@vostride/agent-qa-model 0.2.2 → 0.2.4

package/dist/index.d.cts

@@ -51,54 +51,18 @@ type LogEntry = {
     level?: string;
     message: string;
 };
-/**
- * Compact a log-shaped burst into a deterministic, bounded string (D-13/D-15).
- * Groups typed `{source, level, message}` entries by `source::level::normalized`,
- * counts each class, and renders it with the first ≤5 raw samples seen (monotonic:
- * once a class has 5 samples, growing the burst never changes which 5 are shown;
- * the onset occurrence is always retained). Classes render in sorted-key order.
- */
 declare function compactLogs(entries: LogEntry[]): string;
-/**
- * Bound a single structured value for prompt inclusion (D-14). Safe-stringifies the
- * value verbatim (variable tokens preserved — the planner may need them exactly) and,
- * only if it exceeds `BLOB_CEILING`, slices and appends the visible truncation marker.
- * `undefined` in ⇒ `undefined` out. Never throws on cyclic input (V5).
- */
 declare function guardBlob(value: unknown): string | undefined;
-/**
- * The composed pre-prompt guard for one tool-output value (CTXB-06): a burst of
- * log-shaped entries (console `{level, text}` / network `{url, method, status}` /
- * generic `{source?, level?, message}`) routes through {@link compactLogs};
- * anything else through {@link guardBlob}. Compacted output is still bounded by
- * the blob ceiling so a burst of all-distinct classes cannot flood the prompt.
- */
 declare function guardToolOutput(value: unknown): string | undefined;
 
 declare function parseRealWindow(error: unknown): number | undefined;
 declare function isContextOverflow(error: unknown): boolean;
 
-/**
- * Caller-supplied degrade controls (D-07/D-09). Policy-free: no default/safety margin is
- * baked here — Phase 355 supplies the concrete budget and pin predicate.
- *
- * - `budget` — the token ceiling the degraded prompt must fit under.
- * - `pin` — a caller pin predicate; a message matching it is never evicted or truncated.
- *   The substrate stays tier-ignorant (355 pins its tiers explicitly).
- * - `estimate` — the estimator the damage loop re-runs after each eviction. Optional so
- *   `degrade` stays a standalone-testable pure function; the Plan-03 factory passes its own.
- */
 interface DegradeOptions {
     budget: number;
     pin?: (message: ModelMessage, index: number) => boolean;
     estimate?: (messages: ModelMessage[]) => number;
 }
-/**
- * Discriminated degrade result (D-09). `ok:true` carries the fitted messages; `ok:false`
- * is the structured cannot-fit outcome returned when system + final + pinned content alone
- * exceeds the budget — degrade never throws and never violates a pin, returning the
- * best-effort hygiene-reduced messages instead.
- */
 type DegradeOutcome = {
     ok: true;
     messages: ModelMessage[];
@@ -107,10 +71,6 @@ type DegradeOutcome = {
     reason: 'pinned-content-exceeds-budget';
     messages: ModelMessage[];
 };
-/**
- * The instance shape returned by the Plan-03 factory and consumed through the runner's
- * provider-neutral port (D-05). Prompt tokens only; the substrate is policy-free.
- */
 interface TokenAccounting {
     estimate(messages: ModelMessage[]): number;
     recordUsage(usage: LanguageModelUsage, messages: ModelMessage[]): void;
@@ -120,13 +80,6 @@ interface TokenAccounting {
     };
     degrade(messages: ModelMessage[], options: DegradeOptions): DegradeOutcome;
 }
-/**
- * Factory construction options (Plan 03).
- *
- * - `getCapabilities` — the D-04 lazy ceilings thunk (composes v46 `resolveCapabilities`),
- *   read on every estimate/degrade so a mid-run overflow-learned window shrink applies next turn.
- * - `perImageTokens` — the D-12 injectable flat per-image constant (test/chat-layer tuning).
- */
 interface TokenAccountingOptions {
     getCapabilities: () => ModelCapabilities;
     perImageTokens?: number;
@@ -261,7 +214,6 @@ interface VisionGate {
     captureMode: 'full' | 'textOnly';
     promptNote: string;
 }
-/** 439 D-19: the configured `vision` when set, otherwise what the capability table says. */
 declare function visionFromConfig(configured: boolean | undefined, capabilities: {
     vision: boolean;
 }): boolean;

package/dist/index.d.ts

@@ -51,54 +51,18 @@ type LogEntry = {
     level?: string;
     message: string;
 };
-/**
- * Compact a log-shaped burst into a deterministic, bounded string (D-13/D-15).
- * Groups typed `{source, level, message}` entries by `source::level::normalized`,
- * counts each class, and renders it with the first ≤5 raw samples seen (monotonic:
- * once a class has 5 samples, growing the burst never changes which 5 are shown;
- * the onset occurrence is always retained). Classes render in sorted-key order.
- */
 declare function compactLogs(entries: LogEntry[]): string;
-/**
- * Bound a single structured value for prompt inclusion (D-14). Safe-stringifies the
- * value verbatim (variable tokens preserved — the planner may need them exactly) and,
- * only if it exceeds `BLOB_CEILING`, slices and appends the visible truncation marker.
- * `undefined` in ⇒ `undefined` out. Never throws on cyclic input (V5).
- */
 declare function guardBlob(value: unknown): string | undefined;
-/**
- * The composed pre-prompt guard for one tool-output value (CTXB-06): a burst of
- * log-shaped entries (console `{level, text}` / network `{url, method, status}` /
- * generic `{source?, level?, message}`) routes through {@link compactLogs};
- * anything else through {@link guardBlob}. Compacted output is still bounded by
- * the blob ceiling so a burst of all-distinct classes cannot flood the prompt.
- */
 declare function guardToolOutput(value: unknown): string | undefined;
 
 declare function parseRealWindow(error: unknown): number | undefined;
 declare function isContextOverflow(error: unknown): boolean;
 
-/**
- * Caller-supplied degrade controls (D-07/D-09). Policy-free: no default/safety margin is
- * baked here — Phase 355 supplies the concrete budget and pin predicate.
- *
- * - `budget` — the token ceiling the degraded prompt must fit under.
- * - `pin` — a caller pin predicate; a message matching it is never evicted or truncated.
- *   The substrate stays tier-ignorant (355 pins its tiers explicitly).
- * - `estimate` — the estimator the damage loop re-runs after each eviction. Optional so
- *   `degrade` stays a standalone-testable pure function; the Plan-03 factory passes its own.
- */
 interface DegradeOptions {
     budget: number;
     pin?: (message: ModelMessage, index: number) => boolean;
     estimate?: (messages: ModelMessage[]) => number;
 }
-/**
- * Discriminated degrade result (D-09). `ok:true` carries the fitted messages; `ok:false`
- * is the structured cannot-fit outcome returned when system + final + pinned content alone
- * exceeds the budget — degrade never throws and never violates a pin, returning the
- * best-effort hygiene-reduced messages instead.
- */
 type DegradeOutcome = {
     ok: true;
     messages: ModelMessage[];
@@ -107,10 +71,6 @@ type DegradeOutcome = {
     reason: 'pinned-content-exceeds-budget';
     messages: ModelMessage[];
 };
-/**
- * The instance shape returned by the Plan-03 factory and consumed through the runner's
- * provider-neutral port (D-05). Prompt tokens only; the substrate is policy-free.
- */
 interface TokenAccounting {
     estimate(messages: ModelMessage[]): number;
     recordUsage(usage: LanguageModelUsage, messages: ModelMessage[]): void;
@@ -120,13 +80,6 @@ interface TokenAccounting {
     };
     degrade(messages: ModelMessage[], options: DegradeOptions): DegradeOutcome;
 }
-/**
- * Factory construction options (Plan 03).
- *
- * - `getCapabilities` — the D-04 lazy ceilings thunk (composes v46 `resolveCapabilities`),
- *   read on every estimate/degrade so a mid-run overflow-learned window shrink applies next turn.
- * - `perImageTokens` — the D-12 injectable flat per-image constant (test/chat-layer tuning).
- */
 interface TokenAccountingOptions {
     getCapabilities: () => ModelCapabilities;
     perImageTokens?: number;
@@ -261,7 +214,6 @@ interface VisionGate {
     captureMode: 'full' | 'textOnly';
     promptNote: string;
 }
-/** 439 D-19: the configured `vision` when set, otherwise what the capability table says. */
 declare function visionFromConfig(configured: boolean | undefined, capabilities: {
     vision: boolean;
 }): boolean;

package/dist/index.js

@@ -720,7 +720,7 @@ var LOCAL_OPENAI_WIRE = [
   {
     id: "deepseek-chat",
     label: "DeepSeek Chat",
-    notes: "text-only model; use accessibility tree and DOM context."
+    notes: "Text-only model; use the accessibility tree."
   },
   { id: "deepseek-r1", label: "DeepSeek R1", notes: "Text-only reasoning model." },
   { id: "qwen2.5-vl", label: "Qwen2.5 VL" },
@@ -1256,10 +1256,7 @@ function seedBuiltins() {
     subscription: true,
     modelAdapter: "openai-responses",
     defaultModel: "gpt-6.1-sol",
-    // RUNX-01: the codex/ChatGPT-subscription Responses backend rejects the output cap
-    // (`max_output_tokens`) with HTTP 400. Strip it at egress for this provider ONLY —
-    // standard OpenAI-compatible `.chat()` endpoints and Anthropic support the cap.
-    // 2026-10-20 (446-11, D-18): it also answers 400 "Unsupported parameter: temperature".
+    // The subscription backend rejects `max_output_tokens` and `temperature` with HTTP 400.
     unsupportedParams: ["maxOutputTokens", "temperature"]
   });
   registerProvider({
@@ -1415,9 +1412,9 @@ var MIDDLEWARE = [
   toolSchemaMiddleware,
   deriveExamplesMiddleware(),
   addToolInputExamplesMiddleware(),
-  // RUNX-04: flag-gated debug logging (AGENT_QA_DEBUG_LLM); catches the raw APICallError first.
+  // Flag-gated by `AGENT_QA_DEBUG_LLM`; first, so it catches the raw `APICallError`.
   debugLoggingMiddleware(),
-  // RUNX-01: strip empty-content messages just before the wire.
+  // Strips empty-content messages just before the wire.
   dropEmptyMessagesMiddleware()
 ];
 function createModel(profile, auth, mask, overrides = {}, observation = {}) {
@@ -1566,7 +1563,7 @@ function resolveVisionConfig(servicesVision) {
 // src/vision/resolve-gate.ts
 var TEXT_ONLY_PROMPT_NOTE = [
   "Visual screenshot is unavailable for this step.",
-  "Rely on the accessibility tree and DOM context for element identification, including roles, names, values, refs, and URLs.",
+  "Rely on the accessibility tree for element identification, including roles, names, values, refs, and URLs.",
   "Verify functional state from the available structured context rather than purely visual appearance."
 ].join(" ");
 function visionFromConfig(configured, capabilities) {