lex-llm-ollama 0.2.25 → 0.3.4

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: c9bae0bff004578c1bfcaddc4a504aaed28b1091cb130c73215898bd7548cb62
-  data.tar.gz: 86faf3f3a7aba8f68f3204f2af601fc7fbb652937bd87550949d7fa601c55969
+  metadata.gz: a7cb2a1511c933e0174c414ba400ef0d7fcc700feeb2ac866be66004447f9df9
+  data.tar.gz: 90756cc1de56697e25880b801a07016a6e0d6b5edda22025b053850716e267cf
 SHA512:
-  metadata.gz: 8f394effa9d62cb81de6404783baed233f4a9fb28e039d36b3a479b85854fd7fa719a7786ebdd7dbc350474d755e951fccad1a801130a6eebf3a3a4dc796fbc0
-  data.tar.gz: 478fdb238735152ea67597c88da3b8a251075d57a368604110a76bb6dd87ec2ba3784dfaae8952a11fcfc0babfce26118acd01c3c46ed5481f26a1cbe35260b9
+  metadata.gz: 603aee402d7b5f0733369fb78c45430ac452ae33d2f8cf05adc7871552ba19f556cd43ab9cf3986ad8a9644798fd0924b394c5d39aff4150ffcc5714003ec299
+  data.tar.gz: 27a834f6be3252033fcc7f8028d8fd44f2056202bb7adf441f01b34f35832791aa6ac69c55db782398a509255af45f6c0ede82733d21dc7b19aa166683590b53

data/CHANGELOG.md

@@ -1,5 +1,70 @@
 # Changelog
 
+## [Unreleased]
+
+### Fixed
+- Remove synthetic-default discovery filtering and its once-per-boot warning; configured discovery now passes every instance entry through normal enabled and endpoint validation.
+- **D1 Callable dispatch** — `OllamaCallable` now implements the fleet dispatch ops (`chat`, `stream_chat`, `embed`, `count_tokens`) by delegating to a per-instance `Ollama::Provider` (previously hardcoded fake payloads that swallowed `messages:` into `**`); errors propagate unrescued for `normalize_dispatch_error`; `disconnect` closes the Provider. Optional `provider:` injection seam for specs (production builds the real Provider lazily).
+- **D15 Raw-string model at the dispatch boundary** — the fleet passes `model:` as the offering's raw id (String). Ollama's render path is string-tolerant (`model.respond_to?(:id) ? model.id : model`) for chat and embed, embed places the model verbatim in the `Embedding` response object, and count_tokens ignores it — so the raw string passes through UNWRAPPED on every op (wrapping would serialize a `Data` object into the wire payload or the response object); `Model::Info` instances pass through unchanged.
+- **D4 Initial-failure recovery** — an instance whose initial readiness failed stays claimable: each tick probes while `:initializing` and re-activates via `activate_instance_snapshot` (fresh probe token, current offerings, sequence 0) on the first passing probe. Previously the instance stayed `:initializing` for the process lifetime (`replace_instance_snapshot`/`readiness_succeeded` refuse to operate on an `:initializing` scope).
+- **D4 Tick reconcile** — configured instances are re-scanned every tick: instances configured after boot are claimed without a restart; removed instances are retired from the Registry and their settings health cleared.
+- **D3 Unconfigured phantom** — only operator-configured instances are claimed. The synthetic `instances.default` section (provider_settings nests the extension's own instance defaults there) is skipped with a warn while it is still the unmodified extension default; named instances are always operator-authored; `enabled: false` and endpoint-less entries are skipped with a warn. No host/port fallback in `derive_instance_id` (raises on a missing endpoint instead).
+- **D3 Legacy port-scan phantom** — the entry module's `discover_local_instance` socket probe (fabricated `:local` when 127.0.0.1:11434 answers) is deleted; `Ollama.discover_instances` and the SSOT actor share one source (`Ollama.configured_instances`).
+- **D2 Bridge** — `Publisher` constructed with the `ScopedRefresher::LegacyCoordinatorAdapter` so SSOT commits project into the old `Legion::LLM::Inventory` coordinator during the mixed-version window.
+- **D8/D17 Error shapes** — `normalize_dispatch_error` delegates to the base `Provider#normalize_dispatch_error` classification (every `Llm::*Error` the ErrorMiddleware raises, plus the raw Faraday transport errors it does not wrap: connection failure — the Ollama down-signal — and timeout) and layers provider-specific detection on top: a 5xx body reporting the model not loaded/loading is `:model_not_ready`, and raw Faraday HTTP status errors (which the base leaves as `:provider_error`) are refined by status. Body/status detection reads every real error shape (`Faraday::Response` from Llm errors, `Faraday::Env` from real Faraday 2.x errors, plain Hash) — no `is_a?(Hash)`-only gate. Readiness status detection handles Response/Env/Hash shapes.
+- **D9 Cadence interval** — actor `time` reads the registered `discovery.interval_seconds` (never nil; falls back to the registered default); dead `self.every_seconds` removed.
+- **D12 State store** — `@instance_states` moved from a plain `Hash` to a `Concurrent::Map` (timer/dispatch/shutdown threads).
+- **D13 Fleet dispatch** — fleet `Subscription` actor's `runner_class` returns the runner CONSTANT (a String cannot be `send`-ed by the `runner_class.send(fn, **message)` dispatch path) and the runner accepts the envelope as kwargs (`handle_fleet_request(**opts)`), matching the dispatch shape.
+- **D14 Health display** — after each registry commit the actor writes `settings[:instances][<config_name>][:health]` (legacy 4-key shape: `circuit_state`/`denied`/`available`/`adjustment` + display `reason`/`observed_at`(ISO8601 UTC)/`last_probe_outcome`/`source`) and `[:capabilities]`; cleared on removal/shutdown. Display-only — routing authority stays the in-memory `AvailabilityFact`.
+- **D16 Loud discovery** — `discover_offerings_for_instance` rescues only `Faraday::Error`/`Legion::JSON::ParseError` (transport/parse); programming errors (`NameError`/`NoMethodError`/`ArgumentError`) propagate instead of publishing an activated instance with an empty offering set. The actor-runtime soft guard (`rescue LoadError` + `return unless defined?`) now raises `LoadError` (fail loud).
+- **D3-churn** — offerings are compared on identity/status fields (not `Data#==`, which fresh `Time.now` `observed_at` stamps poison) so an unchanged catalog no longer triggers `replace_instance_snapshot` every tick.
+- **Standard sweep** — `require_relative` → `require`; in-method `require 'faraday'` hoisted to file top; `extend Core if const_defined?` guard removed (hard dep); logic extracted from the actor class into `EvidenceBuilder`, `ValueEvidenceBuilder`, `ModelDiscovery`, `ConfigResolver`, `HttpClient`, `ProbeRunner`, `HealthChecker`, `OfferingComparison`, `DisplayHealth`, `InstanceComponents`, `InstanceLifecycle` modules so all files sit under Metrics limits.
+- **D6 Specs** — conformance harness now returns the PRODUCTION `OllamaCallable` (the `TrackingOllamaCallable` re-implementation is deleted) and delegates identity/draft building to the actor's real helpers; error fixtures use real Faraday 2.x `Faraday::Env` shapes; new D17 Llm-error-shape classification tests; new D15 per-op raw-string render-path tests; kit glob narrowed to the consumed example-group files (lex-llm self-test specs no longer run in this suite); NEW `actor/discovery_refresh_spec.rb` lifecycle coverage (multi-instance claim/activate, D3 phantom handling, D4 recovery, tick reconcile, D9 interval, D14 health, churn, D16 loud errors, shutdown).
+- **D10 Lock** — `Gemfile.lock` regenerated locally (gitignored): lex-llm now resolves 0.7.0 (gemspec requires `>= 0.7.0`; the stale lock pinned 0.6.17, where the SSOT inventory layer does not exist).
+- **Instance identity is now the operator's config NAME** — the discovery actor previously keyed instances by the derived `host:port` string (`derive_instance_id`, now `derive_physical_id`). The derived id silently inerted the router's `instances.<name>` settings lookups (per-instance tuning, weight, preferred context windows) and collapsed distinct config names that share an endpoint. Discovery now publishes `InstanceKey.instance_id` = the config name and carries the derived `host:port` in the secondary `physical_id` field (dedup/diagnostics only — it never participates in equality, hashing, or registry-scope identity). Two config names pointing at the same endpoint stay distinct instances. A modified `instances.default` is skipped in `claimable_instance_config` (the single source shared by the discovery actor and the fleet responder) because the lex-llm foundation reserves `default` as an instance identity; no host/port fallback in `derive_physical_id` (raises on a missing endpoint instead).
+- **Embedding models now authoritatively exclude chat** — an embedding model published `chat: :supported`/`stream_chat: :supported`, so a plain chat request could be misrouted to an embedding-only model. The evidence builder now branches on `embed_supported` (matching bedrock): embedding models publish `chat`/`stream_chat` as `:unsupported` and `embed` as `:supported`; chat models publish `embed` as `:unsupported`.
+- **lex-llm 0.7.1 floor** — the gemspec now requires `lex-llm >= 0.7.1`: the name-identity `InstanceKey` with the secondary `physical_id` field and the reserved-`default` rejection (0.7.0's `InstanceKey` has no `physical_id` and accepts any instance_id).
+- **Single actor registration** — the provider module no longer extends Core at file level, so the boot-time submodule walk skips it and the gem's own top-level extension load is the sole actor registration (eliminates the double-claim / FencedPublisherError).
+- **Synthetic-default skip warn** — the `reason=synthetic_default` skip warn now fires once per boot instead of every discovery tick (was permanent WARN noise).
+
+## [0.3.2] - 2026-08-13
+
+### Fixed
+- **§1 settings access via .dig:** `api_base` now reads `settings[:instances][:default][:endpoint]` using bracket notation instead of the prohibited `.dig` form.
+- **§9 default model substitution removed:** `translator.rb` `render_request` no longer falls back to `'default'` when model is absent from request metadata; the field is omitted (compacted) so the exact selected model must be provided by the executor.
+- **§1 swallowed rescue fixed:** `extract_host_port` in the discovery actor now calls `handle_exception` and re-raises on `URI::InvalidURIError` instead of silently returning `'unknown:0'`.
+- **§1 swallowed rescue fixed:** `check_readiness` rescues now call `handle_exception` (with `handled: true`) before constructing the `ReadinessResult`, preserving the log audit trail.
+- **§1 cop strictness reverted:** `RSpec/MultipleMemoizedHelpers: Max: 7` removed from `.rubocop.yml`; conformance spec refactored to comply with the default maximum of 5 memoized helpers per context.
+
+## [0.3.1] - 2026-08-13
+
+### Fixed
+- **§8 health firewall:** Connection failures, timeouts, and 5xx errors stay request-local and never map to `:instance_unavailable`. Only an explicit flat service-unavailable wire signal maps to that outcome.
+- **§9 no `:default` substitution:** `offering_from_model` derives a stable `instance_id` from the configured endpoint URL (`host:port`) instead of substituting the `:default` symbol.
+- **§1 rubocop:disable removed:** All `rubocop:disable` comments removed from source and spec files; underlying causes fixed.
+- **§1 swallowed rescue fixed:** `coordinator.finish_probe` failure now logged with `handle_exception` rather than silently swallowed.
+- **§1 settings path corrected:** `api_base` reads `settings.dig(:instances, :default, :endpoint)` instead of the non-existent top-level `settings[:endpoint]`, preventing invalid base URL construction.
+- **§2/§5 second publication engine removed:** `readiness` and `discover_live_offerings` no longer call `registry_publisher.publish_readiness_async` or `publish_models_async`; the SSOT v3 actor is the sole publication path.
+- **Provider capabilities corrected:** `vision?`, `functions?`, `embedding?` return `false` (not `true`) as the static predicate; evidence-based values are derived per-model by the discovery actor.
+- **Settings `embedding: true` removed:** Blanket embedding capability removed from usage defaults; embedding is declared per-model by the discovery actor based on `/api/tags` family evidence.
+
+## [0.3.0] - 2026-08-13
+
+### Changed
+- **SSOT v3 provider migration.** Replace legacy `ScopedRefresher`/`Legion::LLM::Call::Registry` discovery with direct `Inventory::Publisher` claim/activate/replace lifecycle using the lex-llm 0.7.0 runtime contract.
+- Discovery actor now claims exact Ollama instances, builds complete `OfferingDraft` snapshots from `/api/tags` and `/api/show`, and probes readiness via `/api/tags` (non-inference, no model load).
+- Per-model operation evidence: chat/stream_chat always supported, embed supported only for embedding models, image/transcribe/translate/speak/moderate unsupported, count_tokens unknown.
+- Per-model capability evidence from `/api/show` response (tools, thinking, vision) with honest unknown when unavailable.
+- Remove `default_model: 'qwen3.5:latest'` from settings. Omitted model reaches the router unconstrained; provider invocation receives the exact selected model.
+- Stable InstanceKey derivation: `host:port` from the configured endpoint URL.
+- Normalized dispatch errors: connection failure stays `connection_failure`, 503 stays `overloaded`, timeouts stay `timeout`. No raw status code produces `instance_unavailable`.
+- Graceful shutdown removes all claimed instances.
+- Raise gemspec floor to `lex-llm >= 0.7.0`.
+
+### Added
+- `OllamaCallable` wrapper implementing `disconnect` and `normalize_dispatch_error(error:)` for the SSOT callable contract.
+- SSOT v3 conformance spec exercising shared provider examples.
+
 ## [0.2.25] - 2026-08-04
 
 ### Changed

data/Gemfile

@@ -2,6 +2,8 @@
 
 source 'https://rubygems.org'
 
+gem 'lex-llm'
+
 gemspec
 
 group :development do

data/lex-llm-ollama.gemspec

@@ -25,7 +25,9 @@ Gem::Specification.new do |spec|
 
   spec.add_dependency 'legion-json', '>= 1.2.1'
   spec.add_dependency 'legion-logging', '>= 1.3.2'
-  spec.add_dependency 'legion-settings', '>= 1.3.14'
+  spec.add_dependency 'legion-settings', '>= 1.4.2'
   spec.add_dependency 'legion-transport', '>= 1.4.14'
-  spec.add_dependency 'lex-llm', '>= 0.6.9'
+  # 0.7.1 carries the name-identity InstanceKey (instance_id = config name +
+  # secondary physical_id field) that the SSOT v3 discovery actor publishes.
+  spec.add_dependency 'lex-llm', '>= 0.7.1'
 end