unhardcoded-use

name

unhardcoded-use

description

Use a running unhardcoded router — as a CALLER (send an OpenAI-compatible request carrying a Σ_pol policy_ir / Σ_flow flow_ir, or a profile/family/pin model string; session pinning; streaming; dry-run before spending) and as an OPERATOR (dashboard, /x/* endpoints, mint keys, hot-add providers, AntSeed wallet). Load this when integrating an app or agent against the router, wiring a client, minting keys, debugging a route, or operating the host. To AUTHOR a policy term against this host's live catalog use the root SKILL.md (sigma-policy-author); to develop the host itself use unhardcoded-contribute.

Using a running unhardcoded router

The router is an OpenAI-compatible endpoint. You do not pick a model name — you send a policy (data) and the host filters/ranks/picks a (provider, model) over the operator's keys, falling back on provider errors. When a policy is present the request's model field is ignored for selection.

There are two roles: caller (hits /v1/* with a bearer key) and operator (the dashboard + the internal /x/* endpoints). This skill orients both; the detail lives in docs/ (linked inline).

As a caller

Auth. Bearer key, prefix llmr_…. Mint one from the dashboard (Consumers) or POST /dashboard/api/keys. Send it as Authorization: Bearer llmr_….

Send a call. POST /v1/chat/completions (also POST /v1/responses), body = standard OpenAI + one of:

• policy_ir: a Σ_pol term — decides which model serves this one call.
• flow_ir: a Σ_flow term (["flow", nodes]) — a DAG of calls, each node carrying its own policy.
• no policy → the host's default policy runs. Or steer via the model string: profile:<name>, family:<f>, pin:<provider>/<family>. You can also path-address a profile: POST /{profile}/v1/chat/completions.

Session pinning. Set X-Unhardcoded-Session: <conversation-id> (or a session body field — body wins). The router keeps the conversation on its cache-hot peer (warm prompt-cache across turns) and meters spend per session.

Streaming. stream: true. Errors before the first token return a normal JSON error (no partial SSE commit), so client-side fallback still works.

Other caller endpoints. GET /v1/models (families served), POST /v1/compact (stateless context sealing — hand it your turns, get a sealed summary back).

Dry-run before you spend (these cost nothing):

• POST /x/policy/normalize — admit a policy_ir + get its fingerprint/version.
• POST /x/rank — the ordered candidates this host would try for a term.
• POST /x/policy/build — elaborate a declarative spec → a policy_ir.
• POST /x/flow/normalize — admit + identify a flow_ir.
• GET /x/fields — the live field schema a policy may reference.

Authoring the policy itself is a separate skill. The root SKILL.md (sigma-policy-author) is the canonical authoring guide: it embeds this host's live model/provider catalog and field vocabulary and is test-gated so it can only teach ops the core actually admits. Load it to write terms; the normative grammar is core/docs/SIGMA-POL.md.

As an operator

The /x/* endpoints are operator-only — the ingress proxy hides them from consumers. The dashboard (Analytics · Debugger · Activity · Catalog · SKILL.md · Config · Settings) wraps most of them:

• Keys: mint/list consumer keys.
• Providers: hot-add or re-key a provider at runtime — POST /x/providers, POST /x/provider-key (persisted as overlays; no redeploy).
• Observe: GET /x/runtime (breakers, EMA metrics, source freshness), GET /x/market (full price book per family), GET /x/calls (recent ledger), GET /x/session/{sid} and GET /x/sessions (per-session meters).
• Reload: POST /x/config/reload, POST /x/codex/reload.
• AntSeed wallet: POST /x/wallet/deposit|withdraw|refresh (funds the marketplace buyer escrow — see the antseed-prod-deposit runbook for prod).

Gotchas

• model is ignored when a policy is present — the policy drives selection.
• The host envelope only narrows. config.policy_envelope is ∧-composed onto every caller policy: a caller can tighten the host's invariants, never widen them. A term that passes your local check can still be narrowed by the host.
• Seed prices are a placeholder. metrics.live.lua is intentionally fake (see docs/METRICS.md); real EMA prices/latency accrue from live calls. Don't trust the seed numbers for cost reasoning on a fresh host.
• Codex ranks near-free by design (subscription = sunk cost, seeded ≈ $0), so a cheapest-first policy prefers it when available — deliberate, not a bug.

Where to read more

• Root SKILL.md — author Σ_pol/Σ_flow against this host's live catalog (the authoring canon).
• docs/DEPLOY.md — the client contract (OpenAI SDK base_url), deploy + provider modes.
• docs/USAGE_ENDPOINTS.md — per-key usage stats API + the consumer self-service model.
• docs/PROVIDERS.md — per-provider auth (bearer/oauth/none), the AntSeed node + wallet.
• docs/OPENAI-CODEX.md — the ChatGPT-subscription (Codex) provider, its wire shape and setup.
• docs/METRICS.md — the metrics seed format and why it's fake.