Siftpath 2.2.0

Siftpath 2.2.0 release notes

Backward-compatible feature release from 2.1.1. Publication is confirmed by the npm registry and Git release, not this source version. Releases require final candidate CI, package artifact/hash checks, npm authentication and explicit maintainer confirmation.

New providers

• Add an opt-in direct DeepSeek Search API backend (connection.transport: 'deepseek'). It stays disabled by default and does not change saved routes, enable services or switch any intent. No anonymous mode; a query outside the documented limits is rejected rather than rewritten or forwarded to a fallback provider.
• Ship a separately packaged native DeepSeek Harness bundle that registers one search provider and siftpath_read({ url, reader?, focus? }) without replacing DSH source files, hijacking a tool name or selecting Siftpath on installation. The bundle is not published: its pinned host dependency audit fails the High/Critical gate. See DSH.
• Keep the two model-facing MCP tools, the Pi extension, the default routes and the 16,000-character default output budget unchanged.

Output and diagnostics

• Report omittedChars whenever the budget clips a body. An agent can tell it holds a prefix and how much of the end was cut instead of guessing; the count is settled together with the prefix so it can never overflow maxOutputChars.
• Classify an observable upstream 402 as BALANCE_EXHAUSTED and an observable 429 as RATE_LIMITED rather than folding both into UPSTREAM_ERROR. Only a status code is read — structuredContent.error.status, or an HTTP <code> prefix in a text block. Provider wording is never scanned, so no ban, plan or entitlement is inferred, consistent with the existing 403 rule.
• Stop status from contradicting itself. A recorded provider failure now appears in Next: as Last test failed: <provider>(<code>), and an untested or failing service is no longer summarised as healthy. A failure recorded before a package upgrade or configuration change is still listed but labelled with why it may no longer describe the current setup.
• Name every attempt when no provider succeeds: Each attempt failed: jina(BALANCE_EXHAUSTED), jina-backup(AUTH_FAILED). A single generic sentence made users re-check everything when the actionable cause was already known.
• Document multi-account redundancy as configuration rather than a feature: a second account is another provider plus a fallback entry, keeping the failing account visible in warnings. No automatic account rotation and no balance polling are added — whether a 402 means transient throttling, an exhausted plan or genuine non-payment is decided by the operator, not guessed locally.
• Document what Siftpath cannot see, because it forwards reader responses instead of fetching pages itself: a 404 body, an anti-bot block or empty boilerplate still arrives as isError: false; raw HTML is passed through uncleaned; a focus miss is not a NOT_FOUND.

Compatibility and upgrade

The web_search / web_read inputs, default routes, explicit Reader selection and fallback rules are unchanged. Saved Keys, custom bindings, routes and legacy configuration compatibility remain; no configuration rewrite or setup rerun is required. Upgrade the package, then restart Pi or /reload.

Consumers matching output fields or error codes should handle the additive omittedChars field and the new BALANCE_EXHAUSTED / RATE_LIMITED codes where a generic UPSTREAM_ERROR could previously appear. omittedChars is present only when the body was actually clipped. Do not parse human-readable error messages as an API. See diagnostics and migration.

Known limits and acceptance

• The default serialized JSON budget remains 16,000 characters, not tokens. Long bodies can lose decisive evidence; truncated describes local clipping only, not upstream completeness, and omittedChars quantifies the cut without recovering it.
• A supported focus binding, an untruncated response or a successful call does not guarantee complete evidence or an accurate model answer. Source indexes can contain unfetched outgoing links.
• Reader-side failures stay invisible by design. Because no local fetch is performed, Siftpath cannot report an HTTP status the provider never sent.
• The shipped Jina focus window stays at topk: 3, chunk_size: 250. Narrower windows return fewer characters but were observed to drop the sentence answering a single-fact focus while still reporting no clipping; see selective reading.
• Current candidate checks and historical evidence are recorded separately in testing.
• Live acceptance for this candidate covers provider connectivity, a successful read from each configured Reader, omittedChars accuracy on a real clipped page, and the Jina focus-window comparison. It does not include a model-driven agent session: whether a model uses omittedChars or focus correctly is still unverified.