feat(routing): weighted policy for load balancing across providers - #553
Conversation
github-actions
Bot
added
the
missing-template
|
Warning Review limit reached
Next review available in: 12 seconds You've used all free OSS reviews for now. Wait for the free limit to reset to keep reviewing this public repository. How can I continue?After more reviews become available, a review can be triggered using the To avoid repeated limits, reduce automatic review volume by pausing incremental auto-reviews earlier, using label-based review opt-in, excluding WIP or generated PR titles, or requesting reviews manually when the PR is ready. If your team needs uninterrupted high-volume reviews, an organization admin can enable usage-based reviews. How do review limits work?CodeRabbit enforces per-developer PR review limits for each organization. Most developers receive the normal plan review availability. For paid Pro and Pro+ PR reviews, CodeRabbit uses adaptive limits for sustained high-volume activity. When a developer's recent PR review activity reaches the 95th percentile or higher among CodeRabbit users, additional reviews become available more gradually as earlier reviews age out of the rolling window. Please refer docs for additional details. Review details⚙️ Run configurationConfiguration used: Path: .coderabbit.yaml Review profile: CHILL Plan: Pro Plus Run ID: 📒 Files selected for processing (1)
WalkthroughThis PR adds weighted provider routing with validated traffic shares, weighted failover, caller-specific explanations, API and CLI metadata, dashboard configuration flows, documentation, generated dashboard assets, and unit and integration coverage. ChangesWeighted routing core
Estimated code review effort: 5 (Critical) | ~120 minutes Possibly related PRs
Suggested labels: Suggested reviewers: 🚥 Pre-merge checks | ✅ 4 | ❌ 1❌ Failed checks (1 warning)
✅ Passed checks (4 passed)
✨ Finishing Touches 💡 2📝 Generate docstrings 💡
🛠️ Fix failing CI checks 💡
🧪 Generate unit tests (beta)
✨ Simplify code
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
github-actions
Bot
removed
the
missing-template
There was a problem hiding this comment.
Pull request overview
Adds a new weighted router backend to the gateway’s routing system so a single policy can load-balance traffic across multiple provider:model candidates according to operator-declared shares, with supporting updates across API surfaces (explain + OpenAPI), CLI output, dashboard UI, and tests.
Changes:
- Introduces
weightedrouting backend with schema validation, deterministic “explain” rendering, and request-path integration. - Updates dashboard routing editor/table to display and author weighted splits; adjusts routing-memory surfaces to apply only to learned (
knn) policies. - Adds unit + integration coverage and updates docs/OpenAPI + rebuilt dashboard bundle artifacts.
Reviewed changes
Copilot reviewed 34 out of 37 changed files in this pull request and generated 2 comments.
Show a summary per file
| File | Description |
|---|---|
| web/src/pages/RoutingPage.tsx | Dashboard routing UI: weighted split authoring, labeling, summaries, Examples gating |
| web/src/pages/RoutingPage.test.tsx | UI tests for weighted routing creation/editing/validation and labeling behavior |
| web/src/api/types.ts | Adds weights to policy select entries and router_weights to explain response typing |
| tests/unit/test_weighted_router.py | Unit tests for weighted schema rules, draw properties, and explain behavior |
| tests/integration/test_routing_weighted.py | End-to-end integration tests for weighted routing behavior and management surfaces |
| src/gateway/services/routing/weighted.py | Implements weighted backend (draw, share normalization, explain ordering) |
| src/gateway/services/routing/decide.py | Passes weights into routing context; adds explain_router_ordering helper; log level control |
| src/gateway/services/routing/backends.py | Registers weighted backend; adds backend capability helpers (learns/pricing/weighted) |
| src/gateway/services/routing/init.py | Re-exports weighted backend constant and backend capability helpers |
| src/gateway/models/routing.py | Schema support for weights, validation rules, and WEIGHTED_BACKEND constant |
| src/gateway/api/routes/routing.py | explain now includes weighted ordering and returns router_weights; pricing validation scoped to learned backend |
| src/gateway/api/routes/routing_memory.py | Treats only learned backends as routing-memory consumers (excludes weighted) |
| src/gateway/cli.py | CLI routing explain renders weighted shares and uses explain-time weighted ordering |
| src/gateway/services/pricing_init_service.py | Startup pricing warnings scoped to backends that actually require pricing |
| docs/routing.md | Documents weighted routing semantics, constraints, and how to inspect behavior |
| docs/dashboard.md | Documents new weighted split control and behavior on the dashboard |
| docs/api-reference.md | Notes router_weights in /v1/routing/policies/explain for weighted policies |
| docs/public/openapi.json | Regenerated OpenAPI schema including router_weights and updated descriptions |
| src/gateway/AGENTS.md | Updates internal routing architecture notes to include weighted backend |
| src/gateway/static/dashboard/index.html | Updated dashboard bundle entrypoint hash |
| src/gateway/static/dashboard/assets/RoutingPage-Dqh6Cebs.js | New compiled dashboard chunk containing weighted routing UI changes |
| src/gateway/static/dashboard/assets/RoutingPage-B0PbG8Zg.js | Removed old compiled routing page chunk (replaced by new hash) |
| src/gateway/static/dashboard/assets/UsersPage-BUrrwJKv.js | Rebuilt dashboard asset (hash/import updates) |
| src/gateway/static/dashboard/assets/ToolsGuardrailsPage-mzjN9zP-.js | Rebuilt dashboard asset (hash/import updates) |
| src/gateway/static/dashboard/assets/TablePagination-CdsMK2Es.js | Rebuilt dashboard asset (hash/import updates) |
| src/gateway/static/dashboard/assets/OverviewPage-BGv4NjCr.js | Rebuilt dashboard asset (hash/import updates) |
| src/gateway/static/dashboard/assets/ModelScopeControl-BR6e2y7T.js | Rebuilt dashboard asset (hash/import updates) |
| src/gateway/static/dashboard/assets/KeysPage-B_esvNYQ.js | Rebuilt dashboard asset (hash/import updates) |
| src/gateway/static/dashboard/assets/ConfirmDialog-CP3FsMvH.js | Rebuilt dashboard asset (hash/import updates) |
Files not reviewed (1)
- src/gateway/static/dashboard/assets/RoutingPage-Dqh6Cebs.js: Generated file
Suppressed comments (1)
web/src/pages/RoutingPage.tsx:153
routerBackendOfreturns the rawrouterstring, but backend resolution is case/whitespace-insensitive. Downstream UI logic (labels, summaries, Examples button gating) can mis-handle valid backends if the config/API stored a different casing or included whitespace.
💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
There was a problem hiding this comment.
Actionable comments posted: 4
🤖 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 `@config.example.yml`:
- Line 142: Update the `on_failure` comment in the configuration example to
specify that it applies to an eligible provider failure occurring before any
response bytes are sent; preserve the existing description that it controls what
is tried next and in what order.
In `@docs/routing.md`:
- Around line 158-160: Qualify zero-share wording in both user-facing
descriptions: in docs/routing.md lines 158-160, change “serving none of the
traffic” to “receiving none of the weighted traffic”; in docs/dashboard.md lines
308-314, change “takes no traffic” to “takes no weighted traffic.”
In `@src/gateway/static/dashboard/assets/index-D0MBX8Ui.js`:
- Around line 1-2: Regenerate the dashboard entrypoint and all dependent chunks
from one shared source revision so their export bindings are consistent. At
src/gateway/static/dashboard/assets/index-D0MBX8Ui.js lines 1-2, rebuild the
shared runtime and commit every regenerated hashed asset; at
src/gateway/static/dashboard/assets/SettingsPage-BtxuYOZY.js line 1, restore the
y binding and align loading, pricing, and select imports; at
src/gateway/static/dashboard/assets/ToolsGuardrailsPage-mzjN9zP-.js line 1,
replace the undeclared O() reference and correct tool query and mutation
imports; at src/gateway/static/dashboard/assets/RoutingPage-Dqh6Cebs.js line 1,
bind $e to the tool-settings query rather than the pricing-delete mutation.
Smoke-test /settings, /tools, and /routing.
In `@src/gateway/static/dashboard/assets/RoutingPage-Dqh6Cebs.js`:
- Line 1: Update Ve() to use the tool-settings query export rather than the
pricing-delete mutation currently bound to $e. Rebuild the chunk against the
current shared export map or import the correct tool-settings hook, preserving
Ve()’s guardrails_url lookup and configured state behavior.
🪄 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: Path: .coderabbit.yaml
Review profile: CHILL
Plan: Pro Plus
Run ID: 114eee14-d94f-4c8b-b678-f817ed403ace
⛔ Files ignored due to path filters (1)
docs/public/openapi.jsonis excluded by!docs/public/openapi.json
📒 Files selected for processing (38)
config.example.ymldocs/api-reference.mddocs/dashboard.mddocs/index.mddocs/routing.mdsrc/gateway/AGENTS.mdsrc/gateway/api/routes/routing.pysrc/gateway/api/routes/routing_memory.pysrc/gateway/cli.pysrc/gateway/models/routing.pysrc/gateway/services/pricing_init_service.pysrc/gateway/services/routing/__init__.pysrc/gateway/services/routing/backends.pysrc/gateway/services/routing/decide.pysrc/gateway/services/routing/weighted.pysrc/gateway/static/dashboard/assets/ActivityPage-C9_HiWjD.jssrc/gateway/static/dashboard/assets/BudgetsPage-BSua8jzY.jssrc/gateway/static/dashboard/assets/ConfirmDialog-CP3FsMvH.jssrc/gateway/static/dashboard/assets/DocsPage-CwwWDa-K.jssrc/gateway/static/dashboard/assets/KeysPage-B_esvNYQ.jssrc/gateway/static/dashboard/assets/ModelScopeControl-BR6e2y7T.jssrc/gateway/static/dashboard/assets/ModelsPage-NwBGkmaA.jssrc/gateway/static/dashboard/assets/OverviewPage-BGv4NjCr.jssrc/gateway/static/dashboard/assets/ProvidersPage-DOyggcy8.jssrc/gateway/static/dashboard/assets/RoutingPage-B0PbG8Zg.jssrc/gateway/static/dashboard/assets/RoutingPage-Dqh6Cebs.jssrc/gateway/static/dashboard/assets/SettingsPage-BtxuYOZY.jssrc/gateway/static/dashboard/assets/TablePagination-CdsMK2Es.jssrc/gateway/static/dashboard/assets/ToolsGuardrailsPage-mzjN9zP-.jssrc/gateway/static/dashboard/assets/UsagePage-MIcf1ELP.jssrc/gateway/static/dashboard/assets/UsersPage-BUrrwJKv.jssrc/gateway/static/dashboard/assets/index-D0MBX8Ui.jssrc/gateway/static/dashboard/index.htmltests/integration/test_routing_weighted.pytests/unit/test_weighted_router.pyweb/src/api/types.tsweb/src/pages/RoutingPage.test.tsxweb/src/pages/RoutingPage.tsx
💤 Files with no reviewable changes (1)
- src/gateway/static/dashboard/assets/RoutingPage-B0PbG8Zg.js
khaledosman
left a comment
khaledosman
left a comment
There was a problem hiding this comment.
Gates pass on 49d5306: make lint (architecture + Ruff), make typecheck, openapi-check, postman-check, tests/unit/test_weighted_router.py, and web RoutingPage.test.tsx. The committed dashboard bundle carries the new RoutingPage/DocsPage output and is referenced from index-D0MBX8Ui.js.
Two inline: one confirmed regression for the noop placeholder backend, one cosmetic note on the share input.
🤖 Review by Claude Opus 5 via Claude Code
GitHub Actions
Failure
|
Note: this comment was drafted by Claude Opus 5 via back-and-forth with @njbrake. The reasoning and decisions are his; the prose is Claude's. All eight threads addressed in 6879b14: six fixes, two disagreements. Resolving them all. @khaledosman on @khaledosman on the share field, and Copilot on non-finite input. Fixed together. The field now holds the typed text and parses once, so Copilot on backend-name normalization. Fixed. CodeRabbit on the zero-share wording. Taken, in both guides: a zero-weight candidate takes no weighted traffic, and still serves opt-out callers when it is also the default target. CodeRabbit on CodeRabbit on the dashboard chunks (Critical and Major). Not reproducible; resolving as false positives. |
GitHub Actions
Failure
There was a problem hiding this comment.
Actionable comments posted: 2
🤖 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 `@src/gateway/static/dashboard/assets/index-DB9mxJVs.js`:
- Line 2: Update the session helpers at and around at(), ot(), and the
AuthProvider callbacks in ut() to serialize authentication requests with an
in-flight guard, ensuring logout completes before clearing state or allowing a
new login. Apply the existing Ee timeout through the same timeout mechanism used
by i() to both POST and DELETE session requests, and await the logout request
before showing the login screen.
In `@web/src/pages/RoutingPage.test.tsx`:
- Around line 550-596: The existing tests only verify display and editability,
not successful serialization. Extend the fractional-share test to open the
editor, save, and assert the POST payload preserves 7.5; extend the loosely
spelled backend test to edit and save, then assert the emitted weighted
specification uses the normalized backend path and expected weights.
🪄 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: Path: .coderabbit.yaml
Review profile: CHILL
Plan: Pro Plus
Run ID: c5cc25cc-5b34-4ffe-ae0f-be1a4af150a4
📒 Files selected for processing (26)
config.example.ymldocs/dashboard.mddocs/routing.mdsrc/gateway/api/routes/routing_memory.pysrc/gateway/services/routing/__init__.pysrc/gateway/services/routing/backends.pysrc/gateway/static/dashboard/assets/ActivityPage-D6-YEyjz.jssrc/gateway/static/dashboard/assets/BudgetsPage-DF21khAQ.jssrc/gateway/static/dashboard/assets/ConfirmDialog-Bk4-ghkz.jssrc/gateway/static/dashboard/assets/DocsPage-BIu60Z5q.jssrc/gateway/static/dashboard/assets/KeysPage-CVbrPDbf.jssrc/gateway/static/dashboard/assets/ModelScopeControl-DWhpSerS.jssrc/gateway/static/dashboard/assets/ModelsPage-47hJWPEK.jssrc/gateway/static/dashboard/assets/OverviewPage-kp3aG0sw.jssrc/gateway/static/dashboard/assets/ProvidersPage-B5b0u3cx.jssrc/gateway/static/dashboard/assets/RoutingPage-C5-iZGTv.jssrc/gateway/static/dashboard/assets/SettingsPage-DUpymCp3.jssrc/gateway/static/dashboard/assets/TablePagination-bQsnzK0q.jssrc/gateway/static/dashboard/assets/ToolsGuardrailsPage-B0vzdnDJ.jssrc/gateway/static/dashboard/assets/UsagePage-CWRc02TG.jssrc/gateway/static/dashboard/assets/UsersPage-DykLjCx3.jssrc/gateway/static/dashboard/assets/index-DB9mxJVs.jssrc/gateway/static/dashboard/index.htmltests/unit/test_weighted_router.pyweb/src/pages/RoutingPage.test.tsxweb/src/pages/RoutingPage.tsx
🚧 Files skipped from review as they are similar to previous changes (6)
- config.example.yml
- src/gateway/api/routes/routing_memory.py
- docs/routing.md
- docs/dashboard.md
- src/gateway/services/routing/init.py
- web/src/pages/RoutingPage.tsx
|
@njbrake apologies for closing the other PR, I ran into rebase/merge conflicts because other commits were touching the static asset files. I decided to start over and first remove the static assets from the Git repository in favor of compiling them at build time. I'd suggest merging #556 before this one. I've already prepared a branch that will merge cleanly once #556 lands: https://github.com/pocketsizesun/otari/tree/feat/weighted-load-balancing |
|
@pss-julien Great, thank you! Yes no worries about this. I will take the PRs from here and get them over the finish line. Thanks for the great ideas and solid contributions. All the PRs have your cmmits in them so you'll get contribution credit for them 🙏 |
A `select` entry can now name the `weighted` router, which draws one
candidate per request in proportion to declared weights, so traffic
splits across providers behind one policy name.
select:
- router: weighted
candidates: [openai:gpt-5, anthropic:claude-sonnet-4-5]
weights: {openai:gpt-5: 70, anthropic:claude-sonnet-4-5: 30}
- default: openai:gpt-5
Design choices worth recording:
* Weights are normalized rather than percentages, so 7:3 and 70:30 are
one split. A candidate left out of `weights` gets zero, which drains a
provider without deleting it: it takes no traffic and stays at the tail
of the plan as a failover target. An all-zero split is refused, since
it could never select anything.
* Each request is an independent draw and nothing is remembered, so the
split is as correct behind many replicas as behind one and none of the
stickiness machinery in docs/routing-scaling.md is needed. The cost is
that the ratio converges statistically.
* The draw continues without replacement, so the whole ordering is the
plan: a retryable failure lands on another weighted provider before
`on_failure` is reached, and a failing provider sheds its share with no
health tracking.
* No pricing is required, unlike the kNN router which scores cost. The
pricing gate is now scoped by `backend_requires_pricing`, so a weighted
policy is not refused for missing rows it never reads.
* A weighted decision is not logged per request (it would be the log at
load balancer volume); the usage row carries `router:weighted` and the
draw is available at debug.
* `explain` and the CLI render the split itself, normalized over the
candidates the caller may use, because a weighted decision needs no
request state. The declared pool is still handed to the compiler so a
filtered candidate keeps appearing in `dropped` with its reason.
* `Otari-Router: off` keeps its meaning and serves the default target,
which pins a caller to one provider during an incident.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
… literals `_rank` takes `dict[str, float]`, and a dict bound to a variable before the call is inferred as `dict[str, int]`, which is invariant and so not accepted. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
A pool with a repeated model was authorable in the policy form: the save validation only checked that each row was non-empty. On a weighted policy the duplicate then collapsed in the weight map, so the two rows the operator saw became one key and the split submitted was not the split shown; the API refuses a repeated candidate on any router entry, so the form was authoring a 400. Uniqueness is now checked over the trimmed, non-empty rows, which leaves a pair of still-empty rows reading as unfinished rather than as a duplicate. Also name an unknown router backend neutrally. The chip and the row summary called every non-weighted backend "Learned", which is true of knn and would be a guess about the first backend added after that line was written; such a policy is already shown read-only, so "Routed" is what the table can honestly say. Adds a unit test for a drained pool of two, which pins a deliberate divergence: declared_shares reports an even split because it has no order to express, while weighted_ordering keeps declared order so the last resort stays predictable. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Five review items on the weighted router. `declared_shares` returned an even split for an all-zero pool while `weighted_ordering` kept declared order, so `explain`, `router_weights`, the CLI and `confidence` all reported 50/50 for a plan that sent every request to the first candidate. Reachable when a caller's allow-list filters out every positively weighted candidate. The reported split now follows the head that actually serves. The routing memory surfaces treated a weighted policy as a learned one, because both name a router. `GET /v1/routing/status` listed it under a warmth report it never reads, and worse, `POST /v1/routing/preferences/rank` let its pool decide which score keys a user may teach, so a workspace whose only router policy was weighted could no longer seed examples for the learned policy it was preparing. Both now gate on `backend_learns`. "Retryable failure" is stale phrasing: #546 made the walker fall back on every pre-response failure and rewrote the docs accordingly. Restored in the new prose, the CLI output and the test comments. The allow-list `explain` transcript omitted the `on_failure` candidate, which is filtered by the same rule, so the example showed one drop where the command prints two. Also rebased onto main, rebuilt the dashboard bundle, regenerated the OpenAPI spec, and named `weighted.py` in the gateway's backend map. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Review findings on #553. When a caller's allow-list filtered out every candidate in a weighted split, `explain_router_ordering` returned nothing, so the compiler fell through to the failure chain and the split disappeared from the plan and from `dropped` alike. The command then printed the learned router's decline text, which talks about cold pools and low confidence on a policy that has neither. It now returns the declared ordering with empty shares, so every filtered candidate is named with its reason, and the CLI keys its closing note on the backend rather than on whether any share survived. The pricing claim was too broad: the weighted router reads no price, but `require_pricing` is on by default, so a metered caller drawn onto an unpriced candidate still gets a 402. Narrowed in the docs and in the module docstring, which the same page already contradicted. Also: the docs said the dashboard renders the explain endpoint's `router_weights`, which it does not (the page computes the declared split client-side); the candidate cap rule still said "for a learned policy"; `docs/index.md` did not mention the weighted split; and `config.example.yml` still said "retryable failure", stale since #546. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
… share field Review findings from @khaledosman and Copilot on #553. `backend_learns` narrowed the routing memory surfaces to the kNN backend, which also dropped `noop`. That backend exists to hold a policy's shape while its pool is being taught, so its candidates are the very ones an operator is seeding: `GET /v1/routing/status` stopped reporting the pool's warmth and `POST /v1/routing/preferences/rank` started refusing the examples meant for it, which is the failure this change was made to prevent, landing on the workflow the backend exists for. Replaced with `backend_pool_is_teachable`, false only for the weighted backend, so `noop` and a name from a newer build keep the behavior they had. The dashboard compared a backend name without normalizing it, so a policy spelled `" Weighted "`, which the gateway resolves happily, was shown read-only and labelled wrong. The share field parsed on every keystroke, so a decimal point was swallowed as it was typed, a cleared field snapped to 0, and "Infinity" passed the form's check to be refused by the API. The field now holds what was typed and parses once, with non-finite and negative values refused where the API refuses them. Also qualified the zero-share wording in both user guides: a zero-weight candidate takes no *weighted* traffic, and still serves opt-out callers when it is the default target. Same for the `on_failure` line in config.example.yml, which did not say failover stops once a response has started. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
6879b14 to
dd1fd57
Compare
…t the form Both new cases stopped at what the form displayed. The fractional-share test now saves and asserts the POST body carries 7.5, and the loosely spelled backend test now saves and asserts the emitted entry, which also pins that the spelling is normalized on the way out. Confirmed both fail against the pre-fix behavior. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
Note: this comment was drafted by Claude Opus 5 via back-and-forth with @njbrake. The reasoning and decisions are his; the prose is Claude's. Two more threads from the latest CodeRabbit pass, both now closed out in 31a161c. Edge-case save paths ( Session request serialization. Out of scope here, so resolving without a change. The logout race is real ( |
Description
Note: this description was drafted by Claude Opus 5 via back-and-forth with @njbrake. The reasoning and decisions are his; the prose is Claude's.
Continues #545 by @pss-julien, whose three commits are preserved here with their authorship. That PR's head branch was the fork's
main, which has since been reset, and GitHub does not honor "allow edits by maintainers" for a fork owned by an organization, so the branch moved here to carry the review fixes.What it adds: a
selectentry can name theweightedrouter, which draws one candidate per request in proportion to declared weights, so traffic splits across providers behind one policy name. The design rationale is recorded in #545 and indocs/routing.md.On top of @pss-julien's work, one review commit:
declared_sharesreported an even split for an all-zero pool while the draw kept declared order, soexplain,router_weights, the CLI, andconfidenceall printed a split the runtime does not serve./v1/routing/statusandpreferences/ranktreated a weighted policy as a learned one. The second is the sharper edge: a workspace whose only router policy was weighted could no longer seed examples for a learned policy it was preparing.explaintranscript omitted theon_failuredrop, which the same filter removes.PR Type
Relevant issues
Continues #545. No issue filed.
Checklist
tests/unit,tests/integration).make lint,make typecheck,make test). Unit and web suites plusopenapi-checkandpostman-checkran locally; the integration suite needs PostgreSQL and ran on CI.uv run python scripts/generate_openapi.py).AI Usage
AI Model/Tool used:
Claude Opus 5 via Claude Code, for the review fixes and this description. The three underlying commits were contributed by @pss-julien, who reported them as fully AI generated with the same tool in #545.
Any additional AI details you'd like to share:
Two places worth a human eye. The
declared_shareschange reverses a behavior @pss-julien deliberately pinned in a test, so it is a decision to confirm rather than a bug fix everyone agrees on. The routing memory change narrows whatPOST /v1/routing/preferences/rankaccepts as a teachable score key, which is a behavior change for any workspace running a weighted policy.🤖 Generated with Claude Code
Summary