Per-session reasoning effort for Claude and Codex

[ryuhaneul]

Summary

This brings reasoning effort to parity with model selection: effort is resolved per session — each forum topic can run at its own effort, and the main chat / DM sets the global default — for both providers (Claude --effort, Codex -c model_reasoning_effort).

The two providers started from very different places, so this is a larger change than "make a setting per-session":

• Claude had no reasoning-effort support at all. claude_provider never emitted --effort and ignored reasoning_effort entirely. This PR adds Claude reasoning effort for the first time — the --effort flag, the /effort command + selector, provider-aware validation, and /status display.
• Codex already consumed the global config.reasoning_effort (-c model_reasoning_effort), but only as a single service-wide default: there was no command to change it at runtime and no per-conversation scope, so changing effort never actually took effect in a given topic. This PR makes a topic's /effort apply to Codex in that topic.

After this change:

• changing effort in a forum topic affects only that topic (Claude and Codex);
• changing it in the main chat / DM updates the global default.

Starting point (v0.18.2)

• config.reasoning_effort exists as one global value (default medium).
• codex_provider reads it → service-wide only; no runtime command, no per-topic scope.
• claude_provider does not read effort at all.
• There is no /effort command and no per-session effort field (only model is per-session).

What changed

• Claude provider — claude_provider now emits --effort <value> (skipped only when the value is default). First-class reasoning effort for Claude.
• /effort command + selector — new command and inline selector in model_selector, offering the efforts valid for the active session's provider/model.
• SessionData.reasoning_effort — new per-session field, captured when a session is created and preserved afterward, mirroring how model is captured (preserve_existing_target). A session keeps the effort it was created with instead of drifting when the global default later changes.
• Per-turn resolution — a reasoning_effort_override is threaded through every turn path (normal flow, named foreground/background sessions, injection, task hub, webhook observer, cron) and resolved in _make_cli as effort = request.effort_override or config.reasoning_effort, exactly mirroring model_override or default_model. Both providers read the single resolved value, so Claude and Codex alike honor per-session effort.
• Effort selection scope — effort is written to the global default only from the main chat / DM; in a topic it is stored on that topic's session. A provider switch re-validates the carried-over effort and resets an unsupported value.
• Named / background carriers — NamedSession, BackgroundSubmit, BackgroundTask, TaskSubmit, TaskEntry now carry reasoning_effort, so named foreground/background runs keep their session's effort.
• /status — shows the effective effort (the session's value in a topic, otherwise the global default).
• Effort selection response — picking an effort echoes the applied value (e.g. Reasoning effort set to high) across all locales instead of a generic "updated" line, matching how model selection confirms the chosen model.
• cron / webhook executors emit the resolved effort as well.
• Coupled /model alignment fix — switch_model shares its resolution path with per-session effort, and this PR also makes it compare the requested model against the active session's model in the main chat / DM, not only in topics (the main chat previously compared against config.model). This lets /model re-align a session whose model has drifted from config.model, and realigns config.provider to the new model's provider. It only has a visible effect when config.model has been changed out of band so that it differs from the active session — e.g. editing config.json directly, or the per-agent config hot-reload used to switch a sub-agent's model. Under normal in-chat /model use the active session and the global default always move together, so this is a no-op. Flagged explicitly as it is a /model behavior change rather than an effort one; it can be split into a separate PR on request.

Provider-aware validation

Effort values are validated against the selected provider's supported set — Claude low/medium/high/xhigh/max; Codex's set is discovered per model, falling back to low/medium/high/xhigh. An effort that isn't valid for a newly selected model/provider is reset rather than passed through.

Tests

New and extended coverage across the model selector, orchestrator core, flows, commands, named-session recovery, tasks, cron execution, the Claude provider, and the param resolver (the model-selector suite alone adds ~830 lines). Full suite is green with no new failures.

Note for merging alongside #156 (interactive REPL provider)

Both this PR and #156 modify the single per-request config builder CLIService._make_cli, so they conflict there. Effort is resolved once in _make_cli (effort = request.effort_override or config.reasoning_effort) and stored on the provider as self._config.reasoning_effort; every execution path reads it from there — the streaming path (_build_command) and #156's _send_interactive alike. When both land, resolve the _make_cli conflict by keeping reasoning_effort=effort (the per-session value) and adding #156's interactive_repl_pool / interactive_enabled fields. Note that _make_cli is shared by both paths, so keeping the global reasoning_effort=self._config.reasoning_effort instead would silently drop per-session effort everywhere.

A pooled interactive REPL (#156) is spawned once and reused, so it keeps the effort it was spawned with. #156 already drops the pooled REPL on a model change (switch_model → kill_interactive_repl); an effort change needs the same, otherwise a topic's REPL keeps answering at the old effort until it happens to respawn. This PR adds that drop in the effort handler through a small getattr-guarded helper — a no-op on this branch (which targets a base without #156) and an active REPL-respawn once #156 lands. Changing effort in a topic running interactive then respawns its REPL at the new effort immediately.

Supersedes #161, which added Claude reasoning effort but applied it globally.