@ggui-ai/negotiator 0.27.0 → 0.28.0

package/dist/llm-rerank.d.ts

@@ -67,7 +67,7 @@ export interface RerankQuery {
     readonly intent: string;
     readonly contractSummary: string;
 }
-declare const RERANK_SYSTEM_PROMPT = "You match user UI requests against previously-generated UI blueprints. Each blueprint was produced for a past request and stored. Decide whether any candidate belongs to the SAME FAMILY as the current request \u2014 i.e. it would serve as a reasonable starting point that the requester can refine, not a pixel-exact replica.\n\nMATCH means the candidate is the same intended user task AND the same broad UI shape (component types, layout pattern). A candidate still MATCHES when the current request adds or omits fields, slots, or actions relative to the cached blueprint \u2014 a superset, a subset, or an overlapping wire surface all still match. Added or omitted fields/slots/actions DO NOT block a match and are NOT yours to judge: those wire-surface deltas are reconciled and reported to the agent separately, after you decide. Judge similarity of task and shape, never coverage of fields.\n\nNO-MATCH means the candidate is a fundamentally different thing \u2014 a different intended task (haiku composer vs tweet draft, login vs signup), a different UI shape (form vs list vs dashboard), OR a conflicting load-bearing fixed VALUE baked into the blueprint (calendar pinned to Jan vs a request for Mar \u2014 same contract shape, but the fixed value conflicts). Reserve NO-MATCH for these; do not decline a candidate merely because its fields, slots, or actions differ from the request.\n\nVisual style differences alone (minimal vs ornate, dense vs spacious) DO NOT block a match \u2014 the user can refine those after they get a working UI.\n\nOutput exactly ONE tool call with your decision. confidence is a number in [0, 1] measuring how strongly the candidate is in the same task-and-shape family as the request (not how completely its fields cover the request). matchId is a string from the candidate ids, or null when no candidate matches. reason is a short sentence the operator can use to debug.";
+declare const RERANK_SYSTEM_PROMPT = "You match user UI requests against previously-generated UI blueprints. Each blueprint was produced for a past request and stored. Decide whether any candidate belongs to the SAME FAMILY as the current request \u2014 i.e. it would serve as a reasonable starting point that the requester can refine, not a pixel-exact replica.\n\nMATCH means the candidate is the same intended user task AND the same broad UI shape (component types, layout pattern). A candidate still MATCHES when the current request adds or omits fields, slots, or actions relative to the cached blueprint \u2014 a superset, a subset, or an overlapping wire surface all still match. Added or omitted fields/slots/actions DO NOT block a match and are NOT yours to judge: those wire-surface deltas are reconciled and reported to the agent separately, after you decide. Judge similarity of task and shape, never coverage of fields.\n\nUI shape includes the interaction model. A surface that observes input and filters data it already has (context only, no action: it narrows what is on the screen as you type and has nothing to press) is a different shape from one that submits an action to fetch or act (a button, a pending state, results that arrive). Do not match across that line, even when the fields look alike.\n\nNO-MATCH means the candidate is a fundamentally different thing \u2014 a different intended task (haiku composer vs tweet draft, login vs signup), a different UI shape (form vs list vs dashboard), OR a conflicting load-bearing fixed VALUE baked into the blueprint (calendar pinned to Jan vs a request for Mar \u2014 same contract shape, but the fixed value conflicts). Reserve NO-MATCH for these; do not decline a candidate merely because its fields, slots, or actions differ from the request.\n\nVisual style differences alone (minimal vs ornate, dense vs spacious) DO NOT block a match \u2014 the user can refine those after they get a working UI.\n\nOutput exactly ONE tool call with your decision. confidence is a number in [0, 1] measuring how strongly the candidate is in the same task-and-shape family as the request (not how completely its fields cover the request). matchId is a string from the candidate ids, or null when no candidate matches. reason is a short sentence the operator can use to debug.";
 declare const RERANK_TOOL: ToolSchema;
 /**
  * Run the rerank judge against a query + candidate list.

package/dist/llm-rerank.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"llm-rerank.d.ts","sourceRoot":"","sources":["../src/llm-rerank.ts"],"names":[],"mappings":"AAeA,OAAO,KAAK,EAAE,SAAS,EAAE,UAAU,EAAE,UAAU,EAAE,MAAM,iBAAiB,CAAC;AAEzE;;;;;;GAMG;AACH,MAAM,WAAW,eAAe;IAC9B,+DAA+D;IAC/D,QAAQ,CAAC,EAAE,EAAE,MAAM,CAAC;IACpB,gEAAgE;IAChE,QAAQ,CAAC,YAAY,EAAE,MAAM,CAAC;IAC9B;;;;;OAKG;IACH,QAAQ,CAAC,qBAAqB,EAAE,MAAM,CAAC;IACvC;;;;;OAKG;IACH,QAAQ,CAAC,MAAM,CAAC,EAAE,MAAM,CAAC;CAC1B;AAED,0CAA0C;AAC1C,MAAM,WAAW,cAAc;IAC7B;;;OAGG;IACH,QAAQ,CAAC,OAAO,EAAE,MAAM,GAAG,IAAI,CAAC;IAChC;;;;;;;;OAQG;IACH,QAAQ,CAAC,UAAU,EAAE,MAAM,CAAC;IAC5B;;;;;OAKG;IACH,QAAQ,CAAC,MAAM,CAAC,EAAE,MAAM,CAAC;IACzB,0CAA0C;IAC1C,QAAQ,CAAC,SAAS,EAAE,MAAM,CAAC;IAC3B;;;;;;;OAOG;IACH,QAAQ,CAAC,SAAS,CAAC,EAAE,UAAU,CAAC;CACjC;AAED,8DAA8D;AAC9D,MAAM,WAAW,WAAW;IAC1B,QAAQ,CAAC,MAAM,EAAE,MAAM,CAAC;IACxB,QAAQ,CAAC,eAAe,EAAE,MAAM,CAAC;CAClC;AAED,QAAA,MAAM,oBAAoB,03DAQ6U,CAAC;AAExW,QAAA,MAAM,WAAW,EAAE,UA4BlB,CAAC;AAkFF;;;;;;;;;;;GAWG;AACH,wBAAsB,gBAAgB,CACpC,IAAI,EAAE;IAAE,QAAQ,CAAC,GAAG,EAAE,SAAS,CAAA;CAAE,EACjC,KAAK,EAAE,WAAW,EAClB,UAAU,EAAE,SAAS,eAAe,EAAE,GACrC,OAAO,CAAC,cAAc,CAAC,CAyDzB;AAGD;;;;;;;GAOG;AACH,MAAM,MAAM,WAAW,GAAG,CACxB,KAAK,EAAE,WAAW,EAClB,UAAU,EAAE,SAAS,eAAe,EAAE,KACnC,OAAO,CAAC,cAAc,CAAC,CAAC;AAE7B;;;;GAIG;AACH,wBAAgB,cAAc,CAAC,GAAG,EAAE,SAAS,GAAG,WAAW,CAE1D;AAED,OAAO,EAAE,oBAAoB,EAAE,WAAW,EAAE,CAAC"}
+{"version":3,"file":"llm-rerank.d.ts","sourceRoot":"","sources":["../src/llm-rerank.ts"],"names":[],"mappings":"AAeA,OAAO,KAAK,EAAE,SAAS,EAAE,UAAU,EAAE,UAAU,EAAE,MAAM,iBAAiB,CAAC;AAEzE;;;;;;GAMG;AACH,MAAM,WAAW,eAAe;IAC9B,+DAA+D;IAC/D,QAAQ,CAAC,EAAE,EAAE,MAAM,CAAC;IACpB,gEAAgE;IAChE,QAAQ,CAAC,YAAY,EAAE,MAAM,CAAC;IAC9B;;;;;OAKG;IACH,QAAQ,CAAC,qBAAqB,EAAE,MAAM,CAAC;IACvC;;;;;OAKG;IACH,QAAQ,CAAC,MAAM,CAAC,EAAE,MAAM,CAAC;CAC1B;AAED,0CAA0C;AAC1C,MAAM,WAAW,cAAc;IAC7B;;;OAGG;IACH,QAAQ,CAAC,OAAO,EAAE,MAAM,GAAG,IAAI,CAAC;IAChC;;;;;;;;OAQG;IACH,QAAQ,CAAC,UAAU,EAAE,MAAM,CAAC;IAC5B;;;;;OAKG;IACH,QAAQ,CAAC,MAAM,CAAC,EAAE,MAAM,CAAC;IACzB,0CAA0C;IAC1C,QAAQ,CAAC,SAAS,EAAE,MAAM,CAAC;IAC3B;;;;;;;OAOG;IACH,QAAQ,CAAC,SAAS,CAAC,EAAE,UAAU,CAAC;CACjC;AAED,8DAA8D;AAC9D,MAAM,WAAW,WAAW;IAC1B,QAAQ,CAAC,MAAM,EAAE,MAAM,CAAC;IACxB,QAAQ,CAAC,eAAe,EAAE,MAAM,CAAC;CAClC;AAED,QAAA,MAAM,oBAAoB,4vEAU6U,CAAC;AAExW,QAAA,MAAM,WAAW,EAAE,UA4BlB,CAAC;AAkFF;;;;;;;;;;;GAWG;AACH,wBAAsB,gBAAgB,CACpC,IAAI,EAAE;IAAE,QAAQ,CAAC,GAAG,EAAE,SAAS,CAAA;CAAE,EACjC,KAAK,EAAE,WAAW,EAClB,UAAU,EAAE,SAAS,eAAe,EAAE,GACrC,OAAO,CAAC,cAAc,CAAC,CAyDzB;AAGD;;;;;;;GAOG;AACH,MAAM,MAAM,WAAW,GAAG,CACxB,KAAK,EAAE,WAAW,EAClB,UAAU,EAAE,SAAS,eAAe,EAAE,KACnC,OAAO,CAAC,cAAc,CAAC,CAAC;AAE7B;;;;GAIG;AACH,wBAAgB,cAAc,CAAC,GAAG,EAAE,SAAS,GAAG,WAAW,CAE1D;AAED,OAAO,EAAE,oBAAoB,EAAE,WAAW,EAAE,CAAC"}

package/dist/llm-rerank.js

@@ -17,6 +17,8 @@ const RERANK_SYSTEM_PROMPT = `You match user UI requests against previously-gene
 
 MATCH means the candidate is the same intended user task AND the same broad UI shape (component types, layout pattern). A candidate still MATCHES when the current request adds or omits fields, slots, or actions relative to the cached blueprint — a superset, a subset, or an overlapping wire surface all still match. Added or omitted fields/slots/actions DO NOT block a match and are NOT yours to judge: those wire-surface deltas are reconciled and reported to the agent separately, after you decide. Judge similarity of task and shape, never coverage of fields.
 
+UI shape includes the interaction model. A surface that observes input and filters data it already has (context only, no action: it narrows what is on the screen as you type and has nothing to press) is a different shape from one that submits an action to fetch or act (a button, a pending state, results that arrive). Do not match across that line, even when the fields look alike.
+
 NO-MATCH means the candidate is a fundamentally different thing — a different intended task (haiku composer vs tweet draft, login vs signup), a different UI shape (form vs list vs dashboard), OR a conflicting load-bearing fixed VALUE baked into the blueprint (calendar pinned to Jan vs a request for Mar — same contract shape, but the fixed value conflicts). Reserve NO-MATCH for these; do not decline a candidate merely because its fields, slots, or actions differ from the request.
 
 Visual style differences alone (minimal vs ornate, dense vs spacious) DO NOT block a match — the user can refine those after they get a working UI.

package/dist/rerank-eval/run-probe-cli.js

@@ -3,7 +3,7 @@
  * Rerank quality probe CLI.
  *
  * Reads `~/.ggui/credentials.json` for the Anthropic API key, runs
- * the probe against `claude-haiku-4-5`, prints the report.
+ * the probe against `claude-haiku-5-5` (ggui#1804), prints the report. Its readings are not comparable to runs made before the switch.
  *
  * Usage:
  *   pnpm -F @ggui-ai/negotiator probe-rerank
@@ -16,7 +16,7 @@
  */
 import { runProbe, formatReport } from './run-probe.js';
 import { buildAnthropicLlmCaller, getTokenUsage, resolveAnthropicKey, } from '../synth-bench/cli-llm.js';
-const DEFAULT_MODEL = 'claude-haiku-4-5';
+const DEFAULT_MODEL = 'claude-haiku-5-5';
 function parseArgs(argv) {
     let limit;
     let threshold;

package/dist/synth-bench/cli-llm.d.ts

@@ -1,6 +1,6 @@
 import type { LLMCaller } from '../llm-caller.js';
 /** Default bench model — the model the `ui-gen-default` seed generator declares. */
-export declare const DEFAULT_MODEL = "claude-haiku-4-5";
+export declare const DEFAULT_MODEL = "claude-haiku-5-5";
 /** Haiku 4.5 token pricing (USD per token) for the cost report. */
 export declare const HAIKU_4_5_PRICE_INPUT_PER_TOKEN: number;
 export declare const HAIKU_4_5_PRICE_OUTPUT_PER_TOKEN: number;

package/dist/synth-bench/cli-llm.js

@@ -11,7 +11,7 @@ import { resolve as pathResolve } from 'node:path';
 import { anthropicRejectsForcedToolChoice } from '@ggui-ai/protocol';
 const ANTHROPIC_API = 'https://api.anthropic.com/v1/messages';
 /** Default bench model — the model the `ui-gen-default` seed generator declares. */
-export const DEFAULT_MODEL = 'claude-haiku-4-5';
+export const DEFAULT_MODEL = 'claude-haiku-5-5';
 /** Haiku 4.5 token pricing (USD per token) for the cost report. */
 export const HAIKU_4_5_PRICE_INPUT_PER_TOKEN = 1.0 / 1_000_000;
 export const HAIKU_4_5_PRICE_OUTPUT_PER_TOKEN = 5.0 / 1_000_000;

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "@ggui-ai/negotiator",
-  "version": "0.27.0",
-  "description": "Contract-synthesis + match-judge engine for ggui's handshake. Synthesizes or repairs a conforming DataContract from an agent's draft, judges blueprint-match candidates for reuse, and validates contract structure + novelty — the primitives composed by decideHandshake in @ggui-ai/mcp-server-handlers. Deployment-agnostic: concrete embedding and vector-store bindings plug in via the storage interfaces from @ggui-ai/mcp-server-core.",
+  "version": "0.28.0",
+  "description": "Contract-synthesis + match-judge engine for GGUI's handshake. Synthesizes or repairs a conforming DataContract from an agent's draft, judges blueprint-match candidates for reuse, and validates contract structure + novelty — the primitives composed by decideHandshake in @ggui-ai/mcp-server-handlers. Deployment-agnostic: concrete embedding and vector-store bindings plug in via the storage interfaces from @ggui-ai/mcp-server-core.",
   "license": "Apache-2.0",
   "keywords": [
     "ggui",
@@ -47,8 +47,8 @@
     }
   },
   "dependencies": {
-    "@ggui-ai/mcp-server-core": "0.27.0",
-    "@ggui-ai/protocol": "0.27.0"
+    "@ggui-ai/mcp-server-core": "0.28.0",
+    "@ggui-ai/protocol": "0.28.0"
   },
   "devDependencies": {
     "@types/node": "^24.0.0",

package/src/llm-rerank.ts

@@ -90,6 +90,8 @@ const RERANK_SYSTEM_PROMPT = `You match user UI requests against previously-gene
 
 MATCH means the candidate is the same intended user task AND the same broad UI shape (component types, layout pattern). A candidate still MATCHES when the current request adds or omits fields, slots, or actions relative to the cached blueprint — a superset, a subset, or an overlapping wire surface all still match. Added or omitted fields/slots/actions DO NOT block a match and are NOT yours to judge: those wire-surface deltas are reconciled and reported to the agent separately, after you decide. Judge similarity of task and shape, never coverage of fields.
 
+UI shape includes the interaction model. A surface that observes input and filters data it already has (context only, no action: it narrows what is on the screen as you type and has nothing to press) is a different shape from one that submits an action to fetch or act (a button, a pending state, results that arrive). Do not match across that line, even when the fields look alike.
+
 NO-MATCH means the candidate is a fundamentally different thing — a different intended task (haiku composer vs tweet draft, login vs signup), a different UI shape (form vs list vs dashboard), OR a conflicting load-bearing fixed VALUE baked into the blueprint (calendar pinned to Jan vs a request for Mar — same contract shape, but the fixed value conflicts). Reserve NO-MATCH for these; do not decline a candidate merely because its fields, slots, or actions differ from the request.
 
 Visual style differences alone (minimal vs ornate, dense vs spacious) DO NOT block a match — the user can refine those after they get a working UI.

package/src/rerank-eval/run-probe-cli.ts

@@ -3,7 +3,7 @@
  * Rerank quality probe CLI.
  *
  * Reads `~/.ggui/credentials.json` for the Anthropic API key, runs
- * the probe against `claude-haiku-4-5`, prints the report.
+ * the probe against `claude-haiku-5-5` (ggui#1804), prints the report. Its readings are not comparable to runs made before the switch.
  *
  * Usage:
  *   pnpm -F @ggui-ai/negotiator probe-rerank
@@ -21,7 +21,7 @@ import {
   resolveAnthropicKey,
 } from '../synth-bench/cli-llm.js';
 
-const DEFAULT_MODEL = 'claude-haiku-4-5';
+const DEFAULT_MODEL = 'claude-haiku-5-5';
 
 function parseArgs(argv: readonly string[]): { limit?: number; threshold?: number; model: string } {
   let limit: number | undefined;

package/src/synth-bench/cli-llm.ts

@@ -14,7 +14,7 @@ import { anthropicRejectsForcedToolChoice } from '@ggui-ai/protocol';
 const ANTHROPIC_API = 'https://api.anthropic.com/v1/messages';
 
 /** Default bench model — the model the `ui-gen-default` seed generator declares. */
-export const DEFAULT_MODEL = 'claude-haiku-4-5';
+export const DEFAULT_MODEL = 'claude-haiku-5-5';
 
 /** Haiku 4.5 token pricing (USD per token) for the cost report. */
 export const HAIKU_4_5_PRICE_INPUT_PER_TOKEN = 1.0 / 1_000_000;