fix(providers): use chat completions for all Actual routes

[somewheresy]

Summary

Use /v1/chat/completions for every Actual generation path, including foreground chat, compaction, titles, async tasks, startup resolution, model switches, fallback chains, session restoration, and credential rotation. Store Actual provider settings in config.yaml and credentials in .env.

Motivation

The transport switch left multiple ways to select Responses: auxiliary wrappers, the hosted hostname mapping, persisted task/custom-provider modes, aliases, and runtime transitions. Bare endpoint overrides could also lose /v1. Environment-backed provider settings could override YAML during setup, credential resolution, or live key reloads.

Links

• E-1047.
• E-1005: scheduled compatibility detector.
• Supersedes the Responses-only approach in fix(providers): restore Responses-API routing for Actual Computer #80248 and duplicate fix(agent): restore the Actual Computer Responses-API routing branch #80266.
• Original provider integration: feat(providers): Actual Computer inference provider (salvage of #26491) #79644.

Approach

• Recognize Actual by its provider aliases or the exact api.actual.inc hostname and enforce Chat Completions when resolving and changing runtimes.
• Override stale Responses settings in Actual auxiliary tasks and custom providers targeting the hosted relay. Auxiliary overrides sharing the active Actual daemon also retain its protocol. Normalize bare Actual endpoints to /v1.
• Guard the main Responses send site and the sync/async auxiliary adapter: a known Actual route raises before HTTP even if stale runtime state or a wrapper bypasses resolution. Preserve Actual identity on resolved auxiliary clients.
• Remove the Actual-specific branch from the Responses transport. Preserve reasoning content, tool-turn replay, supported effort clamping, and one-shot --reasoning propagation.
• Save endpoint selections in model.base_url. Prefer YAML during setup, discovery, credential-pool resolution, and live API-key reloads. Keep ACTUAL_BASE_URL as a legacy fallback; credentials remain in .env.
• Preserve the hosted macOS TLS default and explicit TLS overrides.
• Rebase onto upstream main at 872bafd58d.

Reviewer Focus

The HTTP harness in tests/agent/test_actual_auxiliary_routing.py exercises the real SDK, title generator, compressor, sync/async calls, provider resolution, and live agent transitions. A local server rejects inference routes other than /v1/chat/completions; hosted-route cases resolve api.actual.inc to that server. Coverage includes aliases, custom routes, keyed/keyless local endpoints, stale configuration, bare URLs, unsupported-model fallback, startup fallback, restored sessions, credential rotation, and live key reloads.

Regression cases reproduced the old alias/custom routing, startup fallback, and credential-rotation failures before their fixes. 1,900 tests pass across 109 files, including 121 Actual routing/configuration cases, after rebasing. The 18 forced-Responses cases reproduced outbound HTTP before the send guards and now verify that no HTTP request is sent; normal Actual requests still complete successfully. The suite also exercises the supported Codex Responses paths. Repository Ruff checks and the plugin-compatibility check pass. The routing/configuration review found no remaining material issues. Validation uses local HTTP integration tests; the live hosted Actual service was not exercised in this revision.

GitHub CI, Nix, and Docker workflows at commit 8a6b5b67a7 require upstream workflow approval before running. Current CI run. The PR has no merge conflicts.

Failure Modes

• Stale title work is cancelled before any HTTP request.
• Old Actual Responses settings cannot select /v1/responses, including after startup fallback, a model switch, or session restoration. Forced entry into the Responses transport raises a clear error before any HTTP request.
• Unsupported-model recovery sends sync and async fallback requests through Chat Completions.
• YAML endpoint selections survive stale environment URLs and API-key reloads. Invalid setup URL input preserves the configured endpoint.
• Exact hostname checks reject lookalike hosts and URL-path spoofing.
• Missing reasoning preserves final content; unsupported reasoning levels are clamped to Actual's supported vocabulary.

Breaking Changes

Actual routes now override saved Responses transport selections, including custom configurations targeting api.actual.inc. No config migration is required. ACTUAL_API_MODE was introduced only in an earlier revision of this unmerged PR; existing ACTUAL_BASE_URL settings remain a fallback.

Post-Merge Behavior

Merging triggers upstream CI and the Docker build/publish workflow. Existing Hermes processes receive the fix after updating and restarting. No Actual service deployment or new runtime flag is required. Release smoke: run hermes -z '<prompt>' --provider actual --model <model> --reasoning high, generate a title, and compact a conversation; verify outgoing generation requests use /v1/chat/completions.

Diagrams

flowchart LR
    Config[config.yaml provider and endpoint] --> Runtime[Actual route resolution]
    Key[.env API key] --> Runtime
    Legacy[Legacy URL fallback] --> Runtime
    Task[Auxiliary task or custom provider] --> Runtime
    Runtime --> Chat[Chat and startup]
    Runtime --> Aux[Titles, compaction, sync and async tasks]
    Runtime --> Changes[Switch, fallback, restore, key rotation]
    Chat --> API[POST /v1/chat/completions]
    Aux --> API
    Changes --> API

Loading