What each delivery phase added, and why. For current behaviour, follow the documentation index instead — this page is history.

The project's delivery is organised in phases. Each one is independently deployable and additive; nothing in a later phase is required by an earlier one.

Phase 1 — the gateway#

The synchronous inference path.

Phase 2 — the policy-driven control plane#

Turns the router into something that explains itself. Every request now flows through ten visible stages: authenticate → policy → classify → shape → cache → health and scores → route → execute → fall back → persist.

Phase 3 — management and provisioning#

Turns a view-only gateway into a manageable control plane. Providers, models, tenants, keys, policies, endpoints and overrides are created, edited, disabled and deleted through the API and the dashboard: no config-file edits, no restarts.

Phase 4 — tool calling and agent runs#

Turns the gateway from a request router into an agent runtime. Models can call tools, answers can be contract-checked against a JSON Schema, and every gateway-side run leaves a durable step trace.

Phase 5 — the local response cache#

Makes Synapass faster and cheaper by reusing prior responses wherever it is safe, with safety as the first consideration rather than an afterthought.

Public landing page#

A browser reaching the gateway root now gets a small HTML page naming the real endpoints, and JSON for everything else. A gateway reached from a tunnel URL or typed into a browser should not answer with a 404 that reads as a broken deployment.

Cloudflare tunnels#

Temporary public exposure of one local service through a supervised cloudflared child process. Opt-in twice — feature flag and explicit create — and loopback targets only, so a tunnel can never become a proxy for someone else's infrastructure. See installation/cloudflare-tunnel.md.

Truncation fix#

Long answers were stopping mid-sentence for three independent reasons, all fixed:

1. The soft latency target was reused as a hard per-attempt deadline. The default 10s target cancelled every generation longer than ten seconds. It is now a ranking signal only, and generation budgets come from the timeout policy. An explicitly configured budget is never raised behind the operator's back.
2. The shared transport set ResponseHeaderTimeout to the first-token budget. For a buffered request a provider only sends headers once the whole answer exists, so long completions were aborted before a byte of body arrived. Removed; per-attempt deadlines come from the request context.
3. The policy output ceiling was sent as the client's max_tokens. A client that asked for no limit was given one and nothing reported the cut. A ceiling is now only sent when the client requested one — and Anthropic's mandatory field gets a generous default rather than a small constant.

Truncation is now always reported in synapass.completion (truncated, reason, requested_tokens, applied_tokens, budget_ms), counted in synapass_provider_truncations_total, and never conflated with a completed answer. A stream that has begun is committed: the executor reports a stream-started error rather than appending a second attempt's tokens to the first.

Documentation#

This documentation set. The previous phase-by-phase notes were folded into topic-based pages so each fact is documented once and cross-linked, rather than restated per phase.

---

Related: Documentation index · Architecture · Back to README