feat(copilot): Anthropic Fast Mode for opus/sonnet/haiku 4.x + synthetic -fast model ids

[arminanton]

feat(copilot): Anthropic Fast Mode for opus/sonnet/haiku 4.x + synthetic -fast model ids

Unlock GitHub Copilot's Anthropic Fast Mode (extra_body.speed="fast", ~2.5x
output throughput) for the opus/sonnet/haiku 4.x families, make it selectable
three ways, and fix the TUI status-bar label so "fast" isn't printed twice.

Background

Native Anthropic only honors speed=fast on Opus 4.6 (4.7/4.8 explicitly 400).
But GitHub Copilot's /v1/messages proxy (api.githubcopilot.com) accepts
speed=fast on the whole opus/sonnet/haiku 4.x line and passes it upstream
(empirically verified). The previous gate was hard-coded to Opus-4.6-only, so
Copilot users couldn't reach Fast Mode on the newer models at all.

What changed

(a) Endpoint-aware gate

• agent/anthropic_adapter._supports_fast_mode now takes base_url. Native
  Anthropic stays Opus-4.6-only; Copilot's proxy returns True for opus/sonnet/
  haiku 4.x. A small _is_copilot_base_url(base_url) helper centralizes the
  api.githubcopilot.com check.
• hermes_cli.models._is_anthropic_fast_model widened to the same families so
  the /fast toggle and resolve_fast_mode_overrides apply to them.

(b) Auto-knob — a -fast model id just works

• normalize_copilot_model_id strips a trailing "-fast" to the real base id on
  the wire (claude-opus-4.8-fast -> claude-opus-4.8); the speed param is
  attached separately by the adapter.
• selecting a -fast id is an implicit /fast toggle (sets request_overrides
  even when the toggle is off).

(c) Visibility — -fast variants in the catalog

• fetch_github_model_catalog injects synthetic "-fast" companion
  entries (name " (Fast)") for fast-capable Claude models, so the Fast
  variant is selectable directly from the model picker (the way OpenRouter
  exposes …-fast). Gated off by HERMES_COPILOT_HIDE_FAST_VARIANTS.

(d) TUI label fix

• ui-tui appChrome.shortModelLabel strips a trailing "-fast"/"_fast" suffix so
  the model name is the base model and fast mode is shown only by the badge.
  The label is now identical whether fast came from the -fast id or the /fast
  toggle ("opus 4.8 fast" instead of "opus 4.8 fast fast"). modelLabel is
  exported for the regression test.

Tests

• tests/cli/test_fast_command.py: endpoint-aware policy + TestSyntheticFastVariant
  (normalize, override resolution, catalog injection, copilot-yes/native-no
  adapter gates).
• tests/hermes_cli/test_model_validation.py: probe-catalog assertion includes a
  -fast variant.
• ui-tui appChromeModelLabel.test.ts: double-print regression + toggle/suffix
  parity + fast-off + sonnet/haiku variants.

Relationship to #49184

Fast Mode is only reachable because Claude-on-Copilot is routed to /v1/messages
(#49184). This PR is self-contained (it carries the small _is_copilot_base_url
helper and the catalog-injection logic it needs), but it is best reviewed/merged
after #49184 since they share the Copilot Anthropic-proxy surface.

Wire-model strip (the load-bearing fix)

build_anthropic_kwargs now strips a trailing "-fast" from the model id it puts
on the wire (new _strip_fast_suffix helper, reused by _supports_fast_mode). The
"-fast" id is the Fast Mode KNOB, not a real model — sending it verbatim 400s
"model not supported". Selecting claude-opus-4.8-fast in the TUI hit exactly this
(the picker's per-session model override reaches the adapter without the catalog
normalizer). Stripping at the adapter boundary fixes every path regardless of how
the id was selected. Regression test added (test_fast_suffix_stripped_from_wire_model).

[teknium1]

Thanks for mapping the adapter, picker, and label paths. The Copilot-specific fast-mode gap remains on current main, but this draft needs routing and surface-completeness work before it can safely provide the advertised behavior.

Problems

• The new adapter path is unreachable for Copilot Claude on current main: hermes_cli/runtime_provider.py:320-335 derives Copilot mode through copilot_model_api_mode, whose non-GPT path is chat_completions (hermes_cli/models.py:3293-3300); the changed adapter is reached through anthropic_messages (agent/transports/anthropic.py:41-78). #49184 is still open.
• The implicit -fast override is added only in hermes_cli/cli_agent_setup_mixin.py; gateway turns return {} when /fast is off (gateway/run.py:3924-3933), and TUI creation only enables fast when params.fast is supplied (tui_gateway/server.py:5225-5228).
• The unconditional wire stripping conflicts with main’s separate anthropic/claude-opus-4.8-fast catalog/pricing entry (hermes_cli/models.py:44, agent/usage_pricing.py:154-180). Scope the transform to Copilot synthetic IDs.
• Please replace the new HERMES_COPILOT_HIDE_FAST_VARIANTS setting with config.yaml, per AGENTS.md:102-107, and avoid the exact catalog snapshot assertion in tests/hermes_cli/test_model_validation.py.

Automated hermes-sweeper review.

[teknium1]

This strips -fast for every Anthropic-Messages call, but main treats anthropic/claude-opus-4.8-fast as a distinct real OpenRouter model (hermes_cli/models.py:44; agent/usage_pricing.py:172-180). Restrict stripping to a positively identified Copilot synthetic variant.

[teknium1]

This implements suffix-as-fast only for the classic CLI. Gateway resolution still returns no override when /fast is off (gateway/run.py:3924-3933), and the TUI create path only enables fast from params.fast (tui_gateway/server.py:5225-5228); centralize this behavior before advertising picker support.

[teknium1]

HERMES_COPILOT_HIDE_FAST_VARIANTS is a new user-facing non-secret HERMES_* behavior switch. AGENTS.md:102-107 requires behavioral configuration to live in config.yaml; use a config setting if this opt-out is needed.

[teknium1]

Avoid expanding this exact catalog snapshot. Assert the relationship instead: the returned Claude item has exactly one -fast companion while non-Claude entries do not, so routine catalog changes do not break the test.

[arminanton]

Closing this.

Re-verified against Copilot's live /v1/messages catalog and current upstream main: on Copilot, fast mode is a real entitled catalog model id (claude-opus-4.8-fast, billed 2×), selected by sending that id verbatim — the inverse of native Anthropic (where you strip -fast and attach speed:"fast").

This PR applies the native-Anthropic mechanism to Copilot: it strips -fast on the wire, attaches an inert speed param, and synthesizes <family>-fast companions (claude-sonnet-*-fast, claude-haiku-*-fast) that don't exist in the Copilot catalog and 400 with model_not_supported. That would actively break the working path.

Current main already does the right thing by default:

• claude-opus-4.8-fast is a real catalog + pricing entry — added and merged in feat: add claude-opus-4.8 and claude-opus-4.8-fast #34003 (hermes_cli/models.py:89, agent/usage_pricing.py:215-222, 2× billing).
• Non-GPT Copilot models return the id verbatim, so an entitled -fast id already reaches the proxy unchanged and bills correctly.

Since the model this PR targets is already merged and the mechanism here is counter-productive, there's nothing left to salvage. Anything still worth doing (a guard ensuring -fast is never stripped/mangled on the Copilot path) is a much smaller, separate change; I'll open a focused PR if a real gap remains rather than rework this one.

Thanks for the review.