feat(validation): reject provably-doomed inference requests at ingress by pjb157 · Pull Request #1839 · doublewordai/control-layer

pjb157 · 2026-09-23T16:40:53Z

Summary

Adds an ingress validation stage that rejects inference requests which are certain to fail, before they are forwarded to a provider or enqueued. Rejections are fast, consistent, and use the error shape of the surface the client called. Clients no longer wait in a queue for a provider error, and doomed requests no longer consume provider or queue capacity.

The same stage runs in two places:

Inference endpoints (/v1/chat/completions, /v1/completions, /v1/responses, /v1/messages, /v1/embeddings): in the inference middleware, after the body is parsed (and after previous_response_id hydration) and before the realtime/flex split, so it covers realtime and queued requests alike. It only validates callers whose key may use the requested model (bearer token or x-api-key), checked against the same routing table and request-class pool onwards authorises against. Unknown models are left to the normal routing 404. Everyone else gets the normal authentication or access error, so a rejection never describes a model the caller cannot use.
Batch input files: every line of an uploaded purpose=batch file. A doomed line fails the upload and the error names the line.

Rules

Rule	Fires when
`model_type_mismatch`	The model's type does not match the endpoint (e.g. an embeddings model on chat completions)
`invalid_service_tier`	`service_tier` is not a recognised value
`invalid_max_tokens`	The max-tokens field is not a positive integer
`max_tokens_exceeds_limit`	The max-tokens field exceeds the model's `max_output_tokens` or `context_window`
`unsupported_modality`	Image input is sent to a model whose capabilities don't include `vision` (audio and file inputs are tracked but not gated yet)
`context_length_exceeded`	The prompt is larger than the model's `context_window`

Guardrails:

Precision over recall. A rule fires only when the request provably cannot succeed. Every rule passes when the model metadata it needs is missing.
Shadow first. The feature is off by default. Once enabled, every rule starts in shadow mode: it records dwctl_request_validation_violations_total{rule,surface,source,model,mode} and forwards the request. Each rule can be switched to enforce on its own.
Lossless. Unknown request fields are never a reason to reject.

Context-length check

The check has two stages, so the common case costs nothing:

Byte count. Byte-level BPE never produces more tokens than bytes, so a prompt that fits the window in bytes passes without a count. For completions/embeddings input lists, each element is its own sequence and the largest one is measured.
Exact count. Larger prompts get an exact chat-templated count from the tokenizer service, under a deadline. If the tokenizer is unavailable or too slow, the request passes. A circuit breaker skips the tokenizer for a cooldown after a timeout or outage.

Context rejections always rest on an exact count, so with exact counting disabled the context rule never rejects.

Model metadata

Rules read a per-alias metadata cache (sync/model_metadata.rs). It mirrors zdr_keys: it loads at startup and refreshes on auth_config_changed, with a periodic fallback. Catalog metadata gains max_output_tokens. context_window and max_output_tokens are now range-checked on the admin API and in model provisioning.

Other changes

Malformed JSON on the inference endpoints now returns a 400 error envelope (code: invalid_json) instead of an empty body, and a failed body read returns a body_read_failed envelope. Both apply whether or not validation is enabled.
The batch upload model-type check now uses the same endpoint-to-model-type mapping as the validation rules.

Configuration

request_validation:
  enabled: false            # DWCTL_REQUEST_VALIDATION__ENABLED
  default_mode: shadow      # off | shadow | enforce
  rules: {}                 # e.g. DWCTL_REQUEST_VALIDATION__RULES__MODEL_TYPE_MISMATCH=enforce
  exact_count_enabled: false
  exact_count_deadline_ms: 2000
  exact_count_max_per_batch_file: 1000

Suggested rollout: enable in shadow mode, check the violations metric for false positives, enable exact counting, then enforce rule by rule.

Test plan

Unit tests for request extraction on every surface, each rule and its boundaries, both error envelopes, exact counting (mocked tokenizer: counted, fallback, unmapped memoization, unavailable, timeout), and the metadata cache (including a real-DB NOTIFY refresh test)
End-to-end tests through the full app: enforce/shadow behaviour, callers without model access are left to normal auth errors, the Anthropic envelope on /v1/messages, the malformed-JSON envelope, flex rejection leaving no queued request, modality, context length, model-type mismatch, fail-open without metadata, and batch upload (a doomed line is rejected with its line number; shadow mode accepts the file)
cargo test -p dwctl: 2668 passed
cargo sqlx prepare --check, cargo fmt

🤖 Generated with Claude Code

Engines return a steady stream of 4xx for requests that were never going to succeed (unknown models, text-only models sent images, prompts over the context window), and queued requests for them occupy queue and retry capacity before failing. Add one validation stage and run it wherever requests enter: the inference middleware (after the body is parsed, before the realtime/flex split) and every line of an uploaded batch input file. A doomed request is rejected before it is forwarded or enqueued. - inference/validation: surface-aware request view, rules (model not found, model type vs endpoint, service_tier, max tokens, modality, context length), OpenAI/Anthropic rejection envelopes, and a two-stage context check: prompts that fit the window in bytes pass for free; larger ones get an exact chat-templated count from the tokenizer service under a deadline. - sync/model_metadata: per-alias metadata cache refreshed on auth_config_changed, mirroring zdr_keys. Missing metadata fails open. - Batch upload reuses the same ValidationStage (shared via AppState) and the same endpoint -> model type mapping. - Rules are off unless request_validation.enabled, default to shadow mode (dwctl_request_validation_violations_total, labelled by source), and are enforced per rule. - Malformed JSON now returns a 400 error envelope instead of an empty body. - Catalog metadata gains max_output_tokens; context_window and max_output_tokens are range-checked on the admin API and in provisioning.

cloudflare-workers-and-pages · 2026-09-23T16:41:41Z

Deploying control-layer with Cloudflare Pages

Latest commit:	`f6c1e29`
Status:	✅ Deploy successful!
Preview URL:	https://5f6d005f.control-layer.pages.dev
Branch Preview URL:	https://feat-ingress-validation.control-layer.pages.dev

View logs

Copilot

Copilot review overview

🟡 Changes recommended

Unresolved authorization, exact-counting, context-bound, and validation-correctness issues remain.

Get a fresh assessment by requesting another Copilot review.

Review effort: Lite
Findings: 3

Open (3)

Exact counting allows unauthenticated tokenizer resource exhaustion · New
Model validation leaks metadata before authorization · New
Unsafe heuristic falsely rejects valid large prompts · New

What changed in this PR

Adds configurable ingress validation for inference and batch requests, backed by cached model metadata and optional exact token counting.

Changes:

Adds validation rules, error envelopes, metrics, and shadow/enforce modes.
Integrates validation into inference middleware and batch uploads.
Adds metadata synchronization and provisioning constraints.
Updates configuration, tests, dashboard types, and SQLx metadata.

File	Summary
`model-provisioning.schema.json`	Adds token metadata constraints.
`dwctl/src/test/utils.rs`	Updates validation test configuration.
`dwctl/src/test/request_validation.rs`	Adds end-to-end validation tests.
`dwctl/src/test/mod.rs`	Registers validation tests.
`dwctl/src/sync/model_metadata.rs`	Implements metadata caching and refresh.
`dwctl/src/sync/mod.rs`	Registers metadata synchronization.
`dwctl/src/model_provisioning.rs`	Validates provisioning metadata.
`dwctl/src/metrics/errors.rs`	Adds metadata-sync metrics.
`dwctl/src/lib.rs`	Wires validation and synchronization.
`dwctl/src/inference/validation/view.rs`	Extracts validation inputs.
`dwctl/src/inference/validation/stage.rs`	Coordinates validation behavior.
`dwctl/src/inference/validation/rules.rs`	Implements validation rules.
`dwctl/src/inference/validation/mod.rs`	Defines validation types and configuration.
`dwctl/src/inference/validation/exact.rs`	Adds tokenizer-backed counting.
`dwctl/src/inference/validation/envelope.rs`	Renders client-facing errors.
`dwctl/src/inference/mod.rs`	Exposes validation modules.
`dwctl/src/inference/middleware.rs`	Integrates ingress validation.
`dwctl/src/db/models/deployments.rs`	Adds model token metadata.
`dwctl/src/db/handlers/api_keys.rs`	Updates test configuration.
`dwctl/src/config.rs`	Adds validation configuration.
`dwctl/src/auth/middleware.rs`	Updates test state.
`dwctl/src/api/handlers/files.rs`	Validates batch input lines.
`dwctl/src/api/handlers/deployments.rs`	Validates catalog metadata.
`dashboard/src/api/control-layer/types.ts`	Updates dashboard metadata types.
`config.yaml`	Documents validation settings.
`.sqlx/query-45070a43e7311f3f5b692b4c5d792425e72cc38ef0b2e1940ff750d492f7fdc2.json`	Adds SQLx query metadata.

Files not reviewed (1)

.sqlx/query-45070a43e7311f3f5b692b4c5d792425e72cc38ef0b2e1940ff750d492f7fdc2.json: Generated file

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

pjb157 · 2026-09-24T14:15:32Z

+    if let (Some(validation), Some(surface)) = (&state.validation, surface)
+        && let Some(rejection) = validation.check(surface, &request_value).await
+    {


Fixed in efeed8c. Validation now runs only for callers whose key may use the requested model, checked against the same in-memory routing table onwards authorises against (ModelAccess in validation/stage.rs). Unauthenticated or unauthorised traffic never reaches the tokenizer and gets the normal auth error downstream. The exact counter also has a circuit breaker now (61c21c6): after a timeout or outage it stops calling the tokenizer for a cooldown, which bounds the cost of repeated near-limit bodies.

pjb157 · 2026-09-24T14:15:33Z

+    // Validate on the bare alias (serving-class suffix already stripped) and
+    // before anything is persisted, forwarded or enqueued.
+    if let (Some(validation), Some(surface)) = (&state.validation, surface)
+        && let Some(rejection) = validation.check(surface, &request_value).await
+    {
+        return rejection;


Fixed in efeed8c. Rules run only when the caller's key may use the model (same check as onwards' own routing and /models visibility). Anyone else gets the normal 401/403/404, so a rejection never describes a model the caller can't use. Covered by callers_without_access_are_not_validated (unit and end-to-end).

pjb157 · 2026-09-24T14:15:35Z

+    // With exact counting on, every rejection is backed by a real count: the
+    // size heuristic below assumes no token spans more than
+    // `reject_min_bytes_per_token` bytes, which is likely but not provable.
+    let estimated = bytes / config.reject_min_bytes_per_token;
+    if !config.exact_count_enabled && estimated > limit {
+        evaluation
+            .violations
+            .push(context_length_violation(view.surface, context_window, estimated.floor() as u64));


Agreed. Removed in efeed8c along with reject_min_bytes_per_token. Stage 1 now only passes prompts (bytes <= context_window is a proof, since byte-level BPE gives tokens <= bytes). Anything larger is rejected only on an exact tokenizer count, so with exact counting off the context rule never rejects.

cubic-dev-ai

6 issues found across 26 files

Prompt for AI agents (unresolved issues)


Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.


<file name="dwctl/src/lib.rs">

<violation number="1" location="dwctl/src/lib.rs:3524">
P1: This background task can take the whole server down when its initial listener connection fails, contradicting the fail-open behavior promised for unavailable model metadata. Keep the sync task alive with retry/backoff around listener setup (as well as reload failures) so validation continues using the last or empty cache.</violation>
</file>

<file name="dwctl/src/config.rs">

<violation number="1" location="dwctl/src/config.rs:241">
P2: `request_validation` accepts `reject_min_bytes_per_token: 0` without startup validation, so enabling the feature can reject valid near-limit prompts because the context rule divides by zero. Validate this value as finite and strictly positive before enabling request validation.</violation>
</file>

<file name="dwctl/src/api/handlers/files.rs">

<violation number="1" location="dwctl/src/api/handlers/files.rs:960">
P2: The new validation error reports the accepted-request ordinal instead of the JSONL source line. Track the physical line number separately so blank lines do not misidentify which batch entry failed.</violation>
</file>

<file name="dwctl/src/inference/validation/stage.rs">

<violation number="1" location="dwctl/src/inference/validation/stage.rs:64">
P1: Responses continuations are validated before their prior response is inlined, so context checks and exact counts omit prior-turn tokens and can forward or enqueue a request that exceeds the model window after hydration. Run validation after hydration or validate the hydrated body.</violation>
</file>

<file name="dwctl/src/inference/validation/view.rs">

<violation number="1" location="dwctl/src/inference/validation/view.rs:53">
P1: `service_tier: "standard_only"` is valid on the Anthropic Messages surface, but the global tier rule rejects it after this extraction. Keep the allowlist surface-aware or skip this rule for Messages.</violation>
</file>

<file name="dwctl/src/test/request_validation.rs">

<violation number="1" location="dwctl/src/test/request_validation.rs:651">
P3: The shadow-mode tests never verify that the would-be rejection was actually recorded, so a wiring bug that silently skipped validation in shadow mode (or never recorded the violation) would still pass. `ValidationStage.enforced_violation` records every violation before forwarding (rules::record + tracing), so assert on that observability (a violation counter or log record) in addition to the forwarded response.</violation>
</file>

_{Tip: instead of fixing issues one by one fix them all with cubic

Re-trigger cubic}

cubic-dev-ai · 2026-09-23T16:53:50Z

+        let metadata_cache = cache.clone();
+        let metadata_shutdown = shutdown_token.clone();
+        let metadata_fallback = config.background_services.onwards_sync.fallback_interval_milliseconds;
+        background_tasks.spawn("model-metadata-sync", async move {


P1: This background task can take the whole server down when its initial listener connection fails, contradicting the fail-open behavior promised for unavailable model metadata. Keep the sync task alive with retry/backoff around listener setup (as well as reload failures) so validation continues using the last or empty cache.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At dwctl/src/lib.rs, line 3524: <comment>This background task can take the whole server down when its initial listener connection fails, contradicting the fail-open behavior promised for unavailable model metadata. Keep the sync task alive with retry/backoff around listener setup (as well as reload failures) so validation continues using the last or empty cache.</comment> <file context> @@ -3481,6 +3499,44 @@ async fn setup_background_services(input: BackgroundServicesInput) -> anyhow::Re + let metadata_cache = cache.clone(); + let metadata_shutdown = shutdown_token.clone(); + let metadata_fallback = config.background_services.onwards_sync.fallback_interval_milliseconds; + background_tasks.spawn("model-metadata-sync", async move { + crate::sync::model_metadata::run( + metadata_pool, </file context>

Fixed in d9a8549. run() no longer returns on DB errors: listener setup is retried with capped exponential backoff (1s→30s, honouring shutdown) and only returns on shutdown. Tested with a pool whose first connect fails.

cubic-dev-ai · 2026-09-23T16:53:50Z

+    /// Run every rule on a parsed body, record all violations, and return the
+    /// first enforced one. Callers render it in their own error shape.
+    pub async fn enforced_violation(&self, surface: Surface, body: &Value, source: Source) -> Option<Violation> {
+        let view = view::extract(surface, body);


P1: Responses continuations are validated before their prior response is inlined, so context checks and exact counts omit prior-turn tokens and can forward or enqueue a request that exceeds the model window after hydration. Run validation after hydration or validate the hydrated body.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At dwctl/src/inference/validation/stage.rs, line 64: <comment>Responses continuations are validated before their prior response is inlined, so context checks and exact counts omit prior-turn tokens and can forward or enqueue a request that exceeds the model window after hydration. Run validation after hydration or validate the hydrated body.</comment> <file context> @@ -0,0 +1,196 @@ + /// Run every rule on a parsed body, record all violations, and return the + /// first enforced one. Callers render it in their own error shape. + pub async fn enforced_violation(&self, surface: Surface, body: &Value, source: Source) -> Option<Violation> { + let view = view::extract(surface, body); + let lookup = match view.model.as_deref() { + Some(alias) => self.models.lookup(alias), </file context>

Fixed in efeed8c. The validation call moved after previous_response_id hydration (still before the realtime/flex split), so prior turns count toward the context window.

cubic-dev-ai · 2026-09-23T16:53:50Z

+        // `model` is the bare alias string; non-string values are ignored.
+        model: body.get("model").and_then(Value::as_str).map(str::to_owned),
+        // Raw, so a rule can report a non-string tier rather than silently drop it.
+        service_tier: body.get("service_tier").cloned(),


P1: service_tier: "standard_only" is valid on the Anthropic Messages surface, but the global tier rule rejects it after this extraction. Keep the allowlist surface-aware or skip this rule for Messages.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At dwctl/src/inference/validation/view.rs, line 53: <comment>`service_tier: "standard_only"` is valid on the Anthropic Messages surface, but the global tier rule rejects it after this extraction. Keep the allowlist surface-aware or skip this rule for Messages.</comment> <file context> @@ -0,0 +1,709 @@ + // `model` is the bare alias string; non-string values are ignored. + model: body.get("model").and_then(Value::as_str).map(str::to_owned), + // Raw, so a rule can report a non-string tier rather than silently drop it. + service_tier: body.get("service_tier").cloned(), + ..RequestView::default() + }; </file context>

Fixed in efeed8c. The accepted service_tier values are surface-aware; /v1/messages also accepts standard_only.

cubic-dev-ai · 2026-09-23T16:53:50Z

+    let too_long = "a".repeat(5000);
+    let response = fixture.upload_batch(&batch_jsonl(&alias, &[&too_long])).await;
+
+    assert_eq!(response.status_code(), StatusCode::CREATED, "{}", response.text());


P3: The shadow-mode tests never verify that the would-be rejection was actually recorded, so a wiring bug that silently skipped validation in shadow mode (or never recorded the violation) would still pass. ValidationStage.enforced_violation records every violation before forwarding (rules::record + tracing), so assert on that observability (a violation counter or log record) in addition to the forwarded response.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At dwctl/src/test/request_validation.rs, line 651: <comment>The shadow-mode tests never verify that the would-be rejection was actually recorded, so a wiring bug that silently skipped validation in shadow mode (or never recorded the violation) would still pass. `ValidationStage.enforced_violation` records every violation before forwarding (rules::record + tracing), so assert on that observability (a violation counter or log record) in addition to the forwarded response.</comment> <file context> @@ -0,0 +1,652 @@ + let too_long = "a".repeat(5000); + let response = fixture.upload_batch(&batch_jsonl(&alias, &[&too_long])).await; + + assert_eq!(response.status_code(), StatusCode::CREATED, "{}", response.text()); +} </file context>

Partly addressed: the shadow end-to-end test now uses a known model that breaks a rule (so a rule definitely fires) and asserts the request is forwarded and unmarked. Asserting the counter would need a process-global metrics recorder in the test harness, which other tests share, so I've left that to the unit level (rules::record / describe_metrics).

pg_stat_activity shows the LISTEN statement from the moment it starts; a NOTIFY committed before the LISTEN commits is never delivered, so the barrier must also require the connection to be idle.

cubic-dev-ai

All reported issues were addressed across 1 file (changes from recent commits).

_{Reply with feedback, questions, or to request a fix.

Fix all with cubic | Re-trigger cubic}

- Validate only callers whose key may use the requested model (the same routing table onwards authorises against). Everyone else gets the normal authentication/access error, so rejections never describe a model the caller cannot use and unauthenticated traffic never reaches the tokenizer. - Validate after previous_response_id hydration so prior turns count toward the context window. - Drop the size-only context rejection: no byte ratio proves a prompt is over the window, so context rejections always rest on an exact count (removes reject_min_bytes_per_token). - Completions/embeddings inputs are independent sequences: bound the largest element, not the sum. - Accept Anthropic's `standard_only` service tier on /v1/messages. - max tokens: strict u32 integers on Messages/Responses, integral floats bounded by u32::MAX elsewhere; a null max_completion_tokens falls back to max_tokens. - Responses: only function tools and URL images count toward the view, matching what the translator forwards. - Tests: poll with a short sleep; document the `source` metric label.

Retry listener setup with capped backoff instead of ending the task, reload once after every (re)subscribe so changes committed before LISTEN are not missed, and use saturating_sub for the debounce deadline.

…cuit breaker Completions prompt and embeddings input lists are independent sequences, so the exact count is the largest element, not the sum. After a timeout or outage the exact counter stops calling tokenizer-svc for a cooldown, so an outage cannot cost one deadline per request or per batch line.

…rdai/control-layer into feat/ingress-validation

…into feat/ingress-validation # Conflicts: # dwctl/src/inference/middleware.rs # dwctl/src/lib.rs # dwctl/src/sync/mod.rs # dwctl/src/test/mod.rs

- Drop the model_not_found rule: validation only runs once routing has accepted the alias, so an unknown alias in the metadata cache can only mean that cache is stale, and routing already answers genuinely unknown models. - Authorise validation against the pool onwards resolves for the request class, and accept Anthropic's x-api-key header for the gate. - Stop gating audio input: the catalog has no audio capability vocabulary, so its absence proves nothing. - Only open the exact-count circuit breaker on real outages (transport, 5xx, 408, 429); other tokenizer 4xx responses are about the request. - Cap exact counts per batch file upload (request_validation.exact_count_max_per_batch_file, default 1000). - Tests for x-api-key callers and for disabled validation forwarding a violating request; doc and config.yaml placement fixes.

cubic-dev-ai

1 issue found across 10 files (changes from recent commits).

Prompt for AI agents (unresolved issues)


Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.


<file name="dwctl/src/inference/validation/stage.rs">

<violation number="1" location="dwctl/src/inference/validation/stage.rs:152">
P3: The per-upload budget is consumed in shadow mode too (exact counts still run to record violations there), so a file with >1000 near-limit lines stops producing shadow violation metrics past the budget — silently under-reporting exactly the signal the shadow rollout phase uses to judge enforcement impact. If tokenizer load capping is the goal, consider not debiting the budget when the context rule is only in shadow; if the cap must apply regardless, document the shadow-mode measurement bias.</violation>
</file>

_{Tip: Review your code locally with the cubic CLI to iterate faster.

Re-trigger cubic}

cubic-dev-ai · 2026-09-25T14:55:03Z

+        let decided = rules::first_enforced(&evaluation.violations, &self.config).is_some();
+        if let (Some(needed), Some(exact), Some(alias), false) = (&evaluation.exact_count, &self.exact, view.model.as_deref(), decided)
+            && self.config.mode(RuleId::ContextLengthExceeded) != RuleMode::Off
+            && budget.is_none_or(ExactCountBudget::try_take)


P3: The per-upload budget is consumed in shadow mode too (exact counts still run to record violations there), so a file with >1000 near-limit lines stops producing shadow violation metrics past the budget — silently under-reporting exactly the signal the shadow rollout phase uses to judge enforcement impact. If tokenizer load capping is the goal, consider not debiting the budget when the context rule is only in shadow; if the cap must apply regardless, document the shadow-mode measurement bias.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At dwctl/src/inference/validation/stage.rs, line 152: <comment>The per-upload budget is consumed in shadow mode too (exact counts still run to record violations there), so a file with >1000 near-limit lines stops producing shadow violation metrics past the budget — silently under-reporting exactly the signal the shadow rollout phase uses to judge enforcement impact. If tokenizer load capping is the goal, consider not debiting the budget when the context rule is only in shadow; if the cap must apply regardless, document the shadow-mode measurement bias.</comment> <file context> @@ -114,6 +149,7 @@ impl ValidationStage { let decided = rules::first_enforced(&evaluation.violations, &self.config).is_some(); if let (Some(needed), Some(exact), Some(alias), false) = (&evaluation.exact_count, &self.exact, view.model.as_deref(), decided) && self.config.mode(RuleId::ContextLengthExceeded) != RuleMode::Off + && budget.is_none_or(ExactCountBudget::try_take) && let ExactOutcome::Counted(tokens) = exact.prompt_tokens(alias, surface, body).await && let Some(violation) = rules::exact_count_violation(surface, alias, tokens, needed) </file context>

Keeping the budget in shadow mode on purpose: its job is to cap tokenizer load per upload, and shadow mode does the same tokenizer work. Documented the measurement bias in f6c1e29 (ValidationConfig::exact_count_max_per_batch_file and config.yaml): past the budget, shadow metrics under-count context violations for that file.

…h budget bias in shadow mode

Copilot AI lite review requested due to automatic review settings September 23, 2026 16:40

Copilot started reviewing on behalf of pjb157 September 23, 2026 16:41 View session

Copilot AI reviewed Sep 23, 2026

View reviewed changes

cubic-dev-ai Bot reviewed Sep 23, 2026

View reviewed changes

test(validation): wait for LISTEN to complete before notifying

8cc85c0

pg_stat_activity shows the LISTEN statement from the moment it starts; a NOTIFY committed before the LISTEN commits is never delivered, so the barrier must also require the connection to be idle.

cubic-dev-ai Bot reviewed Sep 23, 2026

View reviewed changes

Comment thread dwctl/src/sync/model_metadata.rs Outdated

pjb157 and others added 9 commits September 24, 2026 13:48

Merge branch 'main' into feat/ingress-validation

30d9dcf

fix(validation): keep model metadata sync alive through DB errors

d9a8549

Retry listener setup with capped backoff instead of ending the task, reload once after every (re)subscribe so changes committed before LISTEN are not missed, and use saturating_sub for the debounce deadline.

test(validation): drop unused import

c71c35a

Merge branch 'feat/ingress-validation' of https://github.com/doublewo…

1d29532

…rdai/control-layer into feat/ingress-validation

Merge branch 'main' into feat/ingress-validation

753a21b

Merge branch 'main' of https://github.com/doublewordai/control-layer …

bc43f9f

…into feat/ingress-validation # Conflicts: # dwctl/src/inference/middleware.rs # dwctl/src/lib.rs # dwctl/src/sync/mod.rs # dwctl/src/test/mod.rs

cubic-dev-ai Bot reviewed Sep 25, 2026

View reviewed changes

fix(validation): accept x-api-key only on /v1/messages; document batc…

f6c1e29

…h budget bias in shadow mode

Conversation

pjb157 commented Sep 23, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Rules

Context-length check

Model metadata

Other changes

Configuration

Test plan

Uh oh!

cloudflare-workers-and-pages Bot commented Sep 23, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Deploying control-layer with Cloudflare Pages

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Copilot review overview

🟡 Changes recommended

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

cubic-dev-ai Bot left a comment • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

cubic-dev-ai Bot Sep 23, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

cubic-dev-ai Bot Sep 23, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

cubic-dev-ai Bot Sep 23, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

cubic-dev-ai Bot Sep 23, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

cubic-dev-ai Bot left a comment • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

cubic-dev-ai Bot left a comment • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

cubic-dev-ai Bot Sep 25, 2026

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

This branch has not been deployed

Reviewers

Assignees

Labels

Projects

pjb157 commented Sep 23, 2026 •

edited

Loading

cloudflare-workers-and-pages Bot commented Sep 23, 2026 •

edited

Loading

cubic-dev-ai Bot left a comment •

edited

Loading

cubic-dev-ai Bot Sep 23, 2026 •

edited

Loading

cubic-dev-ai Bot Sep 23, 2026 •

edited

Loading

cubic-dev-ai Bot Sep 23, 2026 •

edited

Loading

cubic-dev-ai Bot Sep 23, 2026 •

edited

Loading

cubic-dev-ai Bot left a comment •

edited

Loading

cubic-dev-ai Bot left a comment •

edited

Loading