feat(agent): bounded auto and cold fast modes

[100yenadmin]

What does this PR do?

Adds two time-bounded fast-mode policies on top of the existing normal/priority modes:

• auto — fast mode is on for the first N seconds of a session's activity (default 60s), then relaxes;
• cold — fast mode on for a cold start (no prior session activity), off once the session is warm.

The policy is revalidated at every provider-dispatch boundary, so a mode change (or the window closing) takes effect on the next request and can't leak stale fast-mode settings across turns. It's inert by default: for any service_tier outside {auto, cold} the dispatch path is byte-identical to today.

Stacked PR — read this

This branch has three commits, the first two of which are already open as standalone PRs:

1. refactor(anthropic): extract idempotent fast-mode kwargs helper → refactor(anthropic): extract idempotent fast-mode kwargs helper #89982
2. fix(models): fail closed on unverified fast-mode endpoints → fix(models): fail closed on unverified fast-mode endpoints #89960
3. feat(agent): bounded auto and cold fast modes → this PR's actual contribution

The feature needs both (it calls the idempotent helper to apply/revoke per turn, and the endpoint verifier to stay fail-closed). Simplest path: take #89960 and #89982 first (they stand on their own — a fix and a refactor), and this rebases down to the single feature commit. I can rebase it to one commit on request. This is the last piece of the original #74730, split for reviewability.

Related Issue

Completes #64785 / #74730 (minus the user-facing surface — see below).

Type of Change

• ✨ New feature (non-breaking change that adds functionality)

Changes Made (feature commit only)

• agent/fast_mode.py (new) — the policy module: normalize_fast_auto_on_seconds (with a bool-is-not-a-float guard), has_prior_session_activity, begin_fast_mode_turn / invalidate_fast_mode_turn, effective_request_overrides, revalidate_fast_mode_request.
• agent/agent_init.py, run_agent.py — fast_auto_on_seconds param + normalization on the agent.
• agent/conversation_loop.py — begin_fast_mode_turn at run ingress.
• agent/chat_completion_helpers.py — revalidate at the provider-dispatch boundaries; build_api_kwargs reads the effective overrides.
• Tests: 27 unit (tests/agent/test_fast_mode_auto.py) + 5 dispatch-revalidation integration (tests/run_agent/test_run_agent.py).

Now includes the config/CLI surface

Updated: this PR now also exposes the modes on the generic surface (a 4th commit, da7bc56750), so the feature is actually usable, not just wired:

• agent.service_tier accepts auto/cold (alongside normal/fast/priority) in config, CLI, and the gateway.
• agent.fast_auto_on_seconds (default 60) is a config key threaded to the agent.
• /fast [normal|fast|auto|cold|status] works across the CLI, the messaging gateway, and the TUI/desktop backend gateway.

The one piece still deferred is the desktop model-settings menu UI (ui-tui / apps/desktop) — a small follow-up that doesn't affect headless, CLI, or gateway users. Surface tests: tests/cli/test_fast_command.py + tests/gateway/test_fast_command.py + tests/tui_gateway/test_fast_session_scope.py → 36 passed; a 546-test regression sweep across tui_gateway/gateway/cli is green.

How to Test

bash scripts/run_tests.sh tests/agent/test_fast_mode_auto.py tests/run_agent/test_run_agent.py -q
→ 27/27 fast-mode unit + the dispatch-revalidation integration green

(One unrelated red on this machine, test_interruptible_anthropic_interrupt_never_closes_shared_client, is a missing anthropic SDK in the venv — it fails identically at the base commit.)

test_normal_and_fast_modes_remain_backwards_compatible asserts the normal/priority path is unchanged.

Checklist

Code

• I've read the Contributing Guide
• Conventional Commits
• Searched existing PRs
• My PR contains only changes related to this feature (the two prerequisite commits are their own open PRs)
• Ran the fast-mode + run_agent suites
• Added tests
• Tested on my platform: macOS 15

Documentation & Housekeeping

• Docs — deferred with the user-facing surface (nothing user-visible in this PR)
• cli-config.yaml.example — deferred (no config key exposed yet)
• CONTRIBUTING.md/AGENTS.md — N/A
• Cross-platform — N/A
• Tool descriptions/schemas — N/A

---

Salvage note: the feature is one commit (ec09a3af98) on top of #89960 + #89982 — cherry-pick the prerequisites, then this.

[100yenadmin]

Folded the config/CLI surface into this PR (commit da7bc56750) so the feature is complete and usable rather than wired-but-unreachable: service_tier: auto/cold and fast_auto_on_seconds are now settable via config, the CLI, and the gateway /fast command, all with tests. Only the desktop model-settings menu UI remains a follow-up. The PR is now 4 commits — the two prerequisites (#89960, #89982), the policy+wiring, and this surface.

[100yenadmin]

Fixed both real CI failures here (87914a66dd):

1. NameError: CLI_CONFIG — the surface commit read CLI_CONFIG in _init_agent without the function-local import that file uses elsewhere, so agent init failed whenever that path ran (surfaced as test_runtime_resolution_failure_is_not_sticky). Now behind a small _fast_auto_on_seconds_setting() helper that does the local import and falls back to the default.
2. test_bare_fast_sends_picker_when_adapter_supports_it — that test pins the /fast picker to ["fast", "normal"]; this PR intentionally adds auto and cold, so the assertion is updated to match the new choice list. Behavior change is the point of the PR, not a regression.

Green after the fix: 55 passed across test_choice_picker.py, test_cli_provider_resolution.py, test_fast_command.py (cli + gateway), and test_fast_session_scope.py.

The remaining red check (test_image_generation.py::test_upscale_defaults_are_all_off) is the current main-is-red issue #90013 — unrelated to this PR and already has fixes in flight.