Story-3141 Walkthrough

Outcome: every httpx-backed inference route now crosses one shared final-request hook. It removes the two attribution headers, overwrites the framework fingerprint with the single PROVIDER_USER_AGENT constant, and keeps the existing W3C baggage strip.

agent or eval builds its provider model
  → shared egress_http_client()
  → provider SDK adds its request headers
  → sanitize_provider_egress() runs last
  → User-Agent = PROVIDER_USER_AGENT
  → X-Title / HTTP-Referer / baggage removed
  → x-session-id and authorization preserved
  → request leaves for the provider

Boundary. This removes stack fingerprinting and casual attribution. It does not make PRAMAAN anonymous: the paid API key and egress IP still identify the account to the provider.

One change point

agents/_shared/provider_egress.py owns both the hook and the selected UA value. If the founder selects a different value, only the named constant changes; providers and tests import the same definition.

Fleet coverage

Surface	Covered paths
Atlas	OpenRouter gateway
Vani, Vega, context-judge	OpenRouter gateway + provider-direct
Blake, Indigo	OpenRouter gateway + provider-direct
Evals	Indigo bench clients + Blake judge client

Executable proof

The root fleet guard constructs all 11 real agent HTTP provider branches, reaches each model's actual registered httpx hook, and feeds it the old SDK UA, attribution headers, baggage, cache-affinity header, and authorization header. A discovery assertion forces a future agent provider into the roster, while an AST guard rejects raw AsyncOpenAI clients that bypass the egress client.

uv run ruff check .                                      → All checks passed
uv run pytest -q                                         → 457 passed
atlas / context_judge / framer package suites            → 516 / 58 / 164 passed
indigo / mira / registrar package suites                 → 326 / 320 / 516 passed