Phase 4: Hosted server + deployment flip

[neuromechanist]

Part of epic #155.

Goal: Flip the hosted backend and deployment to Anthropic-Haiku by default.

Scope

• src/api/main.py: add an LLM_PROVIDER=anthropic branch mirroring the openrouter branch (main.py:437-500); env-driven Anthropic role models default to claude-haiku-4-5.
• Build the vision LLM from Claude (lift the OpenRouter-only vision guard at main.py:493).
• Maintainer AWS-billing deployment env (ANTHROPIC_API_KEY + ANTHROPIC_BASE_URL + ANTHROPIC_WORKSPACE_ID) in deploy/ and Docker/compose; do not hardcode the workspace id.
• Update DEPLOYMENT.md, .env.example, CI; cut a release.

Definition of Done

• Hosted API serves annotations on Anthropic-Haiku by default; LLM_PROVIDER=openrouter/ollama still selectable.
• Deployment docs + env templates updated; CI green.

[neuromechanist]

Status check for this phase: the code and ops work is done, two items remain and both
need an owner action rather than a commit.

Done in #162 (merged into develop 2026-08-19):

• The server builds every role LLM from the Anthropic factory with env-driven models
  defaulting to claude-haiku-4-5. There is no LLM_PROVIDER branch any more, because
  there is only one provider: a legacy LLM_PROVIDER=openrouter/ollama value is
  coerced to anthropic with a warning (src/api/main.py:328).
• The OpenRouter-only vision guard is gone; vision runs on Claude.
• ANTHROPIC_API_KEY / ANTHROPIC_BASE_URL / ANTHROPIC_WORKSPACE_ID are wired through
  docker-compose.yml, docker-compose.dev.yml, deploy/README.md, deploy/SECURITY.md,
  and .env.example, with no workspace id hardcoded anywhere.
• CI (test.yml, process-feedback.yml) reads the key from secrets and the base URL and
  workspace id from repository variables.
• DEPLOYMENT.md plus the new docs/deployment/claude-platform-aws.md cover setup and
  the migration path from an OpenRouter or Ollama deployment.

Added in #164: GET /metrics (token use, cost, and prompt-cache savings since
startup, server API key required) and HEDIT_PROMPT_CACHE_TTL, both documented in the
deployment guide and docs/prompt-caching.md.

Remaining, needs the repository owner:

1. Confirm the CI configuration from Migrate LLM stack to Anthropic Claude via Claude Platform on AWS #162's checklist: ANTHROPIC_API_KEY secret plus
   ANTHROPIC_BASE_URL and ANTHROPIC_WORKSPACE_ID variables. Worth verifying rather
   than assuming: integration tests skip silently when the key is absent, so a missing
   secret looks like a green run.
2. Delete the OPENROUTER_API_KEY_FOR_TESTING secret and repository variable (the
   variable exposes a key in plaintext) and revoke those OpenRouter keys.
3. Cut the release once Complete Phase 2: prompt-cache measurement, usage reporting, and migration cleanup #164 lands on develop.

[neuromechanist]

Item 1 is confirmed, by evidence rather than inspection.

The integration job on the #164 merge commit (a4f9bfe, run 32436174122) ran for real
against the live endpoint - the log shows "Running integration tests with Claude (Claude
Platform on AWS)..." rather than the "not set, skipping" branch - and came back
17 passed, 18 skipped, including both new prompt-caching tests:

tests/test_integration_anthropic.py::TestPromptCachingIntegration::test_cache_write_then_read_is_recorded PASSED
tests/test_integration_anthropic.py::TestPromptCachingIntegration::test_short_prefix_does_not_cache PASSED

Those two only pass if a real key reached a real workspace endpoint, so the
ANTHROPIC_API_KEY secret and the ANTHROPIC_BASE_URL / ANTHROPIC_WORKSPACE_ID
variables are all present and correct. CI is also now exercising the cache-hit path on
every push that touches agents, validation, or the Anthropic factory.

Still open on this issue:

1. Confirm the CI credentials - done, see above.
2. Delete the OPENROUTER_API_KEY_FOR_TESTING secret and repository variable (the
   variable exposes a key in plaintext), then revoke those OpenRouter keys. Needs the
   owner: repository secrets and variables are not readable or writable with the token
   available to tooling here (HTTP 403).
3. Cut the release.