Changelog

All user-visible changes to InferBridge (formerly Agni AI) land here. Dates are UTC.

v0.2.0 — 2026-04-23

Product rename: Agni AI → InferBridge. The rename is positioning only; behaviour is unchanged. This release ships the new names everywhere and backwards-compatible aliases for the old ones. One surface — the attached response metadata block — is a breaking change; see below.

Breaking change

• The non-standard response metadata block is now keyed inferbridge (was agni). If your code reads response.json()["agni"], rename that lookup to response.json()["inferbridge"]. The block’s contents are unchanged:

  "inferbridge": {
    "provider": "openai",
    "model":    "gpt-4o-mini",
    "mode":     "ib/balanced",
    "cache_hit": false,
    "latency_ms": 734,
    "cost_usd": "0.000142",
    "residency_actual": "global",
    "request_id": "..."
  }

  No alias is shipped for this field — an alias would inflate every response payload permanently, and the fix for callers is a one-line string change.

Backwards-compatible renames (deprecated, removal 2026-07-22)

All of these keep working until 2026-07-22. Structured warnings are logged (logger inferbridge.aliases) whenever the legacy form is used, so you can grep your ops stream to find remaining callers before the removal date.

• API-key prefix. Newly-issued keys are prefixed ib_; keys issued before the rename (prefixed agni_) authenticate indefinitely, since they remain valid until the user rotates them.

• Tier names. The canonical modes are ib/cheap, ib/balanced, ib/premium. Incoming agni/* mode strings are normalized to their ib/* equivalent at the chat endpoint.

• Request headers (legacy → canonical):

  • X-Agni-Residency → X-InferBridge-Residency
  • X-Agni-Override-Model → X-InferBridge-Override-Model
  • X-Agni-Cache → X-InferBridge-Cache
  • X-Agni-Cache-TTL → X-InferBridge-Cache-TTL
  • X-Agni-Timeout → X-InferBridge-Timeout

  If both the new and legacy header are sent, the new header wins.

• Environment variable. AGNI_ENCRYPTION_KEY → INFERBRIDGE_ENCRYPTION_KEY. Both names are read at startup; the new name wins if both are set, the old name logs a deprecation warning and is used as a fallback.

• Python package. The distribution is now named inferbridge (was agni) in pyproject.toml. The in-repo module path stays app/.

Unchanged on purpose

• Database schema. No column or table renames. The users.agni_key_hash column retains its Agni-era name — a column rename would require a schema migration against a production DB, and the column name is private to the server. The column stores SHA-256 hashes of whatever key prefix the user registered with (ib_ or legacy agni_); neither the prefix nor the rename changes the hash format.
• Git repository and Railway URL. Unchanged. A custom inferbridge.* domain will be added in a follow-up; the Railway hostname stays as a stable secondary for existing integrations.

Removal calendar (2026-07-22)

On that date a follow-up PR removes, in a single commit:

• The app/core/aliases.py shim module.
• The x_agni_* Header(...) parameters on POST /v1/chat/completions.
• The AGNI_ENCRYPTION_KEY field and reconciliation validator on Settings.
• The legacy agni_ branch in has_valid_key_prefix (pre-rename keys remain valid — only the prefix check goes; the hash comparison is unchanged).

Expect the 2026-07-22 change to bump us to v0.3.0.

---

v0.1.0 — 2026-04-21

First public release under the Agni AI brand. MVP feature set:

Gateway

• POST /v1/chat/completions — OpenAI-compatible non-streaming and SSE-streaming. Body passes through; response carries an extra agni block with provider, model, mode, cache_hit, latency, cost, residency_actual, and request_id. (Renamed to inferbridge in v0.2.0.)
• Tier-based routing (agni/cheap, agni/balanced, agni/premium) across the candidate providers the caller has registered keys for. (Renamed to ib/* in v0.2.0.)
• Per-request X-Agni-Override-Model: provider:model to pin a specific upstream — including self_hosted endpoints, which are override-only.
• Residency filtering via X-Agni-Residency: india. Returns 422 residency_error if no registered key matches.
• Fallback across candidates on transient errors (5xx, timeouts, 429). Budget: max 2 provider attempts per request. Streaming fallback stops after the first token.
• Opt-in exact-match cache via X-Agni-Cache: true. TTL default 3600 s, configurable via X-Agni-Cache-TTL (clamped [60, 86400]). Redis outages never block requests.
• Per-request upstream timeout via X-Agni-Timeout: <seconds> (clamped [1, 300]). Belt-and-suspenders: asyncio.wait_for backs up the httpx client timeout.

Providers

• OpenAI (gpt-4o, gpt-4o-mini, …)
• Anthropic (claude-opus-4-7, claude-haiku-4-5) — OpenAI-shaped I/O with system-prompt extraction, max_tokens default, and usage.input_tokens / usage.output_tokens normalization.
• Together (meta-llama/Llama-3.3-70B-Instruct-Turbo, …)
• Sarvam (sarvam-m) — India-residency, free-per-token.
• Self-hosted — any OpenAI-compatible endpoint (vLLM, TGI, Ollama, llama.cpp, LM Studio, …). base_url required; HTTPS enforced unless the host is loopback or a private IP.

User + key management

• POST /v1/users — register; returns agni_<32 chars> API key exactly once. Stored as SHA-256 hash. (Prefix changed to ib_ in v0.2.0; keys issued under v0.1 remain valid.)
• POST /v1/keys, GET /v1/keys, DELETE /v1/keys/{id} — BYOK provider-key CRUD. Fernet-encrypted at rest. Partial unique index prevents duplicate commercial-provider keys per user while allowing multiple self_hosted entries.

Observability

• GET /v1/stats?since=&until= — totals, cache_hit_rate, avg_latency_ms (excl. cache hits), and group-by counts (by_mode, by_provider, by_residency, by_status). Max 90-day window.
• GET /v1/logs?since=&limit=&cursor= — reverse-chronological row stream with opaque cursor pagination; limit default 100, max 1000.
• X-Request-ID threads through the response header, structured logs (JSON in production via structlog), and the request_logs row.
• Prompts and completions are never logged. Pinning test (test_prompts_and_completions_never_reach_log_output) prevents regressions.

Ops

• Docker image (multi-stage, uv-built, ~250 MB, non-root).
• scripts/start.sh runs alembic upgrade head before uvicorn.
• Railway deploy: railway.toml + env-var reference at docs/deploy/railway-env.md.
• Rollback runbook: docs/deploy/rollback.md.

Known limitations

• No streaming tool use (function calling) yet.
• No vision inputs.
• No /v1/embeddings endpoint — call the provider directly.
• No semantic cache (exact-match only).
• No gateway-level rate limits in MVP — upstream 429s still surface via rate_limit_error.

GitHub release — placeholder