specrails-desktop 2.46.0 → 2.48.0

package/server/dist/rails-router.js

@@ -2,6 +2,7 @@
 Object.defineProperty(exports, "__esModule", { value: true });
 exports.prDeliveryRevisionAllowed = prDeliveryRevisionAllowed;
 exports.createRailsRouter = createRailsRouter;
+const agent_runtime_settings_1 = require("./agent-runtime-settings");
 const express_1 = require("express");
 const rails_store_1 = require("./rails-store");
 const queue_manager_1 = require("./queue-manager");
@@ -11,7 +12,9 @@ const file_provenance_1 = require("./file-provenance");
 const file_story_1 = require("./file-story");
 const loops_store_1 = require("./loops-store");
 const loop_runs_store_1 = require("./loop-runs-store");
+const registry_1 = require("./providers/registry");
 const providers_1 = require("./providers");
+const loop_role_engines_1 = require("./loop-role-engines");
 const runtime_1 = require("./providers/runtime");
 const agent_defaults_1 = require("./agent-defaults");
 const profile_manager_1 = require("./profile-manager");
@@ -373,8 +376,9 @@ function createRailsRouter() {
         }
         const value = body.aiEngine;
         const c = ctx(req);
-        // null clears the override; a string must be one of the project's providers.
-        if (value !== null) {
+        // null clears the override; a string must be one of the project's providers
+        // — or the `roles` sentinel (per-role engines from Settings ▸ Specrails Agents).
+        if (value !== null && value !== loop_role_engines_1.ROLES_ENGINE) {
             const check = (0, provider_selection_1.validateRequestedProvider)(c.project, value);
             if (!check.ok) {
                 res.status(400).json({ error: check.error });
@@ -432,6 +436,14 @@ function createRailsRouter() {
             res.status(400).json({ error: 'Invalid rail index' });
             return;
         }
+        let runtimeProviderOverride;
+        try {
+            runtimeProviderOverride = (0, agent_runtime_settings_1.validateRuntimeProviderOverride)(req.body?.runtimeProviderOverride);
+        }
+        catch {
+            res.status(400).json({ error: 'invalid_runtime_provider_override' });
+            return;
+        }
         let { mode = 'implement' } = req.body ?? {};
         const { repositoryIds: rawRepositoryIds, baseDeliveryIds, profileName, aiEngine, model, loopId: rawLoopId, reasoning_effort, originConversationId, originSurface, targetPrNumber, revisionOfDeliveryId, revisionNote, baseBranch: rawBaseBranch } = req.body ?? {};
         // Revision launch (nontech-review-experience Wave 3): the user asked for a
@@ -626,7 +638,21 @@ function createRailsRouter() {
         }
         // AI engine precedence: explicit body param > stored rail engine > primary.
         // `undefined`/empty in both means "run on the project's primary provider".
-        const requestedEngine = aiEngine === undefined ? (rail.aiEngine ?? undefined) : aiEngine;
+        const requestedEngineRaw = aiEngine === undefined ? (rail.aiEngine ?? undefined) : aiEngine;
+        // `roles` (hybrid-role-engines): the pipeline roles keep the project's
+        // per-role runtime settings (no flattening override) and the loop's own
+        // steps come from the stored loop-role engines — verifier for every
+        // non-core ai-step, decider for the Loop Decider. The rail's "engine" for
+        // the non-core path is therefore the verifier's provider.
+        const rolesMode = requestedEngineRaw === loop_role_engines_1.ROLES_ENGINE;
+        if (rolesMode && mode === 'freestyle') {
+            res.status(400).json({ error: 'roles_engine_unsupported_mode', detail: 'Freestyle has no roles; pick a concrete engine' });
+            return;
+        }
+        const loopRoleEngines = rolesMode ? (0, loop_role_engines_1.loadLoopRoleEngines)(c.project) : {};
+        const requestedEngine = rolesMode
+            ? (loopRoleEngines.verifier?.provider && (0, provider_selection_1.validateRequestedProvider)(c.project, loopRoleEngines.verifier.provider).ok ? loopRoleEngines.verifier.provider : undefined)
+            : requestedEngineRaw;
         const engineCheck = (0, provider_selection_1.validateRequestedProvider)(c.project, requestedEngine);
         if (!engineCheck.ok) {
             res.status(400).json({ error: engineCheck.error });
@@ -635,6 +661,37 @@ function createRailsRouter() {
         // Only pass a provider override when one was actually requested (keeps
         // single-provider rails on the legacy code path).
         const railProvider = requestedEngine ? engineCheck.provider : undefined;
+        // Mission/MCP and Launch-all submit aiEngine or use the saved rail engine.
+        // Resolve this at the shared boundary, not only in the Dashboard payload.
+        // Local (OpenAI-compatible) engines: the override carries provider + model
+        // only — OpenAI endpoints have no effort knob, so effort is DROPPED (never
+        // forwarded to Core's executor), and cost caps are refused (a local run has
+        // no billable cost to cap against).
+        const localEngine = (0, registry_1.isLocalAdapterId)(engineCheck.provider);
+        if (rolesMode && runtimeProviderOverride) {
+            res.status(400).json({ error: 'runtime_provider_mismatch', detail: 'A roles launch carries no provider override' });
+            return;
+        }
+        if (!runtimeProviderOverride && requestedEngine && !rolesMode) {
+            try {
+                runtimeProviderOverride = (0, agent_runtime_settings_1.validateRuntimeProviderOverride)({ provider: engineCheck.provider, ...(model ? { model } : {}), ...(reasoning_effort && !localEngine ? { effort: reasoning_effort } : {}) });
+            }
+            catch {
+                res.status(400).json({ error: 'invalid_runtime_provider_override' });
+                return;
+            }
+        }
+        if (runtimeProviderOverride && runtimeProviderOverride.provider !== engineCheck.provider) {
+            res.status(400).json({ error: 'runtime_provider_mismatch' });
+            return;
+        }
+        if (localEngine && runtimeProviderOverride?.effort !== undefined) {
+            runtimeProviderOverride = { provider: runtimeProviderOverride.provider, ...(runtimeProviderOverride.model !== undefined ? { model: runtimeProviderOverride.model } : {}) };
+        }
+        if (localEngine && req.body?.maxCostUsd !== undefined && req.body?.maxCostUsd !== null) {
+            res.status(400).json({ error: 'local_engine_no_cost_cap' });
+            return;
+        }
         // Freestyle bypasses the OpenSpec pipeline and hands the raw spec to a
         // provider that explicitly advertises autonomous freestyle support.
         if (mode === 'freestyle' && (0, providers_1.getAdapter)(engineCheck.provider).capabilities.freestyle !== true) {
@@ -755,26 +812,40 @@ function createRailsRouter() {
                         return;
                     }
                 }
+                const verifierEngine = rolesMode ? (0, loop_role_engines_1.resolveLoopRoleEngine)(c.project, 'verifier', { provider: loopProvider, model: loopAdapter.defaultModel() }, loopRoleEngines) : undefined;
                 const loopModel = typeof model === 'string' && model
                     ? model
-                    : (profileModel ?? globalAgentDefaults?.pipelineModel ?? loopAdapter.defaultModel());
+                    : (verifierEngine?.model ?? profileModel ?? globalAgentDefaults?.pipelineModel ?? loopAdapter.defaultModel());
                 let effort;
-                if (reasoning_effort !== undefined && reasoning_effort !== null) {
+                // Local (OpenAI-compatible) engines have no per-invocation effort knob
+                // unless the connection opts in; a requested effort is DROPPED, not a
+                // 400 — the engine selectors send the last-used effort blindly.
+                const effortRequested = localEngine && !loopAdapter.capabilities.supportsReasoningEffort ? undefined : (reasoning_effort ?? verifierEngine?.effort);
+                if (effortRequested !== undefined && effortRequested !== null) {
                     const allowed = (0, providers_1.reasoningEffortsForModel)(loopAdapter, loopModel);
-                    if (typeof reasoning_effort !== 'string' ||
-                        !allowed.includes(reasoning_effort)) {
+                    if (typeof effortRequested !== 'string' ||
+                        !allowed.includes(effortRequested)) {
                         res.status(400).json({
                             error: `reasoning_effort is not valid for provider "${loopProvider}" and model "${loopModel}"`,
                             allowed,
                         });
                         return;
                     }
-                    effort = reasoning_effort;
+                    effort = effortRequested;
                 }
                 else if (globalAgentDefaults?.pipelineEffort
                     && (0, runtime_1.isReasoningEffortValidForModel)(loopAdapter, loopModel, globalAgentDefaults.pipelineEffort)) {
                     effort = globalAgentDefaults.pipelineEffort;
                 }
+                // Loop Decider engine under roles mode (else the rail's engine, as before).
+                // Its provider must enforce a read-only tool policy like the rail's.
+                const deciderEngine = rolesMode && loopGraph.nodes.some((node) => node.type === 'decider')
+                    ? (0, loop_role_engines_1.resolveLoopRoleEngine)(c.project, 'decider', { provider: loopProvider, model: loopModel, ...(effort ? { effort } : {}) }, loopRoleEngines)
+                    : undefined;
+                if (deciderEngine && !(0, providers_1.supportsToolPolicy)((0, providers_1.getAdapter)(deciderEngine.provider), 'read-only')) {
+                    res.status(409).json({ code: 'provider_tool_policy_unsupported', provider: deciderEngine.provider, requiredPolicy: 'read-only', error: `Provider '${deciderEngine.provider}' cannot run Loop Deciders because its headless CLI does not enforce a read-only tool policy.` });
+                    return;
+                }
                 // The loop's command(s) declare ticket scope + required capabilities.
                 const promptsText = loopGraph.nodes
                     .filter((n) => n.type === 'ai-step')
@@ -907,8 +978,10 @@ function createRailsRouter() {
                         }
                         try {
                             const ids = await (0, rail_isolated_launch_1.launchIsolatedRail)({
+                                runtimeProviderOverride,
                                 ctx: c, railIndex, ticketIds: [...rail.ticketIds], repositoryIds, ...repositoryBases, loopId, loopName, loopGraph,
                                 provider: loopProvider, model: loopModel, effort, scope,
+                                ...(deciderEngine ? { deciderEngine } : {}),
                                 profileName: resolvedProfile,
                                 originSurface: originSurface ?? 'dashboard',
                                 originConversationId: originConversationId ?? null,
@@ -1015,6 +1088,7 @@ function createRailsRouter() {
                     };
                     const runPromise = c.loopRunManager
                         .run({
+                        runtimeProviderOverride,
                         runId,
                         loopId,
                         loopName,
@@ -1040,6 +1114,7 @@ function createRailsRouter() {
                         provider: loopProvider,
                         model: loopModel,
                         effort,
+                        ...(deciderEngine ? { deciderEngine } : {}),
                         profileName: resolvedProfile,
                     })
                         .then((r) => {
@@ -1218,6 +1293,13 @@ function createRailsRouter() {
             catch (err) {
                 console.warn(`[rails-router] stop: loop cancel(${runId}) failed: ${err.message}`);
             }
+            // A run the engine no longer owns (reconciled as an orphan by a context
+            // reload while still alive) never settles through onLoopRunFinished, so
+            // its rail entry would pin the tickets "in flight" forever. When the
+            // durable row is already terminal, release the entry here.
+            const row = (0, loop_runs_store_1.getLoopRun)(c.db, runId);
+            if (!row || row.status === 'completed')
+                c.railLoopRuns.delete(runId);
             const stopMsg = {
                 type: 'loop.run_stopped',
                 projectId: c.project.id,

package/server/dist/result-event.js

@@ -23,6 +23,7 @@ Object.defineProperty(exports, "__esModule", { value: true });
 exports.finaliseInvocationResult = finaliseInvocationResult;
 exports.normaliseResultEvent = normaliseResultEvent;
 const pricing_1 = require("./pricing");
+const local_adapter_1 = require("./providers/local-adapter");
 /**
  * Finalise an AI invocation by walking the adapter's parsed events, stamping
  * the model when the adapter does not report one, and applying the local
@@ -68,6 +69,22 @@ function finaliseInvocationResult(adapter, events, opts = {}) {
     // (COST-ACCOUNTING-AUDIT CRIT-1). A genuine native cost — including a
     // numeric 0 — still passes through untouched with estimated=false.
     const nativeCostAbsent = cloned.total_cost_usd === null || cloned.total_cost_usd === undefined;
+    // Local (OpenAI-compatible) engines are NEVER rate-card guessed: cost is the
+    // connection's optional user-supplied rates (flagged estimated) or NULL —
+    // honest "cost unknown", never $0 (design D9).
+    if ((0, local_adapter_1.isLocalAdapter)(adapter)) {
+        if (nativeCostAbsent && hasBillableUsage) {
+            const computed = (0, pricing_1.estimateLocalCostUsd)(adapter.localConnection.rates, {
+                tokens_in: cloned.tokens_in,
+                tokens_out: cloned.tokens_out,
+            });
+            if (computed !== null) {
+                cloned.total_cost_usd = computed;
+                estimated = true;
+            }
+        }
+        return { result: cloned, estimated };
+    }
     if (nativeCostAbsent && hasBillableUsage) {
         const estimator = opts.estimator ?? pricing_1.estimateCostUsd;
         const computed = estimator(adapter.id, cloned.model, {

package/server/dist/review-packet.js

@@ -240,6 +240,7 @@ function composeReviewPacket({ db, row, repositoryId = row.repository_id ?? unde
             title: snap?.title ?? null,
             problem: narrative.problem,
             solution: narrative.solution,
+            solutionOverflow: narrative.overflow,
             labels: snap?.labels ?? [],
             implementationOutcome: unitRows[0]?.implementationOutcome ?? null,
             deliveryOutcome: unitRows[0]?.deliveryOutcome ?? null,

package/server/dist/schemas/agent-runtime.schema.json

@@ -7,10 +7,17 @@
   "required": ["schemaVersion", "providers", "agents", "verification"],
   "properties": {
     "schemaVersion": { "const": 1 },
+    "efficiency": { "type": "object", "additionalProperties": false, "required": ["schemaVersion"], "properties": {
+      "schemaVersion": { "const": 1 }, "contextMode": { "enum": ["full", "incremental"] }, "reviewMode": { "enum": ["full", "incremental"] },
+      "planning": { "enum": ["full", "proportional"] }, "acceptDeveloperChecks": { "type": "boolean" },
+      "verification": { "type": "object", "additionalProperties": false, "properties": { "maxConcurrency": { "type": "integer", "minimum": 1, "maximum": 4 } } }
+    } },
+    "guardrails": { "type": "object", "additionalProperties": false, "properties": { "plan-validation": { "type": "boolean" }, "genuine-blocking-question": { "type": "boolean" }, "plan-language-hint": { "type": "boolean" }, "frozen-plan-writes": { "type": "boolean" }, "primary-language": { "type": "boolean" }, "duplicate-sibling": { "type": "boolean" }, "one-test-per-module": { "type": "boolean" }, "empty-write": { "type": "boolean" }, "evidence-gated-ticking": { "type": "boolean" }, "inventory-retry": { "type": "boolean" }, "synthetic-corrections": { "type": "boolean" }, "silent-group-closure": { "type": "boolean" }, "verify-per-group": { "type": "boolean" }, "test-reachability": { "type": "boolean" }, "exit-code-honesty": { "type": "boolean" }, "environment-repair": { "type": "boolean" }, "lockfile-repair": { "type": "boolean" }, "verify-idle-timeout": { "type": "boolean" } } },
     "rolePrompts": { "type": "object", "additionalProperties": false, "properties": {
       "architect": { "type": "string", "minLength": 1, "maxLength": 20000, "pattern": "\\S" },
       "developer": { "type": "string", "minLength": 1, "maxLength": 20000, "pattern": "\\S" },
-      "reviewer": { "type": "string", "minLength": 1, "maxLength": 20000, "pattern": "\\S" }
+      "reviewer": { "type": "string", "minLength": 1, "maxLength": 20000, "pattern": "\\S" },
+      "fixer": { "type": "string", "minLength": 1, "maxLength": 20000, "pattern": "\\S" }
     } },
     "enabled": { "type": "boolean", "deprecated": true, "description": "Accepted for saved configuration compatibility; implementation always uses the agent runtime." },
     "providers": { "type": "array", "items": { "oneOf": [
@@ -20,9 +27,14 @@
       { "type": "object", "additionalProperties": false, "required": ["id", "kind", "baseUrl"], "properties": {
         "id": { "$ref": "#/definitions/id" }, "kind": { "const": "openai-compatible" },
         "baseUrl": { "type": "string", "pattern": "^https?://[^/?#@]+(?:/[^?#]*)?$" },
-        "apiKeyEnv": { "type": "string", "pattern": "^[A-Za-z_][A-Za-z0-9_]*$" }
+        "apiKeyEnv": { "type": "string", "pattern": "^[A-Za-z_][A-Za-z0-9_]*$" },
+        "supportsReasoningEffort": { "type": "boolean", "description": "The endpoint honours the OpenAI reasoning_effort request field (low|medium|high)." },
+        "agentLoop": { "enum": ["compact", "free"], "description": "compact (default): host-driven pipelines of small structured calls for small local models; free: one agentic loop per role with guardrails." },
+        "contextWindowTokens": { "type": "integer", "minimum": 4096, "description": "Context window used for the compaction budget (chars/4 estimate). Default 32768." },
+        "maxOutputTokens": { "type": "integer", "minimum": 1024, "description": "Output budget (max_tokens) of one tool turn; retried once at twice this when cut off, always bounded by the remaining context window. Default 8192." }
       } }
     ] } },
+    "fixer": { "$ref": "#/definitions/agent" },
     "agents": { "type": "object", "additionalProperties": false, "required": ["architect", "developer", "reviewer"], "properties": {
       "architect": { "$ref": "#/definitions/agent" }, "developer": { "$ref": "#/definitions/agent" }, "reviewer": { "$ref": "#/definitions/agent" }
     } },
@@ -31,6 +43,7 @@
       "maxCostUsd": { "type": "number", "exclusiveMinimum": 0 }, "timeoutMs": { "$ref": "#/definitions/positiveInteger" }
     } },
     "verification": { "type": "array", "items": { "type": "object", "additionalProperties": false, "required": ["repositoryId", "command", "args"], "properties": {
+      "key": { "$ref": "#/definitions/id" }, "label": { "type": "string", "minLength": 1, "maxLength": 256, "pattern": "\\S" }, "policy": { "$ref": "#/definitions/checkPolicy" },
       "repositoryId": { "$ref": "#/definitions/id" }, "command": { "type": "string", "minLength": 1 },
       "args": { "type": "array", "items": { "type": "string" } }, "cwd": { "type": "string", "minLength": 1 },
       "env": { "type": "object", "propertyNames": { "pattern": "^[A-Za-z_][A-Za-z0-9_]*$" }, "additionalProperties": { "type": "string" } },
@@ -49,10 +62,22 @@
     } }
   },
   "definitions": {
+    "effort": { "type": "string", "pattern": "^[a-z][a-z0-9_-]{0,31}$" },
+    "inputList": { "type": "array", "maxItems": 256, "items": { "type": "string", "minLength": 1, "maxLength": 4096, "pattern": "\\S" } },
+    "checkPolicy": { "type": "object", "additionalProperties": false, "properties": {
+      "reuse": { "enum": ["never", "snapshot-local"] }, "inputs": { "$ref": "#/definitions/inputList" },
+      "toolchainInputs": { "$ref": "#/definitions/inputList" }, "resources": { "$ref": "#/definitions/inputList" },
+      "deterministic": { "type": "boolean" }, "readOnly": { "type": "boolean" }, "independentGroup": { "$ref": "#/definitions/id" }
+    } },
     "id": { "type": "string", "pattern": "^[A-Za-z0-9][A-Za-z0-9._-]{0,127}$" },
     "positiveInteger": { "type": "integer", "minimum": 1 },
     "score": { "type": "number", "minimum": 0, "maximum": 100 },
-    "agent": { "type": "object", "additionalProperties": false, "required": ["provider"], "properties": {
+    "agent": { "type": "object", "additionalProperties": false, "required": ["provider"], "dependencies": { "escalation": ["model"] }, "properties": {
+      "effort": { "$ref": "#/definitions/effort" },
+      "thinking": { "type": "string", "enum": ["on", "off"], "description": "OpenAI-compatible engines only: private thinking (chat_template_kwargs.enable_thinking). Default off." },
+      "escalation": { "type": "object", "additionalProperties": false, "required": ["model"], "properties": {
+        "model": { "type": "string", "minLength": 1, "maxLength": 256, "pattern": "^[^-\\r\\n\\u0000][^\\r\\n\\u0000]*$" }, "effort": { "$ref": "#/definitions/effort" }
+      } },
       "provider": { "$ref": "#/definitions/id" }, "model": { "type": "string", "minLength": 1, "maxLength": 256, "pattern": "^[^-\\r\\n\\u0000][^\\r\\n\\u0000]*$" },
       "maxTurns": { "$ref": "#/definitions/positiveInteger" }
     } }