specrails-desktop 2.64.2 → 2.65.0

package/server/dist/modules/execution/runtime/queue-manager.js

@@ -19,6 +19,7 @@ const rail_isolation_1 = require("../../delivery/runtime/rail-isolation");
 const repo_map_1 = require("../../../repo-map");
 const cli_prompt_1 = require("../../../util/cli-prompt");
 const stream_display_1 = require("../../../util/stream-display");
+const audit_1 = require("../../audit");
 const hooks_1 = require("../../../hooks");
 const ai_invocations_1 = require("../../accounting/runtime/ai-invocations");
 const feature_flags_1 = require("../../../feature-flags");
@@ -34,6 +35,7 @@ const db_1 = require("../../../db");
 const interactive_job_session_1 = require("./interactive-job-session");
 const attachment_manager_1 = require("../../../attachment-manager");
 const ticket_store_1 = require("../../specs/runtime/ticket-store");
+const non_goals_1 = require("../../specs/domain/non-goals");
 const spec_addenda_1 = require("../../specs/runtime/spec-addenda");
 const binary_probe_1 = require("../../../binary-probe");
 const workspace_manager_1 = require("../../../workspace-manager");
@@ -130,12 +132,15 @@ class QueueManager {
     _lifecycleGeneration;
     _getCostAlertThreshold;
     _getDesktopDailyBudget;
+    _getPeriodCapBreaches;
     _adapter;
     /** Effective model to use when spawning processes. For Claude the adapter
      *  reads its own config; this is the override that gets passed via `--model`.
      *  For codex it controls the catalog model used at spawn time and as the
      *  fallback model name stamped onto the ai_invocations row. */
     _resolvedModel;
+    /** Agent actions audit sink (project-bound). Observing only: never fails a job. */
+    _agentAudit = null;
     _agentDefaults;
     _onJobFinished;
     /** Project-owned durable effects that are part of every queue admission
@@ -241,6 +246,7 @@ class QueueManager {
         this._lifecycleGeneration = 0;
         this._getCostAlertThreshold = options?.getCostAlertThreshold ?? null;
         this._getDesktopDailyBudget = options?.getDesktopDailyBudget ?? null;
+        this._getPeriodCapBreaches = options?.getPeriodCapBreaches ?? null;
         this._adapter = (0, providers_1.getAdapter)(options?.provider ?? 'claude');
         this._resolvedModel = options?.resolvedModel ?? null;
         this._agentDefaults = options?.agentDefaults ?? null;
@@ -249,6 +255,7 @@ class QueueManager {
         this._onBudgetExceeded = options?.onBudgetExceeded ?? null;
         this._resolvePluginsForSpawn = options?.resolvePluginsForSpawn ?? null;
         this._projectId = options?.projectId ?? null;
+        this._agentAudit = options?.agentAudit ?? null;
         this._desktopPort = options?.desktopPort ?? 4200;
         this._projectSlug = options?.projectSlug ?? null;
         this._jobProfileSelection = new Map();
@@ -1217,8 +1224,7 @@ class QueueManager {
      * injected clock is unnecessary — it reads the real local day.
      */
     _localDayStartIso() {
-        const now = new Date();
-        return new Date(now.getFullYear(), now.getMonth(), now.getDate()).toISOString();
+        return (0, __1.localDayStartIso)(new Date());
     }
     /**
      * Evaluate the per-project and app-level daily budgets and pause the queue when
@@ -1255,6 +1261,30 @@ class QueueManager {
                     }
                 },
             });
+            // Weekly/monthly caps: same effect as the daily budget (pause + event),
+            // reported with the period so a webhook/UI can say which cap was reached.
+            for (const cap of this._getPeriodCapBreaches?.() ?? []) {
+                if (!cap.exhausted || cap.period === 'day')
+                    continue;
+                const wasPaused = this._paused;
+                this._paused = true;
+                if (!wasPaused)
+                    storage.pause();
+                if (cap.scope === 'project') {
+                    this._broadcast({ type: 'daily_budget_exceeded', projectId: '', dailySpend: cap.spentUsd, budget: cap.budgetUsd, queuePaused: true, period: cap.period });
+                    try {
+                        this._onBudgetExceeded?.('daily_budget_exceeded', { dailySpend: cap.spentUsd, budget: cap.budgetUsd, queuePaused: true, period: cap.period });
+                    }
+                    catch { /* webhook delivery is best-effort */ }
+                }
+                else {
+                    this._broadcast({ type: 'desktop_daily_budget_exceeded', projectId: '', desktopDailySpend: cap.spentUsd, desktopBudget: cap.budgetUsd, queuePaused: true, period: cap.period });
+                    try {
+                        this._onBudgetExceeded?.('desktop_daily_budget_exceeded', { desktopDailySpend: cap.spentUsd, desktopBudget: cap.budgetUsd, queuePaused: true, period: cap.period });
+                    }
+                    catch { /* webhook delivery is best-effort */ }
+                }
+            }
         }
         catch (err) {
             console.error('[queue-manager] daily-budget enforcement failed (db unavailable?):', err);
@@ -1336,7 +1366,7 @@ class QueueManager {
                     const ticket = store.tickets[String(ticketId)];
                     if (!ticket)
                         continue;
-                    const body = (ticket.description ?? '').trim();
+                    const body = (0, non_goals_1.descriptionWithNonGoals)(ticket.description, ticket.nonGoals).trim();
                     specs.push(`# Spec #${ticketId}: ${ticket.title}\n\n${body || '_(no description)_'}`);
                 }
             }
@@ -1710,6 +1740,7 @@ class QueueManager {
             // Auto sessions arm a wedge detector on the SAME inactivity budget the
             // one-shot path uses; finalize sessions idle awaiting the human.
             zombieTimeoutMs: settleMode === 'auto' ? this._zombieTimeoutMs : undefined,
+            ...(this._agentAudit ? { onProviderEvent: (0, audit_1.createStreamAuditTap)(this._agentAudit, { runId: jobId, jobId }) } : {}),
         });
         this._interactiveSessions.set(jobId, session);
         session.start(spec, firstPrompt);
@@ -2090,7 +2121,7 @@ class QueueManager {
         //    claude slash command — propose-spec, implement,
         //    explore-spec, retry, …). This is the rail equivalent of the
         //    user typing `$implement #1 --yes` themselves in `codex`.
-        // Freestyle (capability-gated; currently Claude and Kimi): skip the slash
+        // Freestyle (capability-gated; every bundled CLI and local engine): skip the slash
         // command entirely and send the pre-prompt + spec text directly as the
         // prompt. The server route rejects adapters that do not advertise
         // `freestyle` before this point.
@@ -2213,13 +2244,13 @@ class QueueManager {
         // claude honours OTEL_* env vars natively; codex does not and instead
         // gets signals synthesised by the codex-otel-bridge attached below.
         let spawnEnv = process.env;
-        // Per-project worktree/job env passthrough. The setting stores names only;
-        // values are read from the server env (with login-shell recovery for missing
-        // names) at spawn time. Apply this before Specrails' own env overlays so
-        // internal SPECRAILS_* control-plane values always win.
+        // Per-project agent environment: the passthrough overlay (names stored,
+        // values read at spawn time with login-shell recovery) and, in `minimal`
+        // mode, the allowlist filter. Apply this before Specrails' own env overlays
+        // so internal SPECRAILS_* control-plane values always win.
         let environmentNotice = null;
         if (this._db) {
-            spawnEnv = (0, project_env_1.applyWorktreeEnvPassthrough)(this._db, spawnEnv);
+            spawnEnv = (0, project_env_1.resolveProjectAgentEnv)(this._db, { providerAllowlist: (0, runtime_1.providerEnvAllowlist)(adapter) });
             // One value-free line naming unresolved configured names (written to the
             // job log once the readers exist). The run proceeds without them.
             environmentNotice = (0, project_env_1.worktreeEnvPassthroughWarning)(this._db, spawnEnv);
@@ -2359,14 +2390,18 @@ class QueueManager {
         // ─── Relocate-artifacts spawn env ──────────────────────────────────────
         // Merge SPECRAILS_REPO_DIR (+ workspace/tickets/state/etc.) so stage-3
         // `${SPECRAILS_REPO_DIR:-.}` re-pointing drives source/openspec/git I/O back
-        // into the repo. Per provider: gemini trusts the workspace cwd; codex gets
-        // NO CODEX_HOME override (all-or-nothing incl. auth → breaks the rail) and
-        // relies on cwd-based discovery from the workspace. Legacy ⇒ empty env, no-op.
+        // into the repo. Per provider: codex gets NO CODEX_HOME override
+        // (all-or-nothing incl. auth → breaks the rail) and relies on cwd-based
+        // discovery from the workspace. Legacy ⇒ empty env, no-op.
+        // Gemini trusts its spawn cwd in BOTH layouts (the workspace when
+        // relocated, the repo or rail worktree otherwise): headless `gemini -p`
+        // in an untrusted folder ignores the project's settings and subagents.
+        // The loop path gets the same through the runtime bridge (coreProviderEnv).
+        if (adapter.id === 'gemini') {
+            spawnEnv = { ...spawnEnv, GEMINI_CLI_TRUST_WORKSPACE: 'true' };
+        }
         if (execution.relocated) {
             spawnEnv = { ...spawnEnv, ...execution.env };
-            if (adapter.id === 'gemini') {
-                spawnEnv = { ...spawnEnv, GEMINI_CLI_TRUST_WORKSPACE: 'true' };
-            }
             // openspec PATH shim (claude rails only): prepend a per-job shim dir that
             // re-points every BARE `openspec` call at the repo working tree, so a
             // skill- or un-wrapped-template-driven `openspec <verb>` from the workspace
@@ -2613,6 +2648,8 @@ class QueueManager {
             persistEvent({ event_type: 'log', source: 'stderr', payload: JSON.stringify({ line: environmentNotice }) }, false);
             emitLine('stderr', environmentNotice);
         }
+        // Agent actions audit: observes parsed provider frames; never throws.
+        const auditTap = this._agentAudit ? (0, audit_1.createStreamAuditTap)(this._agentAudit, { runId: jobId, jobId }) : null;
         stdoutReader.on('line', (line) => {
             if (this._disposed)
                 return;
@@ -2621,6 +2658,8 @@ class QueueManager {
                 parsed = JSON.parse(line);
             }
             catch { /* plain text */ }
+            if (parsed && auditTap)
+                auditTap(parsed);
             // Feed the adapter for the canonical event shape used by
             // finaliseInvocationResult and (optionally) the OTEL bridge. Done
             // alongside the raw event persistence below, NOT in place of it: the

package/server/dist/modules/execution/runtime/run-duration-stats.js

@@ -3,8 +3,10 @@ Object.defineProperty(exports, "__esModule", { value: true });
 exports.MIN_DURATION_SAMPLES = void 0;
 exports.percentileMs = percentileMs;
 exports.rangeFromSamples = rangeFromSamples;
+exports.loopRunSamples = loopRunSamples;
 exports.getLoopDurationRange = getLoopDurationRange;
 exports.jobCommandShape = jobCommandShape;
+exports.jobCommandSamples = jobCommandSamples;
 exports.getJobCommandDurationRange = getJobCommandDurationRange;
 /** Fewer settled runs than this and no band is returned. */
 exports.MIN_DURATION_SAMPLES = 5;
@@ -34,16 +36,24 @@ function rangeFromSamples(samplesMs) {
 /** Cap the history window so a long-lived project's ancient runs (different
  * codebase, different model) cannot skew the band a user is shown today. */
 const HISTORY_LIMIT = 50;
-/** Measured band for a loop id, or null below the sample floor. */
-function getLoopDurationRange(db, projectId, loopId) {
+/** Settled runs of one loop, newest first, optionally of one provider. */
+function loopRunSamples(db, projectId, loopId, provider) {
     const rows = db.prepare(`
-    SELECT total_duration_ms AS ms
-      FROM loop_runs
-     WHERE project_id = ? AND loop_id = ? AND finished_at IS NOT NULL AND total_duration_ms > 0
-     ORDER BY started_at DESC
+    SELECT lr.total_duration_ms AS ms,
+           lr.total_cost_usd AS cost,
+           EXISTS (SELECT 1 FROM ai_invocations ai
+                    WHERE ai.loop_run_id = lr.id AND ai.total_cost_usd_estimated = 1) AS estimated
+      FROM loop_runs lr
+     WHERE lr.project_id = ? AND lr.loop_id = ? AND lr.finished_at IS NOT NULL AND lr.total_duration_ms > 0
+       AND (? IS NULL OR lr.provider = ?)
+     ORDER BY lr.started_at DESC
      LIMIT ?
-  `).all(projectId, loopId, HISTORY_LIMIT);
-    return rangeFromSamples(rows.map((row) => row.ms));
+  `).all(projectId, loopId, provider ?? null, provider ?? null, HISTORY_LIMIT);
+    return rows.map((row) => ({ ms: row.ms, costUsd: billedCost(row.cost, row.estimated === 1) }));
+}
+/** Measured band for a loop id, or null below the sample floor. */
+function getLoopDurationRange(db, projectId, loopId) {
+    return rangeFromSamples(loopRunSamples(db, projectId, loopId).map((sample) => sample.ms));
 }
 /**
  * Commands carry per-run arguments (`loop:<uuid>`, `/specrails:implement #12`),
@@ -56,16 +66,24 @@ function jobCommandShape(command) {
         return 'loop:';
     return head;
 }
-/** Measured band for a job command shape, or null below the sample floor. */
-function getJobCommandDurationRange(db, commandShape) {
+/** Completed jobs of one command shape, newest first, optionally of one provider. */
+function jobCommandSamples(db, commandShape, provider) {
     const rows = db.prepare(`
-    SELECT duration_ms AS ms, command
+    SELECT duration_ms AS ms, command, total_cost_usd AS cost, total_cost_usd_estimated AS estimated, provider
       FROM jobs
      WHERE status = 'completed' AND duration_ms > 0
      ORDER BY started_at DESC
      LIMIT 500
   `).all();
-    return rangeFromSamples(rows.filter((row) => jobCommandShape(row.command ?? '') === commandShape)
+    return rows
+        .filter((row) => jobCommandShape(row.command ?? '') === commandShape && (!provider || row.provider === provider))
         .slice(0, HISTORY_LIMIT)
-        .map((row) => row.ms));
+        .map((row) => ({ ms: row.ms, costUsd: billedCost(row.cost, row.estimated === 1) }));
+}
+/** Measured band for a job command shape, or null below the sample floor. */
+function getJobCommandDurationRange(db, commandShape) {
+    return rangeFromSamples(jobCommandSamples(db, commandShape).map((sample) => sample.ms));
+}
+function billedCost(cost, estimated) {
+    return !estimated && typeof cost === 'number' && Number.isFinite(cost) && cost > 0 ? cost : null;
 }