@pugi/cli 0.1.0-beta.97 → 0.1.0-beta.99

package/dist/core/engine/budgets.js

@@ -75,12 +75,21 @@ export const beta1DefaultBudgets = {
     // real per-call token use is ~30-40% lower than legacy. Bump headroom
     // so multi-file refactors no longer trip the cap. Anvil clamps per-call
     // max_tokens to 128k (PR) so the engine envelope still safe.
-    fix: { maxTokens: 80_000, maxToolCalls: 20 },
-    code: { maxTokens: 120_000, maxToolCalls: 25 },
-    build: { maxTokens: 200_000, maxToolCalls: 30 },
-    plan: { maxTokens: 200_000, maxToolCalls: 8 },
-    explain: { maxTokens: 60_000, maxToolCalls: 10 },
-    review_triple: { maxTokens: 100_000, maxToolCalls: 10 },
+    // CEO escalation 2026-06-05 round 2: "почему другие модели пишут
+    // днями? ты ограничиваешь токены?" — the upstream / Codex / Aider have
+    // NO per-task budget cap. They auto-compact at ~75% threshold and
+    // continue indefinitely. Pugi's 400K cap was still a regression vs
+    // industry expectation. Raise к the HARD_MAX_TOKENS ceiling (5M) for
+    // code/fix/build so the engine loop effectively runs until completion
+    // or until the operator aborts. Auto-compact (#14, shipped) reclaims
+    // headroom mid-loop. Token cost still tracked + tier-gated upstream;
+    // this lifts only the in-process loop ceiling.
+    fix: { maxTokens: 2_000_000, maxToolCalls: 200 },
+    code: { maxTokens: 5_000_000, maxToolCalls: 500 },
+    build: { maxTokens: 5_000_000, maxToolCalls: 500 },
+    plan: { maxTokens: 500_000, maxToolCalls: 30 },
+    explain: { maxTokens: 200_000, maxToolCalls: 20 },
+    review_triple: { maxTokens: 300_000, maxToolCalls: 20 },
 };
 /**
  * Hard upper bounds. Anything above this is treated as user error

package/dist/core/engine/intensity.js

@@ -49,26 +49,31 @@ const PROFILES = {
         allowParallelAgents: false,
         maxParallelAgents: 0,
     },
+    // CEO 2026-06-05: 80K standard exhausted React multi-file build mid-
+    // turn (120K hardcoded budget). Customers compare to the upstream = 200K
+    // context per session. Bump standard к 200K so default REPL doesn't
+    // trip mid-build; deep к 500K for complex multi-file refactors;
+    // marathon к 1.5M for long-running autonomous work.
     standard: {
         level: 'standard',
-        maxTurns: 15,
-        budgetTokens: 80_000,
+        maxTurns: 30,
+        budgetTokens: 200_000,
         modelTag: 'standard',
         allowParallelAgents: false,
         maxParallelAgents: 0,
     },
     deep: {
         level: 'deep',
-        maxTurns: 50,
-        budgetTokens: 200_000,
+        maxTurns: 80,
+        budgetTokens: 500_000,
         modelTag: 'standard',
         allowParallelAgents: true,
         maxParallelAgents: 3,
     },
     marathon: {
         level: 'marathon',
-        maxTurns: 200,
-        budgetTokens: 800_000,
+        maxTurns: 300,
+        budgetTokens: 1_500_000,
         modelTag: 'heavy',
         allowParallelAgents: true,
         maxParallelAgents: 3,

package/dist/core/repl/session.js

@@ -2831,6 +2831,24 @@ export class ReplSession {
         try {
             if (useDirectEngine) {
                 const persona = personaSlugFor('code');
+                // PR C (PUGI-538-FU): thread the recent conversation
+                // into the engine prompt so multi-turn refinements work. Without
+                // this, the engine sees only the literal current brief — a
+                // follow-up like "react" after "сделай крестики нолики" arrives
+                // as a bare "react" with no prior context, and the engine ships
+                // arbitrary nonsense or asks again ("нет конкретного feature
+                // request"). The CEO reproduction 2026-06-05 (Python tic-tac-toe
+                // shipped когда customer wanted React браузер game, then engine
+                // claimed "нет feature request" on the correction turn) is
+                // exactly this gap.
+                //
+                // Display channels (system line, transcript) keep using the bare
+                // `brief` for UX cleanliness; only the engine's task.prompt gets
+                // the full conversational context via the new `enginePrompt`
+                // field. Engine-bridge falls back to brief when enginePrompt is
+                // undefined (server-emitted parser-built tags), preserving the
+                // legacy behaviour for those surfaces.
+                const enginePrompt = this.buildEnginePromptWithContext(brief);
                 const tag = {
                     command: 'code',
                     brief,
@@ -2842,6 +2860,7 @@ export class ReplSession {
                     signature: signatureForToolRoute('code', persona, brief),
                     start: 0,
                     end: 0,
+                    ...(enginePrompt !== brief ? { enginePrompt } : {}),
                 };
                 await this.runEngineBridge(tag);
             }
@@ -2862,6 +2881,54 @@ export class ReplSession {
             this.markDispatchFailed('post_brief_failed');
         }
     }
+    /**
+     * PR C (PUGI-538-FU): build the engine prompt with recent
+     * conversation context prepended. The current brief is preserved as
+     * the explicit "Current request:" terminal so the engine knows what
+     * the user is asking right now, while the prior turns give it the
+     * stack/framework/format hints from earlier in the dialog.
+     *
+     * Returns `brief` unchanged when there is no prior conversation —
+     * the empty preamble would just waste tokens.
+     *
+     * Window policy: last 4 conversational exchanges (operator + persona
+     * pairs), text truncated к 400 chars per row. Drops the trailing
+     * operator row if it matches `brief` (which has already been appended
+     * to the transcript by `appendOperatorLine` at line 3429 above and
+     * would otherwise duplicate inside the prompt).
+     *
+     * Doc strings stay в English per repo convention; the rendered
+     * preamble uses neutral English labels ("User", "Pugi") so the
+     * engine's model treats it as standard transcript context rather
+     * than a localized field name.
+     */
+    buildEnginePromptWithContext(brief) {
+        const MAX_TURNS = 4;
+        const MAX_ROW_CHARS = 400;
+        const conversational = this.state.transcript.filter((r) => r.source === 'operator' || r.source === 'persona');
+        if (conversational.length === 0)
+            return brief;
+        // Take the last MAX_TURNS * 2 rows (each turn = 1 operator + 1 persona).
+        const recent = conversational.slice(-(MAX_TURNS * 2));
+        // Drop trailing operator row when it equals the brief we're about
+        // to dispatch — the brief is the "current request" and already
+        // landed in the transcript via `appendOperatorLine` earlier in
+        // `dispatchBrief`. Including it twice would confuse the engine.
+        const lastRow = recent[recent.length - 1];
+        const trimmed = lastRow && lastRow.source === 'operator' && lastRow.text === brief
+            ? recent.slice(0, -1)
+            : recent;
+        if (trimmed.length === 0)
+            return brief;
+        const lines = trimmed.map((r) => {
+            const role = r.source === 'operator' ? 'User' : 'Pugi';
+            const truncated = r.text.length > MAX_ROW_CHARS
+                ? r.text.slice(0, MAX_ROW_CHARS) + '...'
+                : r.text;
+            return `- ${role}: ${truncated}`;
+        });
+        return `Recent conversation:\n${lines.join('\n')}\n\nCurrent request: ${brief}`;
+    }
     /**
      * : reset the FSM to `idle` after a terminal transition so the
      * next brief can start. The FSM does not allow direct
@@ -4137,7 +4204,11 @@ export class ReplSession {
             result = await bridge({
                 command: tag.command,
                 persona: tag.persona,
-                brief: tag.brief,
+                // PR C (PUGI-538-FU): prefer the contextualized
+                // engine prompt when the direct-engine path set it. Falls back
+                // к the bare brief for parser-built tags from the server-emitted
+                // envelope path (no conversation context available there).
+                brief: tag.enginePrompt ?? tag.brief,
                 bridgeId,
                 signal: abort.signal,
                 onEvent,

package/dist/core/subagents/dispatcher.js

@@ -142,16 +142,21 @@ const DENY_ALL_WRITES_READONLY = Object.freeze([
 /* ------------------------------------------------------------------ */
 /* Default budgets                                                   */
 /* ------------------------------------------------------------------ */
+// CEO escalation 2026-06-05: 120K coder budget exhausted mid-React-
+// build (120214 > 120000). Match the engine-level `code` task bump
+// (apps/pugi-cli/src/core/engine/budgets.ts:149 — 400K). Subagent
+// dispatches inherit the upstream caller's headroom, so this needs
+// to track the engine envelope.
 const DEFAULT_BUDGETS = Object.freeze({
-    orchestrator: { tokens: 200_000, dollars: 5, wallClockMs: 600_000 },
-    architect: { tokens: 80_000, dollars: 2, wallClockMs: 300_000 },
-    coder: { tokens: 120_000, dollars: 3, wallClockMs: 600_000 },
-    verifier: { tokens: 60_000, dollars: 2, wallClockMs: 300_000 },
-    reviewer: { tokens: 80_000, dollars: 2, wallClockMs: 300_000 },
-    researcher: { tokens: 60_000, dollars: 1.5, wallClockMs: 300_000 },
-    release: { tokens: 40_000, dollars: 1, wallClockMs: 180_000 },
-    devops: { tokens: 60_000, dollars: 2, wallClockMs: 300_000 },
-    design_qa: { tokens: 60_000, dollars: 1.5, wallClockMs: 300_000 },
+    orchestrator: { tokens: 400_000, dollars: 8, wallClockMs: 900_000 },
+    architect: { tokens: 200_000, dollars: 4, wallClockMs: 600_000 },
+    coder: { tokens: 400_000, dollars: 8, wallClockMs: 900_000 },
+    verifier: { tokens: 150_000, dollars: 3, wallClockMs: 600_000 },
+    reviewer: { tokens: 200_000, dollars: 4, wallClockMs: 600_000 },
+    researcher: { tokens: 150_000, dollars: 3, wallClockMs: 600_000 },
+    release: { tokens: 80_000, dollars: 2, wallClockMs: 300_000 },
+    devops: { tokens: 150_000, dollars: 3, wallClockMs: 600_000 },
+    design_qa: { tokens: 150_000, dollars: 3, wallClockMs: 600_000 },
 });
 /**
  * Resolve the effective budget for a dispatch by merging task overrides

package/dist/runtime/version.js

@@ -44,7 +44,7 @@ export function sanitizeSemver(raw) {
  * during import). When bumping the CLI version BOTH literals must be
  * updated; the release smoke-test (`pack:smoke`) verifies they agree.
  */
-export const PUGI_CLI_VERSION = sanitizeSemver('0.1.0-beta.97');
+export const PUGI_CLI_VERSION = sanitizeSemver('0.1.0-beta.99');
 /**
  * Outbound: the CLI's installed semver. Read at request time by
  * `version-interceptor.ts` and injected on every `fetch` call.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@pugi/cli",
-  "version": "0.1.0-beta.97",
+  "version": "0.1.0-beta.99",
   "description": "Pugi CLI - terminal-native software execution system",
   "homepage": "https://pugi.io",
   "repository": {
@@ -62,8 +62,8 @@
     "undici": "^8.3.0",
     "which": "^6.0.0",
     "zod": "^3.23.0",
-    "@pugi/personas": "0.1.2",
-    "@pugi/sdk": "0.1.0-beta.97"
+    "@pugi/sdk": "0.1.0-beta.99",
+    "@pugi/personas": "0.1.2"
   },
   "devDependencies": {
     "@types/node": "^22.0.0",