Note: if PR #165 merges first, change this to context.chat_data
context.user_data["force_new_session"] = True

When #165 merges, that one-line change will be needed here too.

Tests
Added three new tests:

• Closure regression guard: asserts model:sonnet sets model_override, not effort_override
• Effort prefix isolation: asserts effort:high doesn't overwrite model_override
• force_new_session assertion on model switch

Rebase
Rebased onto current main (10 commits ahead). Conflicts resolved in sdk_integration.py, facade.py, orchestrator.py, and test_orchestrator.py — all tests pass (552 total).