fix(memory): classify Ollama-unavailable embed failures actionably - #5398
Merged
M3gA-Mind merged 5 commits intoAug 5, 2026
Merged
Conversation
The memory embedder reported an unusable local Ollama runtime as a generic transient fault. Both failure shapes the embedder produces -- "is Ollama running at <base>?" (daemon not listening) and "Ollama embedding model `<id>` is not installed at <base>" (model never pulled) -- carry no `Embedding API error (<status>)` envelope, so `classify_embed_error_str` fell through to `Transient` and the memory status panel told the user "a temporary error interrupted memory processing, it will retry automatically". Retrying cannot start a daemon or pull a model, and the actual fix was never named. `FailureCode::LocalModelUnavailable` and its translated remediation already existed but had no producer anywhere in the tree. Match the two shapes explicitly so that code is emitted, and mark the semantic-recall surface degraded at classification time so the remediation appears on the first failed embed rather than after the retry budget drains. Keep the code in the transient retry class. Only transient rows are picked up by `requeue_transient_failed`, the automatic self-healing requeue; classifying it unrecoverable would park every affected job until someone pressed "Retry failed" by hand, so a user who simply restarted Ollama would never see ingestion resume. Also bridge the existing health-gate signal to the UI. The gate published `DomainEvent::EmbeddingModelUnhealthy`, but nothing carries the domain bus to the product UI -- `/events/domain` is read only by the developer Event Log panel -- so that event reached no user. Broadcast the condition over the metadata-only `user_error` web-channel path the cron scheduler already uses, which lands a durable UserErrorCenter entry with a deep link to provider settings. The broadcast deliberately sits above the once-per-process Sentry latch: `publish_web_channel_event` is an unbuffered broadcast send, memory is constructed before the renderer socket attaches, and a single dropped send under the latch would be the only attempt ever made. The panel store dedupes on the descriptor identity, so repeats collapse into one entry. The reported root cause -- a deprecated `/api/embeddings` call -- was not present: the embedder has used `POST /api/embed` with `input` since the provider port, and the deprecated path appears nowhere in the tree. Closes tinyhumansai#5354
…ed-error-classification
Contributor
|
No actionable comments were generated in the recent review. 🎉 ℹ️ Recent review info⚙️ Run configurationConfiguration used: Organization UI Review profile: CHILL Plan: Pro Run ID: 📒 Files selected for processing (15)
🚧 Files skipped from review as they are similar to previous changes (14)
📝 WalkthroughWalkthroughThe change detects unavailable local Ollama models, marks memory recall as degraded, broadcasts metadata-only errors, classifies them in the frontend, adds memory error scopes, and provides localized remediation in thirteen languages. ChangesLocal model unavailability handling
Estimated code review effort: 4 (Complex) | ~45 minutes Sequence Diagram(s)sequenceDiagram
participant EmbeddingPath
participant HealthModule
participant UserErrorPublisher
participant SocketService
participant FrontendClassifier
EmbeddingPath->>HealthModule: classify local model failure
HealthModule->>UserErrorPublisher: publish metadata-only user_error
UserErrorPublisher->>SocketService: broadcast memory event
SocketService->>FrontendClassifier: ingest with memory scope
FrontendClassifier->>FrontendClassifier: select provider-settings remediation
Possibly related PRs
Suggested reviewers: Poem
🚥 Pre-merge checks | ✅ 5✅ Passed checks (5 passed)
Comment |
YellowSnnowmann
marked this pull request as ready for review
August 4, 2026 14:11
coderabbitai
Bot
added
bug
memory
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 6fb354eed7
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
|
| Filename | Overview |
|---|---|
| src/openhuman/memory/tree/health/mod.rs | Core classification logic: adds LocalModelUnavailable prose matchers, mark_local_model_unavailable_if_applicable with edge detection, and upgrades degraded-flag stores from Relaxed to Release/Acquire. Logic and memory ordering are correct. |
| src/openhuman/memory/tree/health/user_error.rs | New module: owns the metadata-only user_error payload and publisher. Pinned by three tests (payload shape, wire token, source mapping). Clean. |
| src/openhuman/memory/store/factories.rs | Moves surface_local_model_unavailable_to_clients() above the Sentry latch so socket-attach races before renderer connection don't permanently silence the UserErrorCenter entry. Tested by a new Rust unit test. |
| app/src/lib/userErrors/classify.ts | Adds isLocalModelUnavailable rule last (after credits rules) with token + four anchored prose matchers. Both Rust-side error shapes are covered; negative tests for bare 'daemon unreachable' and credits-error-naming-Ollama guard the blast radius. |
| app/src/services/socketService.ts | Derives scope from error_source (memory → 'memory', else 'cron') instead of hard-coding 'cron'. The binary mapping is acknowledged as intentional; existing cron tests are unmodified and pass. |
| app/src/types/userError.ts | Additively widens UserErrorKind and UserErrorScope unions. No exhaustive branches on kind; no breaking change. |
| src/openhuman/memory/tinycortex/seal.rs | Adds mark_local_model_unavailable_if_applicable call and structured debug log at the seal embed failure site. Clean addition. |
| src/openhuman/memory/tinycortex/queue_driver.rs | Same mark_local_model_unavailable_if_applicable call added at the reembed failure site, with debug log. Clean addition parallel to seal.rs. |
| app/src/lib/userErrors/tests/classify.test.ts | Adds four new test cases: token path, both prose shapes (daemon-down and model-not-pulled), negative for bare 'daemon unreachable', and negative for credits error naming Ollama. Comprehensive coverage. |
| app/src/services/tests/socketService.events.test.ts | New test verifies that memory-sourced user_error events get scope='memory' and sourceDomain='memory', with no raw message forwarded. |
Sequence Diagram
sequenceDiagram
participant E as EmbedderBridge / queue_driver
participant H as health::classify_embed_error
participant M as mark_local_model_unavailable_if_applicable
participant F as SEMANTIC_RECALL_DEGRADED (AtomicBool)
participant W as publish_web_channel_event
participant S as socketService (frontend)
participant C as classifyUserActionableError
participant UI as UserErrorCenter
E->>H: "classify_embed_error(&error)"
H-->>E: "PipelineFailure { code: LocalModelUnavailable }"
E->>M: "mark_local_model_unavailable_if_applicable(&failure)"
M->>F: load(Acquire) → already_surfaced?
alt "First failure in outage (already_surfaced = false)"
M->>F: mark_semantic_recall_degraded (Release)
M->>W: publish_local_model_unavailable_user_error(embed_classify)
W-->>S: "user_error {error_type: local_model_unavailable, error_source: memory}"
S->>C: "classifyUserActionableError({errorType, scope: memory})"
C-->>UI: "UserErrorDescriptor {kind: local_model_unavailable, action: open_provider_settings}"
else "Subsequent failures (already_surfaced = true)"
M->>F: mark_semantic_recall_degraded (Release)
Note over M: No re-broadcast — panel dedupes anyway
end
Note over W,S: health gate (factories.rs) fires surface_local_model_unavailable_to_clients() above the Sentry latch, covering daemon-down races before socket attaches
Reviews (2): Last reviewed commit: "fix(memory): surface local-model errors ..." | Re-trigger Greptile
![greptile-apps[bot]](https://avatars.githubusercontent.com/in/867647?s=60&v=4){kind=small}
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
Contributor
There was a problem hiding this comment.
Actionable comments posted: 7
🤖 Prompt for all review comments with AI agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
Inline comments:
In `@app/src/lib/i18n/it.ts`:
- Around line 7298-7299: Update the Italian translation for
userErrors.localModelUnavailable.body to use “non è installato” instead of
wording that means “never downloaded,” while preserving the existing Ollama and
cloud-provider guidance.
In `@app/src/lib/i18n/pl.ts`:
- Around line 7270-7271: Update the translation value for
userErrors.localModelUnavailable.body to describe the required Ollama model as
not installed at the configured endpoint, rather than saying it was never
downloaded. Preserve the existing guidance to start Ollama, install the model,
or move the task to a cloud provider.
In `@app/src/lib/i18n/zh-CN.ts`:
- Around line 6779-6780: Update the translation value for
userErrors.localModelUnavailable.body to replace the narrow “从未拉取” wording with
wording that covers any model not installed at the configured Ollama endpoint,
while preserving the existing guidance to start Ollama, obtain the model, or
switch to a cloud provider.
In `@app/src/lib/userErrors/classify.ts`:
- Around line 142-149: Update the isLocalModelUnavailable matcher in
classifyUserActionableError to recognize the raw missing-model message pattern
“Ollama embedding model `<id>` is not installed at <base>” in addition to the
existing signals. Add a regression test in classify.test.ts verifying this
message is classified as the expected actionable Ollama error rather than
returning null.
In `@src/openhuman/memory/tree/health/mod.rs`:
- Around line 271-296: Move the local-model error matching from
src/openhuman/memory/tree/health/mod.rs#L271-L296 into a focused sibling module
and re-export its helper from mod.rs. Move the degradation state helpers from
src/openhuman/memory/tree/health/mod.rs#L487-L508 similarly, keeping mod.rs
export-focused. In src/openhuman/memory/store/factories.rs#L37-L60, retain only
thin factory orchestration and delegate health-gate event behavior; move the
event payload and publisher from
src/openhuman/memory/store/factories.rs#L114-L154 into that dedicated module.
Ensure each Rust file remains at or below 500 lines and preserve existing public
behavior through re-exports.
- Around line 502-507: Add debug-level [domain] or [rpc] correlated tracing for
the local-model failure branch and semantic recall state transition in
src/openhuman/memory/tree/health/mod.rs:502-507, passing safe caller correlation
from src/openhuman/memory/tinycortex/queue_driver.rs:185-187 and a safe
seal-operation correlation from src/openhuman/memory/tinycortex/seal.rs:29-33.
In src/openhuman/memory/store/factories.rs:134-139, use the required prefix and
trace the web-channel publish operation. Log only safe correlation fields; never
include raw errors, endpoints, credentials, or input text.
- Around line 487-508: Update mark_semantic_recall_degraded and
current_degraded_state to publish and read SEMANTIC_RECALL_DEGRADED with
SEMANTIC_RECALL_CAUSE as one atomic state, or under a shared lock, so readers
cannot observe a new degraded flag with a stale cause. Preserve existing
cause-clearing and recovery behavior, and add a concurrent test covering the
flag/cause interleaving.
🪄 Autofix
Fix all unresolved CodeRabbit comments on this PR:
- Push a commit to this branch (recommended)
- Create a new PR with the fixes
ℹ️ Review info
⚙️ Run configuration
Configuration used: Organization UI
Review profile: CHILL
Plan: Pro
Run ID: 19c70dea-dd00-4979-91c4-abb5b7be44f9
📒 Files selected for processing (23)
app/src/lib/i18n/ar.tsapp/src/lib/i18n/bn.tsapp/src/lib/i18n/de.tsapp/src/lib/i18n/en.tsapp/src/lib/i18n/es.tsapp/src/lib/i18n/fr.tsapp/src/lib/i18n/hi.tsapp/src/lib/i18n/id.tsapp/src/lib/i18n/it.tsapp/src/lib/i18n/ko.tsapp/src/lib/i18n/pl.tsapp/src/lib/i18n/pt.tsapp/src/lib/i18n/ru.tsapp/src/lib/i18n/zh-CN.tsapp/src/lib/userErrors/__tests__/classify.test.tsapp/src/lib/userErrors/classify.tsapp/src/services/__tests__/socketService.events.test.tsapp/src/services/socketService.tsapp/src/types/userError.tssrc/openhuman/memory/store/factories.rssrc/openhuman/memory/tinycortex/queue_driver.rssrc/openhuman/memory/tinycortex/seal.rssrc/openhuman/memory/tree/health/mod.rs
Review follow-ups on tinyhumansai#5354. The embedder health gate probes `GET /api/tags`, which succeeds whenever the daemon is up. A running daemon whose embedding model was never pulled therefore never tripped that gate, so the "model not installed" half of `LocalModelUnavailable` set the degraded flag but never raised the durable UserErrorCenter entry. The failure classifier now publishes it as well, covering both halves from the one place that has actually classified the failure. It fires on the transition into the state rather than per failed embed, since the re-embed path calls it per row. The payload and publisher move to a dedicated `health::user_error` module so both producers emit one identical, tested shape. Publish the degraded cause before its flag, and pair a Release store with an Acquire load, so a concurrent status read cannot observe a freshly-set flag alongside the previous degradation's cause and render the wrong remediation. Teach the frontend classifier the second Ollama prose shape ("embedding model ... is not installed at"). It only recognised the daemon-down shape, so a raw message carrying the model-not-pulled text fell through to null despite the comment claiming both were covered. Reword the remediation copy from "never pulled" to "not installed" across all fourteen locales, English included: the backend condition also covers a model that was removed or lives on a different endpoint. Add correlated debug traces at both classification sites, carrying only the embedder identity, operation, and typed outcome.
The constant is only referenced across module boundaries from a test, so re-exporting it from the health module left an unused import in the non-test build, which `clippy -D warnings` rejects. Expose the module itself instead and let the one cross-module consumer path to it.
There was a problem hiding this comment.
YellowSnnowmann has reached the 50-credit limit for trial accounts. To continue receiving code reviews, upgrade your plan.
Contributor
There was a problem hiding this comment.
Actionable comments posted: 5
🤖 Prompt for all review comments with AI agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
Inline comments:
In `@app/src/lib/i18n/bn.ts`:
- Line 7166: Update the userErrors.localModelUnavailable message to state that
the required model must be installed at the configured Ollama endpoint,
including when it exists on another Ollama instance. Apply this wording
consistently in the source translation and every locale file, including the
Bengali entry shown here.
In `@app/src/lib/i18n/en.ts`:
- Line 7563: Update the userErrors.localModelUnavailable translation to clarify
that the required model must be installed and available at the configured Ollama
endpoint, including cases where it is missing or hosted on another instance;
retain the guidance to switch to a cloud provider.
In `@app/src/lib/i18n/fr.ts`:
- Line 7346: Update the French `localModelUnavailable` translation so it is
domain-neutral for chat, cron, and memory usage, and explicitly states that the
required model must be installed on the configured Ollama endpoint. Replace the
task-specific remediation wording with endpoint-focused guidance to start the
endpoint and install the model there, or use a cloud provider.
In `@src/openhuman/memory/tree/health/mod.rs`:
- Around line 523-539: Make the check-and-update flow around
SEMANTIC_RECALL_DEGRADED, SEMANTIC_RECALL_CAUSE, and
mark_semantic_recall_degraded atomic so only one concurrent caller publishes the
local-model-unavailable user error. Serialize the state check, update, and
publish with a lock or atomic state machine, preserving the existing failure
classification and logging behavior; add a deterministic concurrent test proving
only one transition is published.
- Around line 523-539: Update the semantic-recall degradation flow around
mark_semantic_recall_degraded and publish_local_model_unavailable_user_error so
a missing-model error is delivered when no client was subscribed at the initial
failure. Persist the active LocalModelUnavailable notification for replay on
socket subscription, or implement bounded re-emission until recovery, while
retaining suppression after successful delivery. Add a test covering failure
before subscription followed by subscription and another failed embed.
🪄 Autofix
Fix all unresolved CodeRabbit comments on this PR:
- Push a commit to this branch (recommended)
- Create a new PR with the fixes
ℹ️ Review info
⚙️ Run configuration
Configuration used: Organization UI
Review profile: CHILL
Plan: Pro
Run ID: b6767bd8-3c6c-4e61-8b6b-f48bcf7dfe6e
📒 Files selected for processing (21)
app/src/lib/i18n/ar.tsapp/src/lib/i18n/bn.tsapp/src/lib/i18n/de.tsapp/src/lib/i18n/en.tsapp/src/lib/i18n/es.tsapp/src/lib/i18n/fr.tsapp/src/lib/i18n/hi.tsapp/src/lib/i18n/id.tsapp/src/lib/i18n/it.tsapp/src/lib/i18n/ko.tsapp/src/lib/i18n/pl.tsapp/src/lib/i18n/pt.tsapp/src/lib/i18n/ru.tsapp/src/lib/i18n/zh-CN.tsapp/src/lib/userErrors/__tests__/classify.test.tsapp/src/lib/userErrors/classify.tssrc/openhuman/memory/store/factories.rssrc/openhuman/memory/tinycortex/queue_driver.rssrc/openhuman/memory/tinycortex/seal.rssrc/openhuman/memory/tree/health/mod.rssrc/openhuman/memory/tree/health/user_error.rs
🚧 Files skipped from review as they are similar to previous changes (12)
- app/src/lib/userErrors/tests/classify.test.ts
- app/src/lib/userErrors/classify.ts
- app/src/lib/i18n/id.ts
- app/src/lib/i18n/ko.ts
- app/src/lib/i18n/ru.ts
- app/src/lib/i18n/pt.ts
- app/src/lib/i18n/es.ts
- app/src/lib/i18n/it.ts
- app/src/lib/i18n/de.ts
- src/openhuman/memory/tinycortex/queue_driver.rs
- app/src/lib/i18n/hi.ts
- src/openhuman/memory/tinycortex/seal.rs
Review round two on tinyhumansai#5354. The transition check read the degraded flag and then set it, so two concurrent embed tasks could both conclude they were first and both publish. Claim the announcement with a compare_exchange on a dedicated latch instead, making check-and-claim indivisible. The latch is released by `clear_semantic_recall_degraded`, which every write-embedder build calls once per seal or re-embed operation. That turns the single announcement into bounded re-emission until recovery, which is what covers a client that was not connected when the outage began: `publish_web_channel_event` is an unbuffered broadcast with no replay, so an announcement made before anyone subscribed is simply gone. Make the remediation copy endpoint-aware and domain-neutral in all fourteen locales. With more than one Ollama instance reachable, "start Ollama and pull the model" can send the user to repair the wrong machine, and the string is shared by the chat, cron, and memory scopes so it must not describe the work as a task.
There was a problem hiding this comment.
YellowSnnowmann has reached the 50-credit limit for trial accounts. To continue receiving code reviews, upgrade your plan.
coderabbitai
Bot
removed
bug
rust-core
M3gA-Mind
approved these changes
Aug 5, 2026
M3gA-Mind
left a comment
M3gA-Mind
left a comment
Collaborator
There was a problem hiding this comment.
Reviewed against upstream/main: CI green on every lane that applies to the changed areas, no unresolved review threads, and the diff does what the title and body claim. Checked correctness, blast radius on shared seams, that the behaviour change is pinned by a test rather than only asserted, and i18n/em-dash compliance on any new user-facing string. Nothing blocking found.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
daemon not listening,model never pulled) aslocal_model_unavailableinstead of a generic transient fault, so the memory status panel names the actual fix.requeue_transient_failedauto-resumes ingestion once Ollama is back, with no manual "Retry failed" click.user_errorweb-channel path, landing a durable UserErrorCenter entry.DomainEvent::EmbeddingModelUnhealthywas published but reached no user.local_model_unavailableuser-error kind +memoryscope, and derive the panel scope from the producing domain instead of hardcodingcron.Problem
The issue reports that the memory embedder still calls the deprecated Ollama
/api/embeddingsendpoint removed in v0.32.x. That root cause is not present in the tree. The embedder has calledPOST /api/embedwith theinputfield since the provider port (vendor/tinyagents/src/harness/embeddings/ollama.rs), and/api/embeddingsappears nowhere in the repository. So acceptance criteria 1 and 2 already held.The real defect is criteria 3 and 4, and it is what the reporter actually experienced. When Ollama is not running, or the configured model was never pulled, the embedder produces two messages that already carry the fix in their text:
Neither carries an
Embedding API error (<status>)envelope — the first is a transport bail, the second a rewritten 404 — soclassify_embed_error_strmatched none of its rules and fell through toTransient. The panel then renderedmemory.health.remediation.transient: "A temporary error interrupted memory processing. It will retry automatically." Retrying cannot start a daemon or pull a model, so the pipeline retried indefinitely and the user was never told what to do.FailureCode::LocalModelUnavailableand its translated remediation ("Install/run Ollama and pull the model…") already existed in all 14 locales — with no producer anywhere in the tree.Separately, the memory-store health gate detects an unreachable daemon and publishes
DomainEvent::EmbeddingModelUnhealthy, but nothing carries the domain bus to the product UI:/events/domainis consumed only by the developer Event Log panel, and there is noapp/srclistener. That notice reached no user either.Solution
Classification (
memory/tree/health/mod.rs) — match the two shapes explicitly, anchored on Ollama-specific wording so a generic cloud-embedder transport failure keeps itsTransientcode. Placed after the client-side bails and before status parsing.Retry class —
LocalModelUnavailablemoves intoFailureClass::Transient. This is the key design decision.requeue_transient_failed— the automatic self-healing requeue — skips rows recorded as unrecoverable. Since the condition clears from outside the app, classifying it unrecoverable would park every affected job until someone pressed "Retry failed" by hand. Transient is also consistent withderive_pipeline_status, which escalates toerroronly on unrecoverable failures precisely because transient ones self-heal. The panel banner renders on the presence of a typed cause, not on its class, so the remediation shows either way.Immediate surfacing — a persisted
failure_reasonis only read back once a job settles terminally, which for a transient class means after the whole retry budget drains.mark_local_model_unavailable_if_applicableflips the semantic-recall flag at classification time so the remediation is on screen from the first failure; it self-clears on the next successful embed.UI bridge (
memory/store/factories.rs) — broadcast the condition as auser_errorweb-channel event, the same pathpublish_cron_user_erroralready uses, whichsocketServiceroutes into the durable UserErrorCenter. Metadata-only: a stable kind token pluserror_source, never raw provider text or the configured endpoint.The broadcast deliberately sits above the once-per-process Sentry latch.
publish_web_channel_eventis an unbufferedbroadcast::send— with no socket client attached the event is dropped with no redelivery — and memory is constructed early, before the renderer's socket attaches. Under the latch, that one dropped send would have been the only attempt ever made and the panel would have stayed empty for the whole outage. Re-broadcasting per failed probe is safe: the panel store dedupes on the descriptor'skind:scope:provideridentity and bumpscount. Only the Sentry report stays latched, which is what the latch was introduced for.Frontend — new
local_model_unavailablekind andmemoryscope, a classifier rule ordered last so an out-of-credits provider keeps its billing remediation, andsocketServicenow derives scope fromerror_source. Unknown domains keep the historicalcrondefault, so existing behaviour is unchanged. Each prose matcher is anchored on full producer wording rather than a baredaemon unreachable at, which backend connection-health logs also emit — the same deliberate anchoring the Rust matcher uses.Submission Checklist
diff-cover) meet the gate enforced by.github/workflows/ci-lite.yml. Runpnpm test:coverageandpnpm test:rustlocally; PRs below 80% on changed lines will not merge.N/A: behaviour-only change to existing memory-tree health classification; no feature rows added, removed, or renamed## RelatedN/A: no release-cut surface touchedCloses #NNNin the## RelatedsectionTests added:
memory/tree/healthLocalModelUnavailable; model-not-pulled → same; through an anyhow context chain; negative: non-Ollama transport error staysTransient; degraded flag set with its own cause; negative: other codes do not flip the flagmemory/store/factoriesuser_errorpayload is metadata-only (no prose, no endpoint); broadcast survives the Sentry latch (subscribes to the bus, asserts both calls emit); kind token matches the frontend discriminatorlib/userErrors/classifydaemon unreachablefrom another domain is not promoted; negative: a credits error naming Ollama keeps billing remediationservices/socketServiceuser_errorscopes tomemory, no raw message forwardedImpact
local_model_unavailableinstead oftransient. The panel status moves from a generic retry notice to the actionable remediation; the pipeline status pill staysdegradedrather thanerror, matchingderive_pipeline_status' contract for self-healing failures.PipelineFailure's wire shape is unchanged;latest_failed_job_failurestill trusts the persisted class column.UserErrorKindandUserErrorScopeunions are widened additively. Nothing branches exhaustively onkind; the UI keys onaction,severity, andscope, all of which are covered. Theuser_errorhandler keeps itscrondefault for unknown domains.Related
is_ollama_user_config_rejection(core/observability.rs) anchors every arm onollama embed failed, which the rewritten model-not-installed message no longer contains. Possibly a stale Sentry matcher; not touched here because I could not prove that message reaches a report site from this path.AI Authored PR Metadata (required for Codex/Linear PRs)
Linear Issue
Commit & Branch
fix/5354-ollama-embed-error-classification135339ba3Validation Run
pnpm --filter openhuman-app format:checkpnpm typecheckcargo test --lib -- memory::tree::health memory::store::factories(56 passed),cargo test --lib memory::tinycortex(32 passed),vitest src/lib/userErrors src/services/__tests__/socketService.events.test.ts(27 passed), plus avitest relatedsweep over the changed files (432 files / 4986 tests passed)cargo fmt --all --checkclean;cargo check --libclean; slim--no-default-features --features tokenjuice-treesittercleanapp/src-taurichange. See Validation Blocked.Validation Blocked
command:pnpm rust:checkerror:failed to load source for dependency 'tauri' … unable to update app/src-tauri/vendor/tauri-cef/crates/tauri … No such file or directory (os error 2)impact:Environment only — thetauri-cefsubmodule is not initialised in this worktree; noapp/src-taurifile is touched by this PR. The risk it stands in for is the shell'sdefault-features = falsebuild of the core, which was checked directly and compiles clean.Behavior Changes
local_model_unavailable(transient class) instead oftransient, sets the semantic-recall degraded flag on first failure, and raises a durable UserErrorCenter entry.Parity Contract
Transient). Theuser_errorsocket handler keeps itscronscope for the existing producer and for payloads with noerror_source, pinned by the two pre-existing tests, which are unmodified.FailureCoderound-trip and the all-codes class/remediation guard both updated in lockstep with the class change.PipelineFailureserde shape unchanged. The Ollama matcher is anchored on producer-specific wording on both sides of the language boundary so a wording drift fails a test rather than silently dropping the signal.Duplicate / Superseded PR Handling
Summary by CodeRabbit
New Features
Localization
Bug Fixes