Provider Hub v0.6.16

Simple LLM-driven specification filling

• Simplifies one-click model specification research: search the web, ask the selected LLM, and persist every usable field the LLM returns.
• Removes strict same-sentence field proof and two-community-domain consensus requirements from the active one-click flow.
• Results from non-official/community evidence are labeled llm-researched and shown as editable in the UI.
• Keeps baseline validation: normalized model identity, positive numeric fields, output not exceeding maximum context, supported reasoning-level keys, and normalized pricing.
• Model identity ignores case and common Unicode hyphen variants (gpt-5.5, GPT-5.5, GPT‑5.5) while still rejecting suffix changes such as gpt-5.5-pro.
• Fixes the research JSON schema wording and merges exact Azure structured fields when the LLM omits them.
• Only explicit maximum context is quartered; legacy contextWindow is treated as the recommended runtime value.
• SSRF/public-DNS safeguards, credential isolation, prompt/response redaction, and research-call auditing remain unchanged.

Verification

• 110/110 tests passed.
• Package dry-run, syntax and diff checks passed.

SHA-256: A1CB2F5E31F816E1003DBD059D3CF105218A630E98892F94954E18BD31776A63

Provider Hub v0.6.15

Research observability and pricing enrichment

• Fixes Azure structured evidence contamination by reducing embedded catalog data to exact-model, exact-field evidence sentences before deterministic validation.
• Adds safe per-model research diagnostics: query count, successful/failed searches, selected/evidence source counts, fetch statuses, evidence size, model-call start/completion and response character count. No prompt or response text is stored.
• Audits Provider Hub route research LLM calls as purpose: model-spec-research, including Key name, route, model, status, latency, usage and cost without persisting prompt/response text, Key values or credential refs.
• Adds a fourth specification search for input/cached-input/output/reasoning per-million Token pricing.
• Stores only independently verified pricing fields in model specifications.
• Cost precedence: provider-reported, route-pricing, then model-spec-pricing; incomplete used categories remain unavailable.
• Request logs now record selected reasoning effort/enabled status and reasoning Token usage.
• UI shows search diagnostics, research-call purpose, reasoning effort, model specification reference pricing, and cost-source semantics.

Verification

• 108/108 tests passed.
• Package dry-run, syntax and diff checks passed.

SHA-256: 16C4779A02097F9354D447DE926CC5231C2C96C4CF58DB0231F40B9A6B59C303

Provider Hub v0.6.14

Research compatibility fixes

• Corrects the AWS Bedrock GPT-5.5 model-card slug (gpt-55) so the official platform source is actually fetched.
• Accepts legacy research-model JSON that still returns contextWindow, while new explicit maximumContextWindow results continue to use the conservative quarter-maximum recommendation policy.
• Preserves strict field-level evidence validation and does not loosen URL/title-only trust.

Verification

• 104/104 tests passed.
• Package dry-run, syntax and diff checks passed.

SHA-256: 8F3C19DE1505A2DE3BCE547F12F47293029CD10CBBE16705D9A316C1986447D4

Provider Hub v0.6.13

Conservative context recommendations

• Separates maximumContextWindow (hard supported limit) from contextWindow (runtime/compaction recommendation).
• If a source explicitly provides a recommended runtime window, Provider Hub uses it.
• If only the maximum is proved, Provider Hub derives loor(maximum / 4) and records contextWindowPolicy: quarter-maximum plus derivation metadata.
• Example: Azure's 1,050,000-token maximum becomes a 262,500-token runtime recommendation; the 1,050,000 maximum remains visible separately.
• Existing legacy specifications with only contextWindow retain their prior value and are not silently changed.
• Adds safe platform-official evidence support for exact OpenAI model records from Azure AI Catalog and AWS Bedrock pages, while URL/title alone remains insufficient.

Verification

• 104/104 tests passed.
• Added platform-official structured-field, failed-fetch, and quarter-maximum recommendation tests.
• Package dry-run, syntax and diff checks passed.

SHA-256: BB76112FDA0EB25DC904C2EB67916E88C04628D6B000CE42E0A62363EDCAACAB

Provider Hub v0.6.12

Managed provider ownership migration

• Fixes a false conflict when an old Provider Hub ownership snapshot predates the managed Relay apiKeyEnv field.
• Migration is deliberately narrow: owned must already be true, every field except apiKeyEnv must match exactly, the credential must be configured, and the current provider apiKeyEnv must exactly equal the active Relay credential reference.
• Any different credential reference or other user edit remains a conflict and is never overwritten.

Verification

• 101/101 tests passed.
• Added successful legacy snapshot migration and different-apiKeyEnv conflict protection tests.
• Package dry-run, syntax and diff checks passed.

SHA-256: 9A69DFE5981ACF42B05C589C13E5683442B077714B9D2C87B76494A28BCCFD90

Provider Hub v0.6.11

SSRF hardening

• Parses IPv6 addresses into normalized numeric segments before classification.
• Detects compressed, expanded, and dotted IPv4-mapped IPv6 forms.
• Converts mapped final 32 bits back to IPv4 and applies the complete non-public IPv4 policy.
• Blocks loopback examples including ::ffff:7f00:1, 0:0:0:0:0:ffff:7f00:1, and ::ffff:127.0.0.1 before any HTTPS source request.
• Retains the existing DNS+HTTPS end-to-end hard deadline and pinned-public-address request design.

Verification

• 99/99 tests passed.
• Added mapped and expanded IPv6 loopback source rejection regression coverage.
• Package dry-run, syntax and diff checks passed.

SHA-256: 67578517FF813BC754EFA1A23070CAD48B918F8F0E3498848FD9BF340CD17D4E

Provider Hub v0.6.10

Connection and pricing correctness

• Cancels non-2xx response bodies before failover so upstream connections are released.
• Cost estimation now requires an explicit rate for every token category actually used; unknown categories are never treated as free.
• Anthropic cache_read_input_tokens is handled as additive input and total tokens are kept internally consistent.
• Startup/config-file modelPricing now uses the same normalization as management saves: numeric strings normalize, negative rates are rejected, and currencies uppercase/fallback consistently.

Verification

• 98/98 tests passed.
• Added non-2xx body cancellation, partial-pricing unknown-cost, Anthropic additive cache, and startup-pricing normalization tests.
• Package dry-run, syntax and diff checks passed.

SHA-256: 4AA94AD30795CEC696F37BC8E3F1F5BBBA3FB867E0D03F551F988394C158F059

Provider Hub v0.6.9

Accounting correctness fixes

• Partial usage frames now only update fields they actually contain, preserving token fields captured from earlier frames.
• Null or absent provider cost is not treated as reported zero; route pricing estimation remains available.
• Summary metrics now distinguish client requests from route attempts and failovers.
• One 503 then successful failover counts as one successful request, two attempts, one failed attempt, and one failover.
• Request success/failure and average latency are grouped by Provider Hub request UUID and use end-to-end client request timing.
• Token and cost totals still include every actual route attempt, because failed attempts can consume billable usage.

Verification

• 94/94 tests passed.
• Added partial usage/null-cost regression and request-level failover summary timing tests.
• Package dry-run, syntax and diff checks passed.

SHA-256: C9177B62802590E5EB8E702982EDE6FE3F3EABA620F439CD77F43AC17B8910C6

Provider Hub v0.6.8

Security and correctness fixes

• Search evidence now has explicit channels: real search citation text and successfully fetched page content. URL/title/fetch-failure status alone can never validate a field.
• Official search citations remain usable when the official page returns 403, while blocked/non-public sources are discarded.
• Added end-to-end hard deadline across DNS resolution and HTTPS source retrieval.
• Client disconnect now aborts upstream generation; relay observes writable backpressure and marks cancelled requests failed.
• Provider Hub generates and stores its own request UUID; caller-controlled X-Request-ID is never persisted.
• SSE parsing joins legal multiline data events before JSON parsing.

Existing v0.6.7 capabilities

• One-click model specification completion with field-level evidence.
• Token, cache, reasoning, latency, rate-limit and cost logs.
• Provider-reported usage/cost preferred; local estimation is explicitly labeled.

Verification

• 92/92 automated tests passed.
• Includes blocked-fetch, 403 citation, private DNS, stalled DNS, malformed search, caller-ID redaction, multiline SSE, downstream cancellation, Token/fee and one-click completion tests.
• Package dry-run, syntax, diff and secret scans passed.

SHA-256: F795049940681C21DD737AD3A0AA2A11A198CA56DE8EC865345059735EC23B26

Provider Hub v0.6.7

Highlights

• Adds a permanent 一键填写规格 action in the Model Specifications page.
• Runs separate field-focused searches for context window, maximum output, and reasoning effort.
• Safely reads accessible HTTPS source pages with public-address validation, DNS pinning, redirect blocking, timeout, content-type, and size limits.
• Preserves official-first and two-independent-community-domain evidence rules with model-local field attribution.
• Adds request-level usage and billing logs: request/provider IDs, key route, model, retries, status, finish reason, input/cache/output/reasoning/total tokens, time to first token, total latency, rate-limit remainder, and cost source.
• Uses provider-reported usage/cost when available; otherwise marks token estimates and only estimates cost when per-model route pricing is configured.
• Adds per-model pricing JSON configuration in currency units per one million tokens.
• Never logs prompts, full responses, API keys, credential references, or full upstream URLs.

Verification

• 85/85 automated tests passed.
• One-click management API acceptance test fills and hot-syncs model specifications from fetched page evidence.
• Production fetch-path tests cover pinned public DNS, HTML extraction, non-public address blocking, and malformed search tolerance.
• Non-stream and SSE pass-through tests cover usage, cost, first-token latency, and redaction.
• Detailed log UI passed strict visual review.

SHA-256: 9BE72110C535362045822615CA17488A00FC6B933EC6C8D0400C2ED5288AB885