v0.9.0: Proxy contract consolidation

[0.9.0] — 2026-08-09

Forge's Proxy has become its most-used integration surface while backend
support has expanded organically. Forge 0.9 consolidates that growth around a
clear sidecar contract, replacing accumulated configuration and metadata seams
before they become a long-term maintenance burden. Guardrail behavior and
native workflows remain intentionally stable; the breaking changes are
concentrated in the Proxy contract. See
Migrating to Forge 0.9 for the complete
old/new/action table.

Breaking — ordinary Proxy deployments

• External Anthropic selection moves from removed backend_protocol /
  --backend-protocol to backend="anthropic" / --backend anthropic.
  Generic external omission remains OpenAI-compatible, with explicit
  backend="openai" also supported.
• /health now forwards backend readiness; Forge liveness is
  /forge/health. /v1/models is the backend's honest catalog rather than a
  synthesized Forge one-model list, and public catalog reads never mutate
  private vLLM routing identity.
• Response model reports the actual pinned, discovered, configured, or
  request-routed wire identity.
• Proxy is unconditionally no-compaction and never budget-compacts or deletes
  caller history. Missing context-window metadata no longer blocks unmanaged
  inference or managed budget_mode=backend.
• Unmanaged startup performs no metadata query. Static-credential unpinned
  vLLM identity discovery moves to first inference and remains retryable.

Added

• Closed read-only forwarding for /health, /v1/health, /v1/models,
  /models, and /props, applying the same one-backend-credential rule as
  inference while preserving exact path/query, backend status/body/content
  type, recalculated length, and Forge CORS; transport failures map to 502.
• Local /forge/health liveness and /forge/usage, which exposes one
  last-completed eligible process-local context snapshot or 204. Session
  attribution prefers a non-empty Claude Code session header over a non-empty
  string LiteLLM body ID; subagent requests do not replace the snapshot.
• Canonical unmanaged openai and anthropic backend profile selectors.

Breaking — specialized valid configurations

• Unmanaged budget_tokens is a reporting denominator only. Managed
  backend, manual, forge-full, and forge-fast remain allocation modes,
  distinct from preserved native compaction.
• Explicit unmanaged/Ollama ports replace the URL authority port while keeping
  normalized prefixes. CLI --extra-flags is now the terminal argv remainder.
• vLLM served identity and reporting window are independent. A model ID without
  max_model_len permits inference; a pin settles identity only, while an
  explicit budget settles reporting only. Clean vLLM passthrough/raw tools and
  tool choice now reach the backend instead of being accepted and dropped.
• litellm_session_id is preserved through applicable conversion/adapters.
• Constructor default introspection may observe omission as None rather than
  old concrete values.

Breaking — direct Python and low-value invalid inputs

• Removed the exported combined discover_backend_metadata() method. Direct
  consumers use get_served_model_name() for vLLM identity and
  get_context_length() for an honest window. vLLM exact-matches the selected
  model; Anthropic no longer fabricates 200K/8192 context defaults.
• Native context accounting prefers backend usage. After compaction invalidates
  that observation, CompactEvent.tokens_after reports a fresh character/4
  heuristic until another backend result supplies usage.
• Profile-invalid identity fields, unmanaged budget_mode, ineffective
  Ollama/unmanaged extra_flags, conflicting managed flags, contradictory
  serialization switches, invalid manual budgets, and negative limits now fail
  before side effects. Python raises ValueError; CLI validation exits 2.

Preserved boundaries

• Proxy keeps flat CLI and ProxyServer configuration surfaces, inference
  aliases, global CORS, closed unknown routes, retry/exhaustion behavior,
  credential isolation, streaming shape, managed ownership/shutdown, and
  managed/unmanaged serialization defaults.
• Native setup_backend(), ContextManager, WorkflowRunner, and built-in or
  custom compaction remain supported and distinct from Proxy no-compaction.
• The optional Anthropic integration now has an anthropic>=0.86.0 floor.