zelari-code 2.19.0 → 2.21.0

package/dist/cli/main.bundled.js

@@ -29977,8 +29977,9 @@ var init_service = __esm({
         const text = (raw.query ?? raw.text ?? "").slice(0, 64e3);
         const limit = Math.max(1, Math.min(raw.limit ?? 8, 100));
         this.emit({ type: "memory_recall_start" });
+        const { weights, ...recallQuery } = raw;
         const candidates = await this.backend.search({
-          ...raw,
+          ...recallQuery,
           text,
           projectId: this.projectId,
           statuses: raw.statuses ?? (raw.includeHistorical ? void 0 : ["active"]),
@@ -30025,7 +30026,7 @@ var init_service = __esm({
             }
           }
         }
-        const ranked = diverse(rankMemoryCandidates([...byId.values()], text), limit);
+        const ranked = diverse(rankMemoryCandidates([...byId.values()], text, weights ? { weights } : {}), limit);
         this.emit({
           type: "memory_recall_end",
           durationMs: Date.now() - started,
@@ -40721,25 +40722,883 @@ var init_skillsMd = __esm({
   }
 });
 
+// src/cli/budget/historySummary.ts
+function extractiveHistorySummary(dropped, opts) {
+  const maxChars = opts?.maxChars ?? MAX_SUMMARY_CHARS;
+  if (dropped.length === 0) return "No prior turns.";
+  const userGoals = [];
+  const assistantNotes = [];
+  const userConstraints = [];
+  const unresolved = [];
+  const verification = [];
+  const decisions = [];
+  const tools = /* @__PURE__ */ new Map();
+  const files = /* @__PURE__ */ new Set();
+  let toolResults = 0;
+  for (const m of dropped) {
+    if (m.role === "user" && m.content.trim()) {
+      const goal = oneLine(m.content, 220);
+      userGoals.push(goal);
+      if (/\b(must|never|required|only|do not|constraint|vincolo|deve|senza)\b/i.test(goal)) userConstraints.push(goal);
+    } else if (m.role === "assistant") {
+      if (m.content.trim()) {
+        const note = oneLine(m.content, 220);
+        assistantNotes.push(note);
+        if (/\b(fail|failed|error|unresolved|remaining|todo|blocked|gap|errore|fallit|irrisolt|manca)\b/i.test(note)) unresolved.push(note);
+        if (/\b(test|typecheck|build|verify|verification|passed|failed|green|red)\b/i.test(note)) verification.push(note);
+        if (/\b(decid|decision|chosen|choose|scelt|adopt|implement)\b/i.test(note)) decisions.push(note);
+      }
+      if (m.toolCalls) {
+        for (const tc of m.toolCalls) {
+          tools.set(tc.name, (tools.get(tc.name) ?? 0) + 1);
+          collectPaths3(tc.args, files);
+        }
+      }
+    } else if (m.role === "tool") {
+      toolResults += 1;
+      collectPathsFromText(m.content, files);
+      if (/\b(fail|failed|error|exception|blocked|errore|fallit)\b/i.test(m.content)) unresolved.push(oneLine(m.content, 220));
+      if (/\b(test|typecheck|build|verify|passed|failed|success)\b/i.test(m.content)) verification.push(oneLine(m.content, 220));
+    }
+  }
+  const parts = [
+    "[history-summary] Earlier turns were compacted to stay within the context budget.",
+    `Dropped ${dropped.length} message(s) (${userGoals.length} user, ${assistantNotes.length} assistant notes, ${toolResults} tool results).`
+  ];
+  if (userGoals.length) {
+    parts.push("## User goals / requests");
+    for (const g of userGoals.slice(-6)) parts.push(`- ${g}`);
+  }
+  if (userConstraints.length) {
+    parts.push("## User constraints (preserve exactly)");
+    for (const item of [...new Set(userConstraints)].slice(-8)) parts.push(`- ${item}`);
+  }
+  if (unresolved.length) {
+    parts.push("## Unresolved failures / pending repair");
+    for (const item of [...new Set(unresolved)].slice(-8)) parts.push(`- ${item}`);
+  }
+  if (verification.length) {
+    parts.push("## Latest verification state");
+    for (const item of [...new Set(verification)].slice(-6)) parts.push(`- ${item}`);
+  }
+  if (decisions.length) {
+    parts.push("## Recent active decisions");
+    for (const item of [...new Set(decisions)].slice(-6)) parts.push(`- ${item}`);
+  }
+  if (assistantNotes.length) {
+    parts.push("## Assistant conclusions (truncated)");
+    for (const a of assistantNotes.slice(-5)) parts.push(`- ${a}`);
+  }
+  if (tools.size) {
+    const ranked = [...tools.entries()].sort((a, b) => b[1] - a[1]).slice(0, 12).map(([n, c]) => `${n}\xD7${c}`);
+    parts.push(`## Tools used: ${ranked.join(", ")}`);
+  }
+  if (files.size) {
+    const list = [...files].slice(0, 24);
+    parts.push(`## Paths mentioned: ${list.join(", ")}`);
+  }
+  parts.push(
+    "Continue from the recent messages below; do not re-ask goals already answered above unless the user changes them."
+  );
+  let out = parts.join("\n");
+  if (out.length > maxChars) {
+    out = `${out.slice(0, maxChars - 1)}\u2026`;
+  }
+  return out;
+}
+function oneLine(s, max) {
+  const t = s.replace(/\s+/g, " ").trim();
+  if (t.length <= max) return t;
+  return `${t.slice(0, max - 1)}\u2026`;
+}
+function collectPaths3(args, out) {
+  if (!args || typeof args !== "object") return;
+  const obj = args;
+  for (const key of ["path", "file", "filepath", "filePath", "target", "cwd"]) {
+    const v = obj[key];
+    if (typeof v === "string" && v.length > 1 && v.length < 260) {
+      out.add(v.replace(/\\/g, "/"));
+    }
+  }
+  if (typeof obj.file_path === "string") out.add(String(obj.file_path));
+}
+function collectPathsFromText(text, out) {
+  const re = /(?:^|[\s"'`])((?:[\w.-]+\/)+[\w.-]+\.\w{1,8})/g;
+  let m;
+  let n = 0;
+  while ((m = re.exec(text)) !== null && n < 8) {
+    out.add(m[1]);
+    n += 1;
+  }
+}
+var MAX_SUMMARY_CHARS;
+var init_historySummary = __esm({
+  "src/cli/budget/historySummary.ts"() {
+    "use strict";
+    MAX_SUMMARY_CHARS = 3500;
+  }
+});
+
+// src/cli/budget/llmCompact.ts
+function isLlmCompactEnabled() {
+  const v = process.env.ZELARI_LLM_COMPACT?.trim().toLowerCase();
+  if (v === "0" || v === "false" || v === "off" || v === "no") return false;
+  return true;
+}
+function compactModelOverride() {
+  const v = process.env.ZELARI_COMPACT_MODEL?.trim();
+  return v ? v : void 0;
+}
+async function llmSummarizeHistoryReplay(input) {
+  const override = input.overrideModel ?? compactModelOverride();
+  const model = override ?? input.model;
+  const cacheReuseExpected = !override;
+  if (!isLlmCompactEnabled()) return { summary: null, model, cacheReuseExpected };
+  if (input.droppedMessages.length === 0) {
+    return { summary: null, model, cacheReuseExpected };
+  }
+  const messages = [
+    ...input.systemMessages,
+    ...input.droppedMessages,
+    {
+      role: "user",
+      content: COMPACTION_INSTRUCTION
+    }
+  ];
+  const controller = new AbortController();
+  const timeout = setTimeout(() => controller.abort(), REPLAY_TIMEOUT_MS);
+  const onOuterAbort = () => controller.abort();
+  input.signal?.addEventListener("abort", onOuterAbort, { once: true });
+  try {
+    let text = "";
+    let emittedToolCall = false;
+    for await (const delta of input.providerStream({
+      provider: input.provider,
+      model,
+      messages,
+      // Tools stay advertised: dropping them would change the prefix token
+      // sequence and destroy cache reuse (explicit DSH decision). They are
+      // sorted canonically (same discipline as the live routed request and
+      // the snapshot fingerprints) so the replay prefix is byte-identical.
+      tools: [...input.tools].sort((a, b) => a.name < b.name ? -1 : a.name > b.name ? 1 : 0),
+      signal: controller.signal,
+      generation: {
+        purpose: "compaction",
+        temperature: 0.1,
+        maxTokens: 900
+      }
+    })) {
+      if (delta.kind === "text") text += delta.delta;
+      if (delta.kind === "tool_call") emittedToolCall = true;
+    }
+    if (emittedToolCall) return { summary: null, model, cacheReuseExpected };
+    if (!text.trim()) return { summary: null, model, cacheReuseExpected };
+    return { summary: text.trim(), model, cacheReuseExpected };
+  } catch {
+    return { summary: null, model, cacheReuseExpected };
+  } finally {
+    clearTimeout(timeout);
+    input.signal?.removeEventListener("abort", onOuterAbort);
+  }
+}
+var COMPACTION_INSTRUCTION, REPLAY_TIMEOUT_MS;
+var init_llmCompact = __esm({
+  "src/cli/budget/llmCompact.ts"() {
+    "use strict";
+    COMPACTION_INSTRUCTION = `
+You are now acting as a compaction engine for this coding-agent session.
+
+Condense the conversation ABOVE into a compact checkpoint sufficient to continue the task.
+
+Preserve:
+- user's goal and evolving intent
+- decisions already made
+- exact file paths and identifiers
+- code changes already completed
+- commands/errors that still matter
+- constraints
+- unfinished work
+- the single most likely next action
+
+Do not call tools.
+Do not mention this summarization request.
+Output only the checkpoint.
+Be concise.
+`.trim();
+    REPLAY_TIMEOUT_MS = 6e4;
+  }
+});
+
+// src/cli/hooks/historyCompaction.ts
+function compactedRangeFromDropped(dropped) {
+  if (dropped.length === 0) return void 0;
+  const seqs = [];
+  const sources = [];
+  for (const m of dropped) {
+    const hasCompactRange = typeof m.compactedFromSeq === "number" && Number.isInteger(m.compactedFromSeq) && m.compactedFromSeq > 0 && typeof m.compactedToSeq === "number" && Number.isInteger(m.compactedToSeq) && m.compactedToSeq >= m.compactedFromSeq;
+    if (hasCompactRange) {
+      seqs.push(m.compactedFromSeq, m.compactedToSeq);
+      sources.push(...m.sourceEventSeqs ?? []);
+      if (typeof m.seq === "number" && Number.isInteger(m.seq) && m.seq > 0) {
+        seqs.push(m.seq);
+        sources.push(m.seq);
+      }
+      continue;
+    }
+    if (typeof m.seq !== "number" || !Number.isInteger(m.seq) || m.seq < 1) return void 0;
+    seqs.push(m.seq);
+    sources.push(m.seq);
+  }
+  return {
+    fromSeq: Math.min(...seqs),
+    toSeq: Math.max(...seqs),
+    sourceEventSeqs: [...new Set(sources)]
+  };
+}
+function withDroppedRange(result, dropped, strategy) {
+  const range = compactedRangeFromDropped(dropped);
+  if (!range) return { ...result, strategy };
+  return { ...result, ...range, strategy };
+}
+function resolveMaxMessages(opts) {
+  const envTurns = envNumber(process.env.ZELARI_HISTORY_TURNS, { default: 6, min: 0 });
+  let turns = opts?.maxMessages ? Math.ceil(opts.maxMessages / 4) : envTurns;
+  if (opts?.durableStatePresent && !opts?.maxMessages && process.env.ZELARI_HISTORY_TURNS === void 0) {
+    turns = Math.min(turns, 3);
+  }
+  if (turns <= 0) return 0;
+  if (opts?.force && opts?.maxMessages) return Math.max(1, opts.maxMessages);
+  return turns * 4;
+}
+function findValidCutIndex(messages, naiveCut) {
+  let cut = naiveCut;
+  while (cut < messages.length) {
+    const kept = messages.slice(cut);
+    const declared = /* @__PURE__ */ new Set();
+    for (const m of kept) {
+      if (m.role === "assistant" && m.toolCalls) {
+        for (const tc of m.toolCalls) declared.add(tc.id);
+      }
+    }
+    let moved = false;
+    for (let k = 0; k < kept.length; k++) {
+      const m = kept[k];
+      if (m.role === "tool" && m.toolCallId && !declared.has(m.toolCallId)) {
+        for (let j = cut - 1; j >= 0; j--) {
+          const prev2 = messages[j];
+          if (prev2.role === "assistant" && prev2.toolCalls && prev2.toolCalls.some((tc) => tc.id === m.toolCallId)) {
+            cut = j;
+            moved = true;
+            break;
+          }
+        }
+        break;
+      }
+    }
+    if (!moved) break;
+  }
+  return cut;
+}
+function resolvePruneLimits(opts) {
+  const maxChars = opts?.maxChars ?? envNumber(process.env.ZELARI_TOOL_RESULT_MAX_CHARS, { default: 8e3, min: 256 });
+  const rawTail = opts?.tailChars ?? envNumber(process.env.ZELARI_TOOL_RESULT_TAIL_CHARS, { default: 1e3, min: 0 });
+  const tailChars = Math.min(rawTail, maxChars);
+  return { maxChars, tailChars };
+}
+function pruneToolResultsDetailed(messages, opts) {
+  const { maxChars, tailChars } = resolvePruneLimits(opts);
+  const headChars = maxChars - tailChars;
+  const stats = { pruned: 0, charsOmitted: 0 };
+  let changed = false;
+  const out = messages.map((m) => {
+    if (m.role !== "tool") return m;
+    const body = m.content ?? "";
+    if (body.length <= maxChars) return m;
+    const head = headChars > 0 ? body.slice(0, headChars) : "";
+    const tail2 = tailChars > 0 ? body.slice(-tailChars) : "";
+    const omitted = body.length - head.length - tail2.length;
+    changed = true;
+    stats.pruned += 1;
+    stats.charsOmitted += omitted;
+    return {
+      ...m,
+      content: [head, "\u2026[pruned " + omitted + " chars]\u2026", tail2].join(String.fromCharCode(10))
+    };
+  });
+  return {
+    messages: changed ? out : messages,
+    stats
+  };
+}
+function compactHistory(messages, opts) {
+  return compactHistoryDetailed(messages, opts).messages;
+}
+function buildCheckpointMessage(summaryText, range) {
+  return {
+    role: "user",
+    content: CHECKPOINT_WRAPPER_PREFIX + "\n\n<compacted-summary>\n" + summaryText + "\n</compacted-summary>",
+    ...range ? {
+      compactedFromSeq: range.fromSeq,
+      compactedToSeq: range.toSeq,
+      sourceEventSeqs: [...range.sourceEventSeqs]
+    } : {}
+  };
+}
+function compactHistoryDetailed(messages, opts) {
+  const maxMessages = resolveMaxMessages(opts);
+  if (maxMessages === 0) {
+    return withDroppedRange(
+      { messages: [], compacted: true, messagesRemoved: messages.length, summary: "" },
+      messages,
+      "extractive"
+    );
+  }
+  if (messages.length <= maxMessages * 2 && !opts?.force) {
+    return {
+      messages,
+      compacted: false,
+      messagesRemoved: 0,
+      summary: ""
+    };
+  }
+  const naiveCut = Math.max(0, messages.length - maxMessages);
+  const cut = findValidCutIndex(messages, naiveCut);
+  if (cut === 0) {
+    return {
+      messages,
+      compacted: false,
+      messagesRemoved: 0,
+      summary: ""
+    };
+  }
+  const droppedMsgs = messages.slice(0, cut);
+  const droppedRange = compactedRangeFromDropped(droppedMsgs);
+  const pruned = pruneToolResultsDetailed(messages.slice(cut));
+  const kept = pruned.messages;
+  const summaryText = extractiveHistorySummary(droppedMsgs);
+  const summary = buildCheckpointMessage(
+    summaryText || `${COMPACT_MARKER} ${cut} earlier message(s) dropped.`,
+    droppedRange
+  );
+  return withDroppedRange(
+    {
+      messages: [summary, ...kept],
+      compacted: true,
+      messagesRemoved: cut,
+      summary: summary.content,
+      prunedToolResults: pruned.stats.pruned
+    },
+    droppedMsgs,
+    "extractive"
+  );
+}
+async function compactHistoryAsync(messages, opts) {
+  const base2 = compactHistoryDetailed(messages, opts);
+  if (!base2.compacted || base2.messagesRemoved === 0) return base2;
+  const cut = base2.messagesRemoved;
+  const droppedMsgs = messages.slice(0, cut);
+  const extractive = extractiveHistorySummary(droppedMsgs);
+  let summaryText = extractive;
+  let cacheReuseExpected;
+  let replayExactPrefix;
+  const canReplay = !!(opts?.providerStream && opts?.requestSnapshot);
+  if (canReplay) {
+    try {
+      const replay = await llmSummarizeHistoryReplay({
+        providerStream: opts.providerStream,
+        provider: opts.requestSnapshot.provider,
+        model: opts.requestSnapshot.model,
+        systemMessages: opts.requestSnapshot.systemMessages,
+        tools: opts.requestSnapshot.tools,
+        droppedMessages: droppedMsgs,
+        signal: opts?.signal
+      });
+      cacheReuseExpected = replay.cacheReuseExpected;
+      if (replay.summary && replay.summary.trim().length > 40) {
+        const sourceTokens = roughTokens(droppedMsgs);
+        const summaryTok = Math.ceil(replay.summary.length / 4);
+        if (summaryTok < sourceTokens) {
+          summaryText = replay.summary.trim();
+        }
+      }
+    } catch {
+    }
+  }
+  const pruned = pruneToolResultsDetailed(messages.slice(cut));
+  const kept = pruned.messages;
+  const summary = buildCheckpointMessage(summaryText, compactedRangeFromDropped(droppedMsgs));
+  const usedLlm = summaryText !== extractive && summaryText.trim().length > 40;
+  return withDroppedRange(
+    {
+      messages: [summary, ...kept],
+      compacted: true,
+      messagesRemoved: cut,
+      summary: summaryText,
+      prunedToolResults: pruned.stats.pruned,
+      cacheReuseExpected,
+      replayExactPrefix
+    },
+    droppedMsgs,
+    usedLlm ? "llm" : "extractive"
+  );
+}
+function roughTokens(msgs) {
+  let n = 0;
+  for (const m of msgs) {
+    n += Math.ceil((m.content ?? "").length / 4);
+    if (m.toolCalls) {
+      for (const tc of m.toolCalls) {
+        n += Math.ceil(JSON.stringify(tc.args ?? {}).length / 4);
+      }
+    }
+  }
+  return Math.max(1, n);
+}
+var COMPACT_MARKER, CHECKPOINT_WRAPPER_PREFIX;
+var init_historyCompaction = __esm({
+  "src/cli/hooks/historyCompaction.ts"() {
+    "use strict";
+    init_historySummary();
+    init_llmCompact();
+    init_envNumber();
+    COMPACT_MARKER = "[history] Earlier turns were compacted to stay within the context budget.";
+    CHECKPOINT_WRAPPER_PREFIX = "This is an automatically generated checkpoint of earlier conversation. Treat it as established context and continue directly.";
+  }
+});
+
+// src/cli/provider/capabilities.ts
+function frozenProfile(input) {
+  if (input.reasoning.levels) Object.freeze(input.reasoning.levels);
+  Object.freeze(input.reasoning);
+  Object.freeze(input.promptCache);
+  Object.freeze(input.toolCalling);
+  Object.freeze(input.buildRecovery);
+  Object.freeze(input.sampling);
+  Object.freeze(input.compaction);
+  return Object.freeze(input);
+}
+function resolveHarnessProfile(model, providerId) {
+  const provider = providerId?.trim().toLowerCase();
+  if (provider === "deepseek") return "deepseek-v4";
+  if (provider === "grok") return "grok";
+  if (provider === "minimax") return "minimax";
+  if (provider === "glm") return "glm";
+  if (model && DEEPSEEK_RE.test(model)) return "deepseek-v4";
+  if (model && GROK_RE.test(model)) return "grok";
+  if (model && MINIMAX_RE.test(model)) return "minimax";
+  if (model && GLM_RE.test(model)) return "glm";
+  return "default";
+}
+function capabilitiesFor(model, providerId) {
+  switch (resolveHarnessProfile(model, providerId)) {
+    case "deepseek-v4":
+      return DEEPSEEK_V4_CAPS;
+    case "grok":
+      return GROK_CAPS;
+    case "minimax":
+      return model && MINIMAX_M3_RE.test(model) ? MINIMAX_M3_CAPS : MINIMAX_M2_CAPS;
+    case "glm":
+      return GLM_CAPS;
+    default:
+      return DEFAULT_CAPS;
+  }
+}
+var SHARED_COMPACTION, DEFAULT_CAPS, DEEPSEEK_V4_CAPS, GROK_CAPS, MINIMAX_M3_CAPS, MINIMAX_M2_CAPS, GLM_CAPS, DEEPSEEK_RE, GROK_RE, MINIMAX_RE, MINIMAX_M3_RE, GLM_RE;
+var init_capabilities = __esm({
+  "src/cli/provider/capabilities.ts"() {
+    "use strict";
+    SHARED_COMPACTION = { warnAt: 0.7, compactAt: 0.85, hardAt: 0.95 };
+    DEFAULT_CAPS = frozenProfile({
+      contextWindow: 4e5,
+      reasoning: { supported: true, levels: ["low", "medium", "high"], replayReasoning: true },
+      promptCache: { supported: true, pricedCacheRead: false },
+      toolCalling: { parallel: true },
+      buildRecovery: { forceToolChoice: false, maxForcedTurns: 0 },
+      sampling: { temperature: 0.7 },
+      compaction: { ...SHARED_COMPACTION },
+      profile: "default"
+    });
+    DEEPSEEK_V4_CAPS = frozenProfile({
+      contextWindow: 1e6,
+      reasoning: { supported: true, levels: ["high", "max"], replayReasoning: true },
+      promptCache: { supported: true, pricedCacheRead: true },
+      toolCalling: { parallel: true },
+      buildRecovery: { forceToolChoice: false, maxForcedTurns: 0 },
+      sampling: { temperature: 0.7 },
+      compaction: { ...SHARED_COMPACTION },
+      profile: "deepseek-v4"
+    });
+    GROK_CAPS = frozenProfile({
+      contextWindow: 5e5,
+      // grok-4.x truncation fix: explicit conversation ceiling (see interface
+      // doc). Override at runtime via ZELARI_MAX_OUTPUT_TOKENS.
+      maxOutputTokens: 32768,
+      reasoning: {
+        supported: true,
+        levels: ["low", "medium", "high", "xhigh"],
+        replayReasoning: false
+      },
+      promptCache: {
+        supported: true,
+        pricedCacheRead: true,
+        conversationAffinityHeader: "x-grok-conv-id"
+      },
+      toolCalling: { parallel: true },
+      buildRecovery: { forceToolChoice: true, maxForcedTurns: 1 },
+      sampling: { temperature: 0.7 },
+      compaction: { ...SHARED_COMPACTION },
+      profile: "grok"
+    });
+    MINIMAX_M3_CAPS = frozenProfile({
+      contextWindow: 1e6,
+      reasoning: { supported: true, replayReasoning: true },
+      promptCache: { supported: false, pricedCacheRead: false },
+      toolCalling: { parallel: true },
+      buildRecovery: { forceToolChoice: false, maxForcedTurns: 0 },
+      sampling: { temperature: 0.7 },
+      compaction: { ...SHARED_COMPACTION },
+      profile: "minimax"
+    });
+    MINIMAX_M2_CAPS = frozenProfile({
+      contextWindow: 204800,
+      reasoning: { supported: true, replayReasoning: true },
+      promptCache: { supported: false, pricedCacheRead: false },
+      toolCalling: { parallel: true },
+      buildRecovery: { forceToolChoice: false, maxForcedTurns: 0 },
+      sampling: { temperature: 0.7 },
+      compaction: { ...SHARED_COMPACTION },
+      profile: "minimax"
+    });
+    GLM_CAPS = frozenProfile({
+      contextWindow: 2e5,
+      reasoning: { supported: true, replayReasoning: true },
+      promptCache: { supported: true, pricedCacheRead: false },
+      toolCalling: { parallel: true },
+      buildRecovery: { forceToolChoice: false, maxForcedTurns: 0 },
+      sampling: { temperature: 0.7 },
+      compaction: { ...SHARED_COMPACTION },
+      profile: "glm"
+    });
+    DEEPSEEK_RE = /^deepseek-(?:v4(?:\.|-|$)|chat$|reasoner$)/i;
+    GROK_RE = /^grok(?:\.|-|$)/i;
+    MINIMAX_RE = /^minimax(?:\.|-|$)/i;
+    MINIMAX_M3_RE = /^minimax-m3(?:\.|-|$)/i;
+    GLM_RE = /^glm(?:\.|-|$)/i;
+  }
+});
+
+// src/cli/budget/tokenBudget.ts
+function mergeCompactRange(into, r) {
+  if (r.fromSeq !== void 0 && r.toSeq !== void 0) {
+    into.fromSeq = into.fromSeq === void 0 ? r.fromSeq : Math.min(into.fromSeq, r.fromSeq);
+    into.toSeq = into.toSeq === void 0 ? r.toSeq : Math.max(into.toSeq, r.toSeq);
+    if (r.sourceEventSeqs) into.sourceSeqs.push(...r.sourceEventSeqs);
+  }
+  if (r.strategy === "llm") into.strategy = "llm";
+  else if (r.strategy && !into.strategy) into.strategy = r.strategy;
+}
+function estimateTokens2(text) {
+  if (!text) return 0;
+  return Math.max(1, Math.ceil(text.length / 4));
+}
+function estimateHistoryTokens(messages) {
+  let n = 0;
+  for (const m of messages) {
+    n += estimateTokens2(m.content);
+    if (m.toolCalls) {
+      for (const tc of m.toolCalls) {
+        n += estimateTokens2(tc.name) + estimateTokens2(JSON.stringify(tc.args ?? {}));
+      }
+    }
+  }
+  return n;
+}
+function defaultContextLimitForModel(model, provider) {
+  return capabilitiesFor(model, provider).contextWindow;
+}
+function resolveContextLimit(model, provider) {
+  return envNumber(process.env.ZELARI_CONTEXT_LIMIT, {
+    default: defaultContextLimitForModel(model, provider),
+    min: 4e3,
+    max: 2e6
+  });
+}
+function phaseKnobs(phase2) {
+  return {
+    historyTurns: phase2 === "plan" ? envNumber(process.env.ZELARI_HISTORY_TURNS, { default: 8, min: 0 }) : envNumber(process.env.ZELARI_HISTORY_TURNS, { default: 6, min: 0 }),
+    maxToolLoopIterations: phase2 === "plan" ? envNumber(process.env.ZELARI_MAX_TOOL_LOOP_ITERATIONS, { default: 60, min: 1 }) : envNumber(process.env.ZELARI_MAX_TOOL_LOOP_ITERATIONS, {
+      default: 120,
+      min: 1
+    })
+  };
+}
+async function applyBudgetPolicyAsync(history2, phase2, opts) {
+  const contextLimit = resolveContextLimit(opts?.model, opts?.provider);
+  const compact3 = capabilitiesFor(opts?.model, opts?.provider).compaction;
+  const sessionExtra = opts?.sessionTokens ?? 0;
+  const warnings = [];
+  let { historyTurns, maxToolLoopIterations } = phaseKnobs(phase2);
+  const envelope = opts?.requestSnapshot ?? null;
+  const surface = envelope?.snapshot ?? opts?.requestSurface ?? null;
+  const replayBase = surface ? {
+    provider: surface.provider,
+    model: surface.model,
+    systemMessages: surface.systemMessages,
+    tools: surface.tools
+  } : opts?.providerStream ? { provider: "local", model: opts?.model ?? "unknown", systemMessages: [], tools: [] } : null;
+  const headerTokens = surface ? estimateSystemTokensLite(surface.systemMessages) + estimateToolSchemaTokensLite(surface.tools) : 0;
+  const convTokensOf = (h) => estimateConversationTokensLite(h);
+  let hist = history2;
+  let estimated = surface ? headerTokens + convTokensOf(hist) + RESERVED_OUTPUT_TOKENS : estimateHistoryTokens(hist) + sessionExtra;
+  let occupancy = Math.min(1, estimated / contextLimit);
+  let compactSummary = "";
+  let messagesRemoved = 0;
+  let cacheReuseExpected;
+  let prunedTotal = 0;
+  const compactRange = { sourceSeqs: [] };
+  if (occupancy >= compact3.warnAt && occupancy < compact3.compactAt) {
+    warnings.push(
+      `[budget] context ~${Math.round(occupancy * 100)}% full (${estimated}/${contextLimit} tok full-request est.) \u2014 consider /compact or shorter replies.`
+    );
+  }
+  if (occupancy >= 0.8) {
+    const pruned = pruneToolResultsDetailed(hist);
+    if (pruned.stats.pruned > 0) {
+      hist = pruned.messages;
+      prunedTotal += pruned.stats.pruned;
+      estimated = surface ? headerTokens + convTokensOf(hist) + RESERVED_OUTPUT_TOKENS : estimateHistoryTokens(hist) + sessionExtra;
+      occupancy = Math.min(1, estimated / contextLimit);
+      warnings.push(
+        `[budget] pruned ${pruned.stats.pruned} oversized tool result(s) \u2192 ${Math.round(occupancy * 100)}% (${estimated} tok).`
+      );
+    }
+  }
+  const fold = (r, label, forcedTurns) => {
+    hist = applySessionSurface(r.messages);
+    if (r.compacted) {
+      messagesRemoved += r.messagesRemoved;
+      if (r.summary) compactSummary = r.summary;
+      mergeCompactRange(compactRange, r);
+      if (r.cacheReuseExpected !== void 0) {
+        cacheReuseExpected = r.cacheReuseExpected;
+      }
+    }
+    estimated = surface ? headerTokens + convTokensOf(hist) + RESERVED_OUTPUT_TOKENS : estimateHistoryTokens(hist) + sessionExtra;
+    occupancy = Math.min(1, estimated / contextLimit);
+    warnings.push(
+      `[budget] ${label} \u2014 kept ~${forcedTurns} turns (${estimated} tok est.` + (r.messagesRemoved ? `, removed ${r.messagesRemoved} msgs` : "") + (r.cacheReuseExpected === false ? ", cache reuse NOT expected (model override)" : "") + ")."
+    );
+  };
+  if (occupancy >= compact3.compactAt) {
+    const forcedTurns = Math.max(1, Math.floor(historyTurns / 2));
+    historyTurns = forcedTurns;
+    maxToolLoopIterations = Math.min(
+      maxToolLoopIterations,
+      phase2 === "plan" ? 24 : 40
+    );
+    let r = await compactHistoryAsync(hist, {
+      maxMessages: Math.max(2, forcedTurns * 4),
+      force: true,
+      signal: opts?.signal,
+      ...replayBase ? { requestSnapshot: replayBase, providerStream: opts?.providerStream } : {}
+    });
+    if (!r.compacted) {
+      r = await compactHistoryAsync(hist, {
+        maxMessages: 2,
+        force: true,
+        signal: opts?.signal,
+        ...replayBase ? { requestSnapshot: replayBase, providerStream: opts?.providerStream } : {}
+      });
+    }
+    const label = r.cacheReuseExpected === false ? "llm-compact (model override)" : "auto-compact";
+    fold(r, label, forcedTurns);
+  }
+  if (occupancy >= compact3.hardAt) {
+    const hard = await compactHistoryAsync(hist, {
+      maxMessages: 2,
+      force: true,
+      signal: opts?.signal
+    });
+    fold(hard, hard.summary.includes("\xB7 llm") ? "llm-compact" : "HARD trim", 2);
+    historyTurns = 2;
+    maxToolLoopIterations = Math.min(maxToolLoopIterations, 16);
+    warnings.push(
+      "[budget] HARD context pressure (\u226595%) \u2014 prefer /clear or a new session if quality drops."
+    );
+  }
+  const cacheMetricsLine = envelope ? [
+    "compaction meter:",
+    `provider/model: ${envelope.snapshot.provider}/${envelope.snapshot.model}`,
+    `headerFingerprint: ${envelope.snapshot.headerFingerprint.slice(0, 12)}`,
+    `occupancy: ${Math.round(occupancy * 100)}% (${estimated}/${contextLimit})`,
+    ...envelope.usage?.cachedPromptTokens !== void 0 ? [`cachedPromptTokens: ${envelope.usage.cachedPromptTokens}`] : [],
+    ...cacheReuseExpected !== void 0 ? [`cacheReuseExpected: ${cacheReuseExpected}`] : []
+  ].join(" | ") : void 0;
+  return {
+    history: hist,
+    warnings,
+    maxToolLoopIterations,
+    historyTurns,
+    estimatedHistoryTokens: surface ? convTokensOf(hist) : estimated,
+    contextLimit,
+    occupancy,
+    compactSummary: compactSummary || void 0,
+    messagesRemoved: messagesRemoved || void 0,
+    ...compactRange.fromSeq !== void 0 && compactRange.toSeq !== void 0 ? {
+      compactedFromSeq: compactRange.fromSeq,
+      compactedToSeq: compactRange.toSeq,
+      compactSourceSeqs: compactRange.sourceSeqs,
+      compactStrategy: compactRange.strategy
+    } : {},
+    ...surface ? { contextPressureTokens: estimated } : {},
+    ...cacheReuseExpected !== void 0 ? { cacheReuseExpected } : {},
+    ...cacheMetricsLine ? { cacheMetricsLine } : {}
+  };
+}
+function estimateSystemTokensLite(systemMessages) {
+  let n = 0;
+  for (const m of systemMessages) n += 4 + Math.ceil((m.content ?? "").length / 4);
+  return n;
+}
+function estimateToolSchemaTokensLite(tools) {
+  let n = 0;
+  for (const t of tools) {
+    n += Math.ceil((t.name ?? "").length / 4);
+    n += Math.ceil((t.description ?? "").length / 4);
+    n += Math.ceil(JSON.stringify(t.parameters ?? {}).length / 4);
+  }
+  return n + tools.length * 4;
+}
+function estimateConversationTokensLite(messages) {
+  let n = 0;
+  for (const m of messages) {
+    n += 4 + Math.ceil((m.content ?? "").length / 4);
+    if (m.toolCalls) {
+      for (const tc of m.toolCalls) {
+        n += Math.ceil(tc.name.length / 4) + Math.ceil(tc.id.length / 4);
+        n += Math.ceil(JSON.stringify(tc.args ?? {}).length / 4);
+      }
+    }
+    if (m.reasoningContent) n += Math.ceil(m.reasoningContent.length / 4);
+    if (m.toolCallId) n += Math.ceil(m.toolCallId.length / 4);
+  }
+  return n;
+}
+var RESERVED_OUTPUT_TOKENS;
+var init_tokenBudget = __esm({
+  "src/cli/budget/tokenBudget.ts"() {
+    "use strict";
+    init_envNumber();
+    init_historyCompaction();
+    init_observationStore();
+    init_capabilities();
+    RESERVED_OUTPUT_TOKENS = 8192;
+  }
+});
+
+// src/cli/budget/retrievalPolicy.ts
+var retrievalPolicy_exports = {};
+__export(retrievalPolicy_exports, {
+  BASELINE_RETRIEVAL_POLICY: () => BASELINE_RETRIEVAL_POLICY,
+  RETRIEVAL_PRESSURE_THRESHOLDS: () => RETRIEVAL_PRESSURE_THRESHOLDS,
+  estimateHistoryOccupancy: () => estimateHistoryOccupancy,
+  filterSkillCatalogByBand: () => filterSkillCatalogByBand,
+  resolveRetrievalPolicy: () => resolveRetrievalPolicy,
+  retrievalBand: () => retrievalBand
+});
+function retrievalBand(occupancy) {
+  const occ = Number.isFinite(occupancy) ? Math.max(0, Math.min(1, occupancy)) : 0;
+  if (occ >= RETRIEVAL_PRESSURE_THRESHOLDS.high) return "high";
+  if (occ >= RETRIEVAL_PRESSURE_THRESHOLDS.medium) return "medium";
+  return "low";
+}
+function resolveRetrievalPolicy(occupancy) {
+  const band = retrievalBand(occupancy);
+  const packing = PRESSURE_PACKING[band];
+  const weights = PRESSURE_WEIGHTS[band];
+  return {
+    band,
+    ...packing,
+    ...weights ? { weights } : {}
+  };
+}
+function estimateHistoryOccupancy(history2, opts) {
+  const limit = resolveContextLimit(opts?.model, opts?.provider);
+  return Math.min(1, estimateHistoryTokens(history2) / limit);
+}
+function filterSkillCatalogByBand(skills, band) {
+  if (band !== "high") return [...skills];
+  return skills.filter((s) => s.estimatedCost !== "high");
+}
+var RETRIEVAL_PRESSURE_THRESHOLDS, BASELINE_RETRIEVAL_POLICY, PRESSURE_WEIGHTS, PRESSURE_PACKING;
+var init_retrievalPolicy = __esm({
+  "src/cli/budget/retrievalPolicy.ts"() {
+    "use strict";
+    init_tokenBudget();
+    RETRIEVAL_PRESSURE_THRESHOLDS = { medium: 0.5, high: 0.8 };
+    BASELINE_RETRIEVAL_POLICY = {
+      band: "low",
+      maxChars: 2e3,
+      maxMemories: 8
+    };
+    PRESSURE_WEIGHTS = {
+      low: void 0,
+      medium: {
+        semanticRelevance: 0.35,
+        lexicalRelevance: 0.2,
+        importance: 0.15,
+        confidence: 0.1,
+        recency: 0.05,
+        graphProximity: 0.1,
+        verificationBonus: 0.05
+      },
+      high: {
+        semanticRelevance: 0.4,
+        lexicalRelevance: 0.25,
+        importance: 0.15,
+        confidence: 0.1,
+        recency: 0,
+        graphProximity: 0.05,
+        verificationBonus: 0.05
+      }
+    };
+    PRESSURE_PACKING = {
+      low: { maxChars: 2e3, maxMemories: 8 },
+      medium: { maxChars: 1200, maxMemories: 6 },
+      high: { maxChars: 600, maxMemories: 4 }
+    };
+  }
+});
+
 // src/cli/tools/skillTool.ts
-function formatAvailableSkillsCatalog(cwd = process.cwd()) {
+function formatAvailableSkillsCatalog(cwd = process.cwd(), opts) {
   try {
     loadSkillMdSkills(cwd);
   } catch {
   }
-  const skills = listCodingSkills();
-  if (skills.length === 0) {
+  const band = opts?.pressureBand;
+  const all = listCodingSkills();
+  if (all.length === 0) {
     return "No skills registered.";
   }
-  const lines = skills.slice(0, 80).map((s) => {
+  const gated = band ? filterSkillCatalogByBand(all, band) : all;
+  const hidden = all.length - gated.length;
+  const lines = gated.slice(0, 80).map((s) => {
     const desc = (s.description || s.id).replace(/\s+/g, " ").trim().slice(0, 160);
     return `- ${s.id}: ${desc}`;
   });
+  if (hidden > 0) {
+    lines.push(
+      `(plus ${hidden} high-cost skill(s) hidden under context pressure \u2014 still loadable by name if truly needed)`
+    );
+  }
   return lines.join("\n");
 }
 function createSkillTool(opts) {
   const cwd = opts?.cwd ?? process.cwd();
-  const catalog = formatAvailableSkillsCatalog(cwd);
+  const catalog = formatAvailableSkillsCatalog(cwd, {
+    pressureBand: opts?.pressureBand
+  });
   return {
     name: "skill",
     description: "Load the full instructions for a named coding skill into this turn. Call when a skill matches the current task. Available skills:\n" + catalog,
@@ -40781,6 +41640,7 @@ var init_skillTool = __esm({
     init_skills2();
     init_toolTypes();
     init_skillsMd();
+    init_retrievalPolicy();
     SkillArgsSchema = external_exports.object({
       name: external_exports.string().min(1).describe("Skill id/name to load (from the available skills list).")
     });
@@ -43862,124 +44722,6 @@ var init_semantic2 = __esm({
   }
 });
 
-// src/cli/provider/capabilities.ts
-function frozenProfile(input) {
-  if (input.reasoning.levels) Object.freeze(input.reasoning.levels);
-  Object.freeze(input.reasoning);
-  Object.freeze(input.promptCache);
-  Object.freeze(input.toolCalling);
-  Object.freeze(input.buildRecovery);
-  Object.freeze(input.sampling);
-  Object.freeze(input.compaction);
-  return Object.freeze(input);
-}
-function resolveHarnessProfile(model, providerId) {
-  const provider = providerId?.trim().toLowerCase();
-  if (provider === "deepseek") return "deepseek-v4";
-  if (provider === "grok") return "grok";
-  if (provider === "minimax") return "minimax";
-  if (provider === "glm") return "glm";
-  if (model && DEEPSEEK_RE.test(model)) return "deepseek-v4";
-  if (model && GROK_RE.test(model)) return "grok";
-  if (model && MINIMAX_RE.test(model)) return "minimax";
-  if (model && GLM_RE.test(model)) return "glm";
-  return "default";
-}
-function capabilitiesFor(model, providerId) {
-  switch (resolveHarnessProfile(model, providerId)) {
-    case "deepseek-v4":
-      return DEEPSEEK_V4_CAPS;
-    case "grok":
-      return GROK_CAPS;
-    case "minimax":
-      return model && MINIMAX_M3_RE.test(model) ? MINIMAX_M3_CAPS : MINIMAX_M2_CAPS;
-    case "glm":
-      return GLM_CAPS;
-    default:
-      return DEFAULT_CAPS;
-  }
-}
-var SHARED_COMPACTION, DEFAULT_CAPS, DEEPSEEK_V4_CAPS, GROK_CAPS, MINIMAX_M3_CAPS, MINIMAX_M2_CAPS, GLM_CAPS, DEEPSEEK_RE, GROK_RE, MINIMAX_RE, MINIMAX_M3_RE, GLM_RE;
-var init_capabilities = __esm({
-  "src/cli/provider/capabilities.ts"() {
-    "use strict";
-    SHARED_COMPACTION = { warnAt: 0.7, compactAt: 0.85, hardAt: 0.95 };
-    DEFAULT_CAPS = frozenProfile({
-      contextWindow: 4e5,
-      reasoning: { supported: true, levels: ["low", "medium", "high"], replayReasoning: true },
-      promptCache: { supported: true, pricedCacheRead: false },
-      toolCalling: { parallel: true },
-      buildRecovery: { forceToolChoice: false, maxForcedTurns: 0 },
-      sampling: { temperature: 0.7 },
-      compaction: { ...SHARED_COMPACTION },
-      profile: "default"
-    });
-    DEEPSEEK_V4_CAPS = frozenProfile({
-      contextWindow: 1e6,
-      reasoning: { supported: true, levels: ["high", "max"], replayReasoning: true },
-      promptCache: { supported: true, pricedCacheRead: true },
-      toolCalling: { parallel: true },
-      buildRecovery: { forceToolChoice: false, maxForcedTurns: 0 },
-      sampling: { temperature: 0.7 },
-      compaction: { ...SHARED_COMPACTION },
-      profile: "deepseek-v4"
-    });
-    GROK_CAPS = frozenProfile({
-      contextWindow: 5e5,
-      reasoning: {
-        supported: true,
-        levels: ["low", "medium", "high", "xhigh"],
-        replayReasoning: false
-      },
-      promptCache: {
-        supported: true,
-        pricedCacheRead: true,
-        conversationAffinityHeader: "x-grok-conv-id"
-      },
-      toolCalling: { parallel: true },
-      buildRecovery: { forceToolChoice: true, maxForcedTurns: 1 },
-      sampling: { temperature: 0.7 },
-      compaction: { ...SHARED_COMPACTION },
-      profile: "grok"
-    });
-    MINIMAX_M3_CAPS = frozenProfile({
-      contextWindow: 1e6,
-      reasoning: { supported: true, replayReasoning: true },
-      promptCache: { supported: false, pricedCacheRead: false },
-      toolCalling: { parallel: true },
-      buildRecovery: { forceToolChoice: false, maxForcedTurns: 0 },
-      sampling: { temperature: 0.7 },
-      compaction: { ...SHARED_COMPACTION },
-      profile: "minimax"
-    });
-    MINIMAX_M2_CAPS = frozenProfile({
-      contextWindow: 204800,
-      reasoning: { supported: true, replayReasoning: true },
-      promptCache: { supported: false, pricedCacheRead: false },
-      toolCalling: { parallel: true },
-      buildRecovery: { forceToolChoice: false, maxForcedTurns: 0 },
-      sampling: { temperature: 0.7 },
-      compaction: { ...SHARED_COMPACTION },
-      profile: "minimax"
-    });
-    GLM_CAPS = frozenProfile({
-      contextWindow: 2e5,
-      reasoning: { supported: true, replayReasoning: true },
-      promptCache: { supported: true, pricedCacheRead: false },
-      toolCalling: { parallel: true },
-      buildRecovery: { forceToolChoice: false, maxForcedTurns: 0 },
-      sampling: { temperature: 0.7 },
-      compaction: { ...SHARED_COMPACTION },
-      profile: "glm"
-    });
-    DEEPSEEK_RE = /^deepseek-(?:v4(?:\.|-|$)|chat$|reasoner$)/i;
-    GROK_RE = /^grok(?:\.|-|$)/i;
-    MINIMAX_RE = /^minimax(?:\.|-|$)/i;
-    MINIMAX_M3_RE = /^minimax-m3(?:\.|-|$)/i;
-    GLM_RE = /^glm(?:\.|-|$)/i;
-  }
-});
-
 // src/cli/provider/openai-compatible.ts
 var openai_compatible_exports = {};
 __export(openai_compatible_exports, {
@@ -44163,6 +44905,12 @@ ${notes}
   }
   return { role: m.role, content: m.content };
 }
+function positiveEnvInt(name) {
+  const raw = process.env[name];
+  if (!raw) return void 0;
+  const n = Number.parseInt(raw, 10);
+  return Number.isFinite(n) && n > 0 ? n : void 0;
+}
 function openaiCompatibleProvider(config2) {
   return async function* (params) {
     const capabilities = capabilitiesFor(params.model, config2.providerId);
@@ -44195,6 +44943,8 @@ function openaiCompatibleProvider(config2) {
     };
     if (typeof generation?.maxTokens === "number" && generation.maxTokens > 0) {
       body.max_tokens = generation.maxTokens;
+    } else if (typeof capabilities.maxOutputTokens === "number" && capabilities.maxOutputTokens > 0) {
+      body.max_tokens = positiveEnvInt("ZELARI_MAX_OUTPUT_TOKENS") ?? capabilities.maxOutputTokens;
     }
     const thinkingSpec = config2.thinking ?? "auto";
     if (config2.providerId === "deepseek" && thinkingSpec === "auto") {
@@ -46386,8 +47136,8 @@ function defaultPermissionPolicy(overrides) {
   return {
     read: parseAction(process.env.ZELARI_PERMISSION_READ, "allow"),
     write: parseAction(process.env.ZELARI_PERMISSION_WRITE, "allow"),
-    execute: parseAction(process.env.ZELARI_PERMISSION_EXECUTE, "allow"),
-    network: parseAction(process.env.ZELARI_PERMISSION_NETWORK, "allow"),
+    execute: parseAction(process.env.ZELARI_PERMISSION_EXECUTE, "ask"),
+    network: parseAction(process.env.ZELARI_PERMISSION_NETWORK, "ask"),
     ui: "allow",
     auto: isAutoPermissions(),
     ...overrides
@@ -47353,7 +48103,7 @@ function matchResourceClaimLayered(layers, precedence, claim, root) {
 function resolveClaimsVerdict(layers, precedence, toolName, args, root) {
   const claims = resourceClaimsFor(toolName, args);
   const matchedRules = [];
-  if (!layers || claims.length === 0) return { matchedRules };
+  if (!layers || claims.length === 0) return { claims, matchedRules };
   const effects = [];
   for (const claim of claims) {
     const hit = matchResourceClaimLayered(layers, precedence, claim, root);
@@ -47362,7 +48112,26 @@ function resolveClaimsVerdict(layers, precedence, toolName, args, root) {
       matchedRules.push(hit);
     }
   }
-  return effects.length > 0 ? { effect: intersectEffects(...effects), matchedRules } : { matchedRules };
+  return effects.length > 0 ? { effect: intersectEffects(...effects), claims, matchedRules } : { claims, matchedRules };
+}
+function describeResourceClaim(claim) {
+  const cap3 = (s) => s.length > 160 ? `${s.slice(0, 157)}\u2026` : s;
+  switch (claim.kind) {
+    case "path":
+      return cap3(`path ${claim.operation}: ${claim.path}`);
+    case "process":
+      return cap3(`process: ${[claim.executable, ...claim.argv ?? []].join(" ")}`);
+    case "network":
+      return cap3(`network: ${claim.host}${claim.port !== void 0 ? `:${claim.port}` : ""}`);
+    case "mcp":
+      return cap3(`mcp: ${claim.server}/${claim.tool}`);
+    case "ssh":
+      return cap3(`ssh: ${claim.target}${claim.command ? ` (${claim.command})` : ""}`);
+    case "ui":
+      return cap3(`ui: ${claim.action}`);
+    case "agent":
+      return cap3(`agent: ${claim.role}`);
+  }
 }
 var MAX_RAW_SHELL_STRIP_DEPTH, RAW_SHELL_WRAPPERS, RAW_SHELL_INTERPRETERS;
 var init_resourceClaims = __esm({
@@ -47819,7 +48588,7 @@ function createBuiltinToolRegistry(options = {}) {
     registry4.register(withPerm(askUserTool));
   }
   const enableSkill = options.enableSkill !== false && options.readOnly !== true && !gauntletParent && profile !== "explore" && profile !== "verify";
-  const skillTool = enableSkill ? withPerm(createSkillTool({ cwd: root })) : null;
+  const skillTool = enableSkill ? withPerm(createSkillTool({ cwd: root, pressureBand: options.pressureBand })) : null;
   if (skillTool) {
     registry4.register(skillTool);
   }
@@ -48113,11 +48882,17 @@ function wrapWithPermissions(original, policy, onAsk, agentLayers, precedence =
           );
         }
         try {
+          const expanded = claims?.claims ?? resourceClaimsFor(original.name, input ?? {});
           const ok = await onAsk({
             toolName: original.name,
             reason: decision.reason,
             categories: decision.categories,
-            args: input
+            args: input,
+            policyNote: rulePrefix || void 0,
+            claims: expanded.slice(0, 6).map((c) => ({
+              kind: c.kind,
+              summary: describeResourceClaim(c)
+            }))
           });
           if (!ok) {
             return typedErr(
@@ -50500,450 +51275,6 @@ var init_messageHelpers = __esm({
   }
 });
 
-// src/cli/budget/historySummary.ts
-function extractiveHistorySummary(dropped, opts) {
-  const maxChars = opts?.maxChars ?? MAX_SUMMARY_CHARS;
-  if (dropped.length === 0) return "No prior turns.";
-  const userGoals = [];
-  const assistantNotes = [];
-  const userConstraints = [];
-  const unresolved = [];
-  const verification = [];
-  const decisions = [];
-  const tools = /* @__PURE__ */ new Map();
-  const files = /* @__PURE__ */ new Set();
-  let toolResults = 0;
-  for (const m of dropped) {
-    if (m.role === "user" && m.content.trim()) {
-      const goal = oneLine(m.content, 220);
-      userGoals.push(goal);
-      if (/\b(must|never|required|only|do not|constraint|vincolo|deve|senza)\b/i.test(goal)) userConstraints.push(goal);
-    } else if (m.role === "assistant") {
-      if (m.content.trim()) {
-        const note = oneLine(m.content, 220);
-        assistantNotes.push(note);
-        if (/\b(fail|failed|error|unresolved|remaining|todo|blocked|gap|errore|fallit|irrisolt|manca)\b/i.test(note)) unresolved.push(note);
-        if (/\b(test|typecheck|build|verify|verification|passed|failed|green|red)\b/i.test(note)) verification.push(note);
-        if (/\b(decid|decision|chosen|choose|scelt|adopt|implement)\b/i.test(note)) decisions.push(note);
-      }
-      if (m.toolCalls) {
-        for (const tc of m.toolCalls) {
-          tools.set(tc.name, (tools.get(tc.name) ?? 0) + 1);
-          collectPaths3(tc.args, files);
-        }
-      }
-    } else if (m.role === "tool") {
-      toolResults += 1;
-      collectPathsFromText(m.content, files);
-      if (/\b(fail|failed|error|exception|blocked|errore|fallit)\b/i.test(m.content)) unresolved.push(oneLine(m.content, 220));
-      if (/\b(test|typecheck|build|verify|passed|failed|success)\b/i.test(m.content)) verification.push(oneLine(m.content, 220));
-    }
-  }
-  const parts = [
-    "[history-summary] Earlier turns were compacted to stay within the context budget.",
-    `Dropped ${dropped.length} message(s) (${userGoals.length} user, ${assistantNotes.length} assistant notes, ${toolResults} tool results).`
-  ];
-  if (userGoals.length) {
-    parts.push("## User goals / requests");
-    for (const g of userGoals.slice(-6)) parts.push(`- ${g}`);
-  }
-  if (userConstraints.length) {
-    parts.push("## User constraints (preserve exactly)");
-    for (const item of [...new Set(userConstraints)].slice(-8)) parts.push(`- ${item}`);
-  }
-  if (unresolved.length) {
-    parts.push("## Unresolved failures / pending repair");
-    for (const item of [...new Set(unresolved)].slice(-8)) parts.push(`- ${item}`);
-  }
-  if (verification.length) {
-    parts.push("## Latest verification state");
-    for (const item of [...new Set(verification)].slice(-6)) parts.push(`- ${item}`);
-  }
-  if (decisions.length) {
-    parts.push("## Recent active decisions");
-    for (const item of [...new Set(decisions)].slice(-6)) parts.push(`- ${item}`);
-  }
-  if (assistantNotes.length) {
-    parts.push("## Assistant conclusions (truncated)");
-    for (const a of assistantNotes.slice(-5)) parts.push(`- ${a}`);
-  }
-  if (tools.size) {
-    const ranked = [...tools.entries()].sort((a, b) => b[1] - a[1]).slice(0, 12).map(([n, c]) => `${n}\xD7${c}`);
-    parts.push(`## Tools used: ${ranked.join(", ")}`);
-  }
-  if (files.size) {
-    const list = [...files].slice(0, 24);
-    parts.push(`## Paths mentioned: ${list.join(", ")}`);
-  }
-  parts.push(
-    "Continue from the recent messages below; do not re-ask goals already answered above unless the user changes them."
-  );
-  let out = parts.join("\n");
-  if (out.length > maxChars) {
-    out = `${out.slice(0, maxChars - 1)}\u2026`;
-  }
-  return out;
-}
-function oneLine(s, max) {
-  const t = s.replace(/\s+/g, " ").trim();
-  if (t.length <= max) return t;
-  return `${t.slice(0, max - 1)}\u2026`;
-}
-function collectPaths3(args, out) {
-  if (!args || typeof args !== "object") return;
-  const obj = args;
-  for (const key of ["path", "file", "filepath", "filePath", "target", "cwd"]) {
-    const v = obj[key];
-    if (typeof v === "string" && v.length > 1 && v.length < 260) {
-      out.add(v.replace(/\\/g, "/"));
-    }
-  }
-  if (typeof obj.file_path === "string") out.add(String(obj.file_path));
-}
-function collectPathsFromText(text, out) {
-  const re = /(?:^|[\s"'`])((?:[\w.-]+\/)+[\w.-]+\.\w{1,8})/g;
-  let m;
-  let n = 0;
-  while ((m = re.exec(text)) !== null && n < 8) {
-    out.add(m[1]);
-    n += 1;
-  }
-}
-var MAX_SUMMARY_CHARS;
-var init_historySummary = __esm({
-  "src/cli/budget/historySummary.ts"() {
-    "use strict";
-    MAX_SUMMARY_CHARS = 3500;
-  }
-});
-
-// src/cli/budget/llmCompact.ts
-function isLlmCompactEnabled() {
-  const v = process.env.ZELARI_LLM_COMPACT?.trim().toLowerCase();
-  if (v === "0" || v === "false" || v === "off" || v === "no") return false;
-  return true;
-}
-function compactModelOverride() {
-  const v = process.env.ZELARI_COMPACT_MODEL?.trim();
-  return v ? v : void 0;
-}
-async function llmSummarizeHistoryReplay(input) {
-  const override = input.overrideModel ?? compactModelOverride();
-  const model = override ?? input.model;
-  const cacheReuseExpected = !override;
-  if (!isLlmCompactEnabled()) return { summary: null, model, cacheReuseExpected };
-  if (input.droppedMessages.length === 0) {
-    return { summary: null, model, cacheReuseExpected };
-  }
-  const messages = [
-    ...input.systemMessages,
-    ...input.droppedMessages,
-    {
-      role: "user",
-      content: COMPACTION_INSTRUCTION
-    }
-  ];
-  const controller = new AbortController();
-  const timeout = setTimeout(() => controller.abort(), REPLAY_TIMEOUT_MS);
-  const onOuterAbort = () => controller.abort();
-  input.signal?.addEventListener("abort", onOuterAbort, { once: true });
-  try {
-    let text = "";
-    let emittedToolCall = false;
-    for await (const delta of input.providerStream({
-      provider: input.provider,
-      model,
-      messages,
-      // Tools stay advertised: dropping them would change the prefix token
-      // sequence and destroy cache reuse (explicit DSH decision). They are
-      // sorted canonically (same discipline as the live routed request and
-      // the snapshot fingerprints) so the replay prefix is byte-identical.
-      tools: [...input.tools].sort((a, b) => a.name < b.name ? -1 : a.name > b.name ? 1 : 0),
-      signal: controller.signal,
-      generation: {
-        purpose: "compaction",
-        temperature: 0.1,
-        maxTokens: 900
-      }
-    })) {
-      if (delta.kind === "text") text += delta.delta;
-      if (delta.kind === "tool_call") emittedToolCall = true;
-    }
-    if (emittedToolCall) return { summary: null, model, cacheReuseExpected };
-    if (!text.trim()) return { summary: null, model, cacheReuseExpected };
-    return { summary: text.trim(), model, cacheReuseExpected };
-  } catch {
-    return { summary: null, model, cacheReuseExpected };
-  } finally {
-    clearTimeout(timeout);
-    input.signal?.removeEventListener("abort", onOuterAbort);
-  }
-}
-var COMPACTION_INSTRUCTION, REPLAY_TIMEOUT_MS;
-var init_llmCompact = __esm({
-  "src/cli/budget/llmCompact.ts"() {
-    "use strict";
-    COMPACTION_INSTRUCTION = `
-You are now acting as a compaction engine for this coding-agent session.
-
-Condense the conversation ABOVE into a compact checkpoint sufficient to continue the task.
-
-Preserve:
-- user's goal and evolving intent
-- decisions already made
-- exact file paths and identifiers
-- code changes already completed
-- commands/errors that still matter
-- constraints
-- unfinished work
-- the single most likely next action
-
-Do not call tools.
-Do not mention this summarization request.
-Output only the checkpoint.
-Be concise.
-`.trim();
-    REPLAY_TIMEOUT_MS = 6e4;
-  }
-});
-
-// src/cli/hooks/historyCompaction.ts
-function compactedRangeFromDropped(dropped) {
-  if (dropped.length === 0) return void 0;
-  const seqs = [];
-  const sources = [];
-  for (const m of dropped) {
-    const hasCompactRange = typeof m.compactedFromSeq === "number" && Number.isInteger(m.compactedFromSeq) && m.compactedFromSeq > 0 && typeof m.compactedToSeq === "number" && Number.isInteger(m.compactedToSeq) && m.compactedToSeq >= m.compactedFromSeq;
-    if (hasCompactRange) {
-      seqs.push(m.compactedFromSeq, m.compactedToSeq);
-      sources.push(...m.sourceEventSeqs ?? []);
-      if (typeof m.seq === "number" && Number.isInteger(m.seq) && m.seq > 0) {
-        seqs.push(m.seq);
-        sources.push(m.seq);
-      }
-      continue;
-    }
-    if (typeof m.seq !== "number" || !Number.isInteger(m.seq) || m.seq < 1) return void 0;
-    seqs.push(m.seq);
-    sources.push(m.seq);
-  }
-  return {
-    fromSeq: Math.min(...seqs),
-    toSeq: Math.max(...seqs),
-    sourceEventSeqs: [...new Set(sources)]
-  };
-}
-function withDroppedRange(result, dropped, strategy) {
-  const range = compactedRangeFromDropped(dropped);
-  if (!range) return { ...result, strategy };
-  return { ...result, ...range, strategy };
-}
-function resolveMaxMessages(opts) {
-  const envTurns = envNumber(process.env.ZELARI_HISTORY_TURNS, { default: 6, min: 0 });
-  let turns = opts?.maxMessages ? Math.ceil(opts.maxMessages / 4) : envTurns;
-  if (opts?.durableStatePresent && !opts?.maxMessages && process.env.ZELARI_HISTORY_TURNS === void 0) {
-    turns = Math.min(turns, 3);
-  }
-  if (turns <= 0) return 0;
-  if (opts?.force && opts?.maxMessages) return Math.max(1, opts.maxMessages);
-  return turns * 4;
-}
-function findValidCutIndex(messages, naiveCut) {
-  let cut = naiveCut;
-  while (cut < messages.length) {
-    const kept = messages.slice(cut);
-    const declared = /* @__PURE__ */ new Set();
-    for (const m of kept) {
-      if (m.role === "assistant" && m.toolCalls) {
-        for (const tc of m.toolCalls) declared.add(tc.id);
-      }
-    }
-    let moved = false;
-    for (let k = 0; k < kept.length; k++) {
-      const m = kept[k];
-      if (m.role === "tool" && m.toolCallId && !declared.has(m.toolCallId)) {
-        for (let j = cut - 1; j >= 0; j--) {
-          const prev2 = messages[j];
-          if (prev2.role === "assistant" && prev2.toolCalls && prev2.toolCalls.some((tc) => tc.id === m.toolCallId)) {
-            cut = j;
-            moved = true;
-            break;
-          }
-        }
-        break;
-      }
-    }
-    if (!moved) break;
-  }
-  return cut;
-}
-function resolvePruneLimits(opts) {
-  const maxChars = opts?.maxChars ?? envNumber(process.env.ZELARI_TOOL_RESULT_MAX_CHARS, { default: 8e3, min: 256 });
-  const rawTail = opts?.tailChars ?? envNumber(process.env.ZELARI_TOOL_RESULT_TAIL_CHARS, { default: 1e3, min: 0 });
-  const tailChars = Math.min(rawTail, maxChars);
-  return { maxChars, tailChars };
-}
-function pruneToolResultsDetailed(messages, opts) {
-  const { maxChars, tailChars } = resolvePruneLimits(opts);
-  const headChars = maxChars - tailChars;
-  const stats = { pruned: 0, charsOmitted: 0 };
-  let changed = false;
-  const out = messages.map((m) => {
-    if (m.role !== "tool") return m;
-    const body = m.content ?? "";
-    if (body.length <= maxChars) return m;
-    const head = headChars > 0 ? body.slice(0, headChars) : "";
-    const tail2 = tailChars > 0 ? body.slice(-tailChars) : "";
-    const omitted = body.length - head.length - tail2.length;
-    changed = true;
-    stats.pruned += 1;
-    stats.charsOmitted += omitted;
-    return {
-      ...m,
-      content: [head, "\u2026[pruned " + omitted + " chars]\u2026", tail2].join(String.fromCharCode(10))
-    };
-  });
-  return {
-    messages: changed ? out : messages,
-    stats
-  };
-}
-function compactHistory(messages, opts) {
-  return compactHistoryDetailed(messages, opts).messages;
-}
-function buildCheckpointMessage(summaryText, range) {
-  return {
-    role: "user",
-    content: CHECKPOINT_WRAPPER_PREFIX + "\n\n<compacted-summary>\n" + summaryText + "\n</compacted-summary>",
-    ...range ? {
-      compactedFromSeq: range.fromSeq,
-      compactedToSeq: range.toSeq,
-      sourceEventSeqs: [...range.sourceEventSeqs]
-    } : {}
-  };
-}
-function compactHistoryDetailed(messages, opts) {
-  const maxMessages = resolveMaxMessages(opts);
-  if (maxMessages === 0) {
-    return withDroppedRange(
-      { messages: [], compacted: true, messagesRemoved: messages.length, summary: "" },
-      messages,
-      "extractive"
-    );
-  }
-  if (messages.length <= maxMessages * 2 && !opts?.force) {
-    return {
-      messages,
-      compacted: false,
-      messagesRemoved: 0,
-      summary: ""
-    };
-  }
-  const naiveCut = Math.max(0, messages.length - maxMessages);
-  const cut = findValidCutIndex(messages, naiveCut);
-  if (cut === 0) {
-    return {
-      messages,
-      compacted: false,
-      messagesRemoved: 0,
-      summary: ""
-    };
-  }
-  const droppedMsgs = messages.slice(0, cut);
-  const droppedRange = compactedRangeFromDropped(droppedMsgs);
-  const pruned = pruneToolResultsDetailed(messages.slice(cut));
-  const kept = pruned.messages;
-  const summaryText = extractiveHistorySummary(droppedMsgs);
-  const summary = buildCheckpointMessage(
-    summaryText || `${COMPACT_MARKER} ${cut} earlier message(s) dropped.`,
-    droppedRange
-  );
-  return withDroppedRange(
-    {
-      messages: [summary, ...kept],
-      compacted: true,
-      messagesRemoved: cut,
-      summary: summary.content,
-      prunedToolResults: pruned.stats.pruned
-    },
-    droppedMsgs,
-    "extractive"
-  );
-}
-async function compactHistoryAsync(messages, opts) {
-  const base2 = compactHistoryDetailed(messages, opts);
-  if (!base2.compacted || base2.messagesRemoved === 0) return base2;
-  const cut = base2.messagesRemoved;
-  const droppedMsgs = messages.slice(0, cut);
-  const extractive = extractiveHistorySummary(droppedMsgs);
-  let summaryText = extractive;
-  let cacheReuseExpected;
-  let replayExactPrefix;
-  const canReplay = !!(opts?.providerStream && opts?.requestSnapshot);
-  if (canReplay) {
-    try {
-      const replay = await llmSummarizeHistoryReplay({
-        providerStream: opts.providerStream,
-        provider: opts.requestSnapshot.provider,
-        model: opts.requestSnapshot.model,
-        systemMessages: opts.requestSnapshot.systemMessages,
-        tools: opts.requestSnapshot.tools,
-        droppedMessages: droppedMsgs,
-        signal: opts?.signal
-      });
-      cacheReuseExpected = replay.cacheReuseExpected;
-      if (replay.summary && replay.summary.trim().length > 40) {
-        const sourceTokens = roughTokens(droppedMsgs);
-        const summaryTok = Math.ceil(replay.summary.length / 4);
-        if (summaryTok < sourceTokens) {
-          summaryText = replay.summary.trim();
-        }
-      }
-    } catch {
-    }
-  }
-  const pruned = pruneToolResultsDetailed(messages.slice(cut));
-  const kept = pruned.messages;
-  const summary = buildCheckpointMessage(summaryText, compactedRangeFromDropped(droppedMsgs));
-  const usedLlm = summaryText !== extractive && summaryText.trim().length > 40;
-  return withDroppedRange(
-    {
-      messages: [summary, ...kept],
-      compacted: true,
-      messagesRemoved: cut,
-      summary: summaryText,
-      prunedToolResults: pruned.stats.pruned,
-      cacheReuseExpected,
-      replayExactPrefix
-    },
-    droppedMsgs,
-    usedLlm ? "llm" : "extractive"
-  );
-}
-function roughTokens(msgs) {
-  let n = 0;
-  for (const m of msgs) {
-    n += Math.ceil((m.content ?? "").length / 4);
-    if (m.toolCalls) {
-      for (const tc of m.toolCalls) {
-        n += Math.ceil(JSON.stringify(tc.args ?? {}).length / 4);
-      }
-    }
-  }
-  return Math.max(1, n);
-}
-var COMPACT_MARKER, CHECKPOINT_WRAPPER_PREFIX;
-var init_historyCompaction = __esm({
-  "src/cli/hooks/historyCompaction.ts"() {
-    "use strict";
-    init_historySummary();
-    init_llmCompact();
-    init_envNumber();
-    COMPACT_MARKER = "[history] Earlier turns were compacted to stay within the context budget.";
-    CHECKPOINT_WRAPPER_PREFIX = "This is an automatically generated checkpoint of earlier conversation. Treat it as established context and continue directly.";
-  }
-});
-
 // src/cli/budget/requestSnapshotStore.ts
 function recordRequestSnapshot(sessionId2, snapshot) {
   store5.set(sessionId2, { snapshot });
@@ -51253,215 +51584,6 @@ var init_phase = __esm({
   }
 });
 
-// src/cli/budget/tokenBudget.ts
-function mergeCompactRange(into, r) {
-  if (r.fromSeq !== void 0 && r.toSeq !== void 0) {
-    into.fromSeq = into.fromSeq === void 0 ? r.fromSeq : Math.min(into.fromSeq, r.fromSeq);
-    into.toSeq = into.toSeq === void 0 ? r.toSeq : Math.max(into.toSeq, r.toSeq);
-    if (r.sourceEventSeqs) into.sourceSeqs.push(...r.sourceEventSeqs);
-  }
-  if (r.strategy === "llm") into.strategy = "llm";
-  else if (r.strategy && !into.strategy) into.strategy = r.strategy;
-}
-function estimateTokens2(text) {
-  if (!text) return 0;
-  return Math.max(1, Math.ceil(text.length / 4));
-}
-function estimateHistoryTokens(messages) {
-  let n = 0;
-  for (const m of messages) {
-    n += estimateTokens2(m.content);
-    if (m.toolCalls) {
-      for (const tc of m.toolCalls) {
-        n += estimateTokens2(tc.name) + estimateTokens2(JSON.stringify(tc.args ?? {}));
-      }
-    }
-  }
-  return n;
-}
-function defaultContextLimitForModel(model, provider) {
-  return capabilitiesFor(model, provider).contextWindow;
-}
-function resolveContextLimit(model, provider) {
-  return envNumber(process.env.ZELARI_CONTEXT_LIMIT, {
-    default: defaultContextLimitForModel(model, provider),
-    min: 4e3,
-    max: 2e6
-  });
-}
-function phaseKnobs(phase2) {
-  return {
-    historyTurns: phase2 === "plan" ? envNumber(process.env.ZELARI_HISTORY_TURNS, { default: 8, min: 0 }) : envNumber(process.env.ZELARI_HISTORY_TURNS, { default: 6, min: 0 }),
-    maxToolLoopIterations: phase2 === "plan" ? envNumber(process.env.ZELARI_MAX_TOOL_LOOP_ITERATIONS, { default: 60, min: 1 }) : envNumber(process.env.ZELARI_MAX_TOOL_LOOP_ITERATIONS, {
-      default: 120,
-      min: 1
-    })
-  };
-}
-async function applyBudgetPolicyAsync(history2, phase2, opts) {
-  const contextLimit = resolveContextLimit(opts?.model, opts?.provider);
-  const compact3 = capabilitiesFor(opts?.model, opts?.provider).compaction;
-  const sessionExtra = opts?.sessionTokens ?? 0;
-  const warnings = [];
-  let { historyTurns, maxToolLoopIterations } = phaseKnobs(phase2);
-  const envelope = opts?.requestSnapshot ?? null;
-  const surface = envelope?.snapshot ?? opts?.requestSurface ?? null;
-  const replayBase = surface ? {
-    provider: surface.provider,
-    model: surface.model,
-    systemMessages: surface.systemMessages,
-    tools: surface.tools
-  } : opts?.providerStream ? { provider: "local", model: opts?.model ?? "unknown", systemMessages: [], tools: [] } : null;
-  const headerTokens = surface ? estimateSystemTokensLite(surface.systemMessages) + estimateToolSchemaTokensLite(surface.tools) : 0;
-  const convTokensOf = (h) => estimateConversationTokensLite(h);
-  let hist = history2;
-  let estimated = surface ? headerTokens + convTokensOf(hist) + RESERVED_OUTPUT_TOKENS : estimateHistoryTokens(hist) + sessionExtra;
-  let occupancy = Math.min(1, estimated / contextLimit);
-  let compactSummary = "";
-  let messagesRemoved = 0;
-  let cacheReuseExpected;
-  let prunedTotal = 0;
-  const compactRange = { sourceSeqs: [] };
-  if (occupancy >= compact3.warnAt && occupancy < compact3.compactAt) {
-    warnings.push(
-      `[budget] context ~${Math.round(occupancy * 100)}% full (${estimated}/${contextLimit} tok full-request est.) \u2014 consider /compact or shorter replies.`
-    );
-  }
-  if (occupancy >= 0.8) {
-    const pruned = pruneToolResultsDetailed(hist);
-    if (pruned.stats.pruned > 0) {
-      hist = pruned.messages;
-      prunedTotal += pruned.stats.pruned;
-      estimated = surface ? headerTokens + convTokensOf(hist) + RESERVED_OUTPUT_TOKENS : estimateHistoryTokens(hist) + sessionExtra;
-      occupancy = Math.min(1, estimated / contextLimit);
-      warnings.push(
-        `[budget] pruned ${pruned.stats.pruned} oversized tool result(s) \u2192 ${Math.round(occupancy * 100)}% (${estimated} tok).`
-      );
-    }
-  }
-  const fold = (r, label, forcedTurns) => {
-    hist = applySessionSurface(r.messages);
-    if (r.compacted) {
-      messagesRemoved += r.messagesRemoved;
-      if (r.summary) compactSummary = r.summary;
-      mergeCompactRange(compactRange, r);
-      if (r.cacheReuseExpected !== void 0) {
-        cacheReuseExpected = r.cacheReuseExpected;
-      }
-    }
-    estimated = surface ? headerTokens + convTokensOf(hist) + RESERVED_OUTPUT_TOKENS : estimateHistoryTokens(hist) + sessionExtra;
-    occupancy = Math.min(1, estimated / contextLimit);
-    warnings.push(
-      `[budget] ${label} \u2014 kept ~${forcedTurns} turns (${estimated} tok est.` + (r.messagesRemoved ? `, removed ${r.messagesRemoved} msgs` : "") + (r.cacheReuseExpected === false ? ", cache reuse NOT expected (model override)" : "") + ")."
-    );
-  };
-  if (occupancy >= compact3.compactAt) {
-    const forcedTurns = Math.max(1, Math.floor(historyTurns / 2));
-    historyTurns = forcedTurns;
-    maxToolLoopIterations = Math.min(
-      maxToolLoopIterations,
-      phase2 === "plan" ? 24 : 40
-    );
-    let r = await compactHistoryAsync(hist, {
-      maxMessages: Math.max(2, forcedTurns * 4),
-      force: true,
-      signal: opts?.signal,
-      ...replayBase ? { requestSnapshot: replayBase, providerStream: opts?.providerStream } : {}
-    });
-    if (!r.compacted) {
-      r = await compactHistoryAsync(hist, {
-        maxMessages: 2,
-        force: true,
-        signal: opts?.signal,
-        ...replayBase ? { requestSnapshot: replayBase, providerStream: opts?.providerStream } : {}
-      });
-    }
-    const label = r.cacheReuseExpected === false ? "llm-compact (model override)" : "auto-compact";
-    fold(r, label, forcedTurns);
-  }
-  if (occupancy >= compact3.hardAt) {
-    const hard = await compactHistoryAsync(hist, {
-      maxMessages: 2,
-      force: true,
-      signal: opts?.signal
-    });
-    fold(hard, hard.summary.includes("\xB7 llm") ? "llm-compact" : "HARD trim", 2);
-    historyTurns = 2;
-    maxToolLoopIterations = Math.min(maxToolLoopIterations, 16);
-    warnings.push(
-      "[budget] HARD context pressure (\u226595%) \u2014 prefer /clear or a new session if quality drops."
-    );
-  }
-  const cacheMetricsLine = envelope ? [
-    "compaction meter:",
-    `provider/model: ${envelope.snapshot.provider}/${envelope.snapshot.model}`,
-    `headerFingerprint: ${envelope.snapshot.headerFingerprint.slice(0, 12)}`,
-    `occupancy: ${Math.round(occupancy * 100)}% (${estimated}/${contextLimit})`,
-    ...envelope.usage?.cachedPromptTokens !== void 0 ? [`cachedPromptTokens: ${envelope.usage.cachedPromptTokens}`] : [],
-    ...cacheReuseExpected !== void 0 ? [`cacheReuseExpected: ${cacheReuseExpected}`] : []
-  ].join(" | ") : void 0;
-  return {
-    history: hist,
-    warnings,
-    maxToolLoopIterations,
-    historyTurns,
-    estimatedHistoryTokens: surface ? convTokensOf(hist) : estimated,
-    contextLimit,
-    occupancy,
-    compactSummary: compactSummary || void 0,
-    messagesRemoved: messagesRemoved || void 0,
-    ...compactRange.fromSeq !== void 0 && compactRange.toSeq !== void 0 ? {
-      compactedFromSeq: compactRange.fromSeq,
-      compactedToSeq: compactRange.toSeq,
-      compactSourceSeqs: compactRange.sourceSeqs,
-      compactStrategy: compactRange.strategy
-    } : {},
-    ...surface ? { contextPressureTokens: estimated } : {},
-    ...cacheReuseExpected !== void 0 ? { cacheReuseExpected } : {},
-    ...cacheMetricsLine ? { cacheMetricsLine } : {}
-  };
-}
-function estimateSystemTokensLite(systemMessages) {
-  let n = 0;
-  for (const m of systemMessages) n += 4 + Math.ceil((m.content ?? "").length / 4);
-  return n;
-}
-function estimateToolSchemaTokensLite(tools) {
-  let n = 0;
-  for (const t of tools) {
-    n += Math.ceil((t.name ?? "").length / 4);
-    n += Math.ceil((t.description ?? "").length / 4);
-    n += Math.ceil(JSON.stringify(t.parameters ?? {}).length / 4);
-  }
-  return n + tools.length * 4;
-}
-function estimateConversationTokensLite(messages) {
-  let n = 0;
-  for (const m of messages) {
-    n += 4 + Math.ceil((m.content ?? "").length / 4);
-    if (m.toolCalls) {
-      for (const tc of m.toolCalls) {
-        n += Math.ceil(tc.name.length / 4) + Math.ceil(tc.id.length / 4);
-        n += Math.ceil(JSON.stringify(tc.args ?? {}).length / 4);
-      }
-    }
-    if (m.reasoningContent) n += Math.ceil(m.reasoningContent.length / 4);
-    if (m.toolCallId) n += Math.ceil(m.toolCallId.length / 4);
-  }
-  return n;
-}
-var RESERVED_OUTPUT_TOKENS;
-var init_tokenBudget = __esm({
-  "src/cli/budget/tokenBudget.ts"() {
-    "use strict";
-    init_envNumber();
-    init_historyCompaction();
-    init_observationStore();
-    init_capabilities();
-    RESERVED_OUTPUT_TOKENS = 8192;
-  }
-});
-
 // src/cli/budget/requestMeter.ts
 function estimateTokensLocal(text) {
   if (!text) return 0;
@@ -55182,39 +55304,224 @@ var init_toolRegistry2 = __esm({
   }
 });
 
+// src/cli/mcp/httpTransport.ts
+var SSE_DATA_RE, ABORT_GRACE_MS, McpHttpTransport;
+var init_httpTransport = __esm({
+  "src/cli/mcp/httpTransport.ts"() {
+    "use strict";
+    SSE_DATA_RE = /^data:\s?(.*)$/;
+    ABORT_GRACE_MS = 250;
+    McpHttpTransport = class {
+      constructor(opts) {
+        this.opts = opts;
+      }
+      sessionId = null;
+      closed = false;
+      reinit = null;
+      controllers = /* @__PURE__ */ new Set();
+      get hasSession() {
+        return this.sessionId !== null;
+      }
+      /**
+       * Deliver one JSON-RPC message. Resolves once the response (if any) has
+       * been fed to onMessage. Transport-level failures are converted into
+       * JSON-RPC error responses so the client's pending map rejects cleanly
+       * through the same pump used for stdio.
+       */
+      async send(msg, timeoutMs2) {
+        if (this.closed) throw new Error(`[mcp:${this.opts.serverName}] transport closed`);
+        try {
+          await this.post(msg, timeoutMs2, false);
+        } catch (err) {
+          if (msg.id !== void 0) {
+            this.opts.onMessage({
+              jsonrpc: "2.0",
+              id: msg.id,
+              error: {
+                code: -32e3,
+                message: err instanceof Error ? err.message : String(err)
+              }
+            });
+          }
+        }
+      }
+      /** Best-effort session teardown (HTTP DELETE), then abort in-flight POSTs. */
+      close() {
+        this.closed = true;
+        for (const ac of this.controllers) ac.abort();
+        this.controllers.clear();
+        const sid = this.sessionId;
+        this.sessionId = null;
+        if (!sid) return;
+        const headers2 = {
+          ...this.opts.headers ?? {},
+          "mcp-session-id": sid
+        };
+        void fetch(this.opts.url, { method: "DELETE", headers: headers2 }).catch(() => {
+        });
+      }
+      // ── internals ────────────────────────────────────────────────────────
+      async post(msg, timeoutMs2, replayed) {
+        const ac = new AbortController();
+        this.controllers.add(ac);
+        const timer = setTimeout(() => ac.abort(), timeoutMs2 + ABORT_GRACE_MS);
+        const hadSession = this.sessionId !== null;
+        try {
+          const headers2 = {
+            "content-type": "application/json",
+            accept: "application/json, text/event-stream",
+            ...this.opts.headers ?? {}
+          };
+          if (this.sessionId) headers2["mcp-session-id"] = this.sessionId;
+          const res = await fetch(this.opts.url, {
+            method: "POST",
+            headers: headers2,
+            signal: ac.signal,
+            body: JSON.stringify({ jsonrpc: "2.0", ...msg })
+          });
+          const sid = res.headers.get("mcp-session-id");
+          if (sid) this.sessionId = sid;
+          if (res.status === 404 && hadSession && !replayed && msg.id !== void 0) {
+            this.sessionId = null;
+            await this.ensureSession();
+            return this.post(msg, timeoutMs2, true);
+          }
+          if (!res.ok && res.status !== 202) {
+            throw new Error(`HTTP ${res.status} ${res.statusText}`.trim());
+          }
+          if (msg.id === void 0) {
+            await res.body?.cancel();
+            return;
+          }
+          const contentType = res.headers.get("content-type") ?? "";
+          if (contentType.includes("text/event-stream")) {
+            await this.readSse(res);
+          } else {
+            const body = await res.text();
+            if (body) this.opts.onMessage(JSON.parse(body));
+          }
+        } finally {
+          clearTimeout(timer);
+          this.controllers.delete(ac);
+        }
+      }
+      /** Dedupe concurrent session-loss recoveries into one handshake. */
+      async ensureSession() {
+        if (!this.reinit) {
+          this.reinit = this.opts.onSessionLost().finally(() => {
+            this.reinit = null;
+          }).catch(() => {
+          });
+        }
+        await this.reinit;
+      }
+      /** Minimal SSE reader: dispatch complete `data:` events to onMessage. */
+      async readSse(res) {
+        const reader = res.body?.getReader();
+        if (!reader) return;
+        const decoder = new TextDecoder();
+        let buf = "";
+        let data = "";
+        const dispatch = () => {
+          const payload = data.trim();
+          data = "";
+          if (!payload) return;
+          try {
+            this.opts.onMessage(JSON.parse(payload));
+          } catch {
+          }
+        };
+        for (; ; ) {
+          const { done, value } = await reader.read();
+          if (done) break;
+          buf += decoder.decode(value, { stream: true });
+          let nl;
+          while ((nl = buf.indexOf("\n")) >= 0) {
+            const line = buf.slice(0, nl).replace(/\r$/, "");
+            buf = buf.slice(nl + 1);
+            if (line === "") {
+              dispatch();
+              continue;
+            }
+            const m = SSE_DATA_RE.exec(line);
+            if (m) data += (data ? "\n" : "") + m[1];
+          }
+        }
+        dispatch();
+      }
+    };
+  }
+});
+
 // src/cli/mcp/mcpClient.ts
 import { spawn as spawn14 } from "node:child_process";
-var DEFAULT_REQUEST_TIMEOUT_MS, INIT_TIMEOUT_MS, MCP_PROTOCOL_VERSION, McpClient;
+var DEFAULT_REQUEST_TIMEOUT_MS, INIT_TIMEOUT_MS, MAX_LIST_PAGES, MCP_PROTOCOL_VERSION, McpClient;
 var init_mcpClient = __esm({
   "src/cli/mcp/mcpClient.ts"() {
     "use strict";
     init_cmdline();
     init_updater();
+    init_httpTransport();
     DEFAULT_REQUEST_TIMEOUT_MS = 3e4;
     INIT_TIMEOUT_MS = 15e3;
+    MAX_LIST_PAGES = 50;
     MCP_PROTOCOL_VERSION = "2025-03-26";
     McpClient = class {
       constructor(serverName, config2) {
         this.serverName = serverName;
         this.config = config2;
+        this.defaultTimeoutMs = config2.timeoutMs ?? DEFAULT_REQUEST_TIMEOUT_MS;
+        this.serial = config2.serial ?? this.transportKind() === "http";
       }
       child = null;
+      transport = null;
       nextId = 1;
       pending = /* @__PURE__ */ new Map();
       stdoutBuffer = "";
       closed = false;
-      /** Spawn the server process and run the MCP initialize handshake. */
+      /** Tail of the serial queue (config.serial); requests run one at a time. */
+      queueTail = Promise.resolve();
+      /** True while the 404-recovery handshake runs: bypasses the serial queue
+       *  (the queued request that triggered the 404 is waiting on this handshake). */
+      recovering = false;
+      serial;
+      defaultTimeoutMs;
+      transportKind() {
+        if (this.config.type) return this.config.type;
+        return this.config.url && !this.config.command ? "http" : "stdio";
+      }
+      /** Connect (spawn / HTTP session) and run the MCP initialize handshake. */
       async start() {
-        if (this.child) return;
+        if (this.child || this.transport) return;
+        if (this.transportKind() === "http") {
+          const url2 = this.config.url;
+          if (!url2) {
+            throw new Error(`[mcp:${this.serverName}] http server requires a url`);
+          }
+          this.transport = new McpHttpTransport({
+            serverName: this.serverName,
+            url: url2,
+            // env may carry an explicit Authorization header for remote servers.
+            headers: this.config.env?.AUTHORIZATION ? { Authorization: this.config.env.AUTHORIZATION } : void 0,
+            onMessage: (msg) => this.handleMessage(msg),
+            onSessionLost: () => this.handshake()
+          });
+          await this.handshake();
+          return;
+        }
+        const command = this.config.command;
+        if (!command) {
+          throw new Error(`[mcp:${this.serverName}] stdio server requires a command`);
+        }
         const spawnOpts = {
           stdio: ["pipe", "pipe", "pipe"],
           env: { ...process.env, ...this.config.env ?? {} },
           windowsHide: true
         };
-        const child = process.platform === "win32" ? spawn14(buildCmdLine(this.config.command, this.config.args ?? []), {
+        const child = process.platform === "win32" ? spawn14(buildCmdLine(command, this.config.args ?? []), {
           ...spawnOpts,
           shell: true
-        }) : spawn14(this.config.command, this.config.args ?? [], spawnOpts);
+        }) : spawn14(command, this.config.args ?? [], spawnOpts);
         this.child = child;
         child.stdout.setEncoding("utf8");
         child.stdout.on("data", (chunk) => this.onStdout(chunk));
@@ -55233,37 +55540,60 @@ var init_mcpClient = __esm({
             );
           }
         });
-        await this.request(
-          "initialize",
-          {
-            protocolVersion: MCP_PROTOCOL_VERSION,
-            capabilities: {},
-            clientInfo: { name: "zelari-code", version: getCurrentVersion() }
-          },
-          INIT_TIMEOUT_MS
-        );
-        this.notify("notifications/initialized", {});
+        await this.handshake();
+      }
+      async handshake() {
+        const prev2 = this.recovering;
+        this.recovering = true;
+        try {
+          await this.request(
+            "initialize",
+            {
+              protocolVersion: MCP_PROTOCOL_VERSION,
+              capabilities: {},
+              clientInfo: { name: "zelari-code", version: getCurrentVersion() }
+            },
+            INIT_TIMEOUT_MS
+          );
+          this.notify("notifications/initialized", {});
+        } finally {
+          this.recovering = prev2;
+        }
       }
-      /** Discover the server's tools. */
+      /**
+       * Discover the server's tools. Follows `nextCursor` pagination so
+       * large servers (hundreds of tools) are listed completely.
+       */
       async listTools() {
-        const res = await this.request("tools/list", {});
-        return (res.tools ?? []).filter(
-          (t) => !!t.name
-        ).map((t) => ({
-          name: t.name,
-          description: t.description ?? "",
-          inputSchema: t.inputSchema ?? { type: "object", properties: {} }
-        }));
+        const tools = [];
+        let cursor;
+        for (let page = 0; page < MAX_LIST_PAGES; page++) {
+          const res = await this.request(
+            "tools/list",
+            cursor ? { cursor } : {}
+          );
+          for (const t of res.tools ?? []) {
+            if (!t.name) continue;
+            tools.push({
+              name: t.name,
+              description: t.description ?? "",
+              inputSchema: t.inputSchema ?? { type: "object", properties: {} }
+            });
+          }
+          cursor = typeof res.nextCursor === "string" ? res.nextCursor : void 0;
+          if (!cursor) break;
+        }
+        return tools;
       }
       /**
        * Call a tool. Returns the concatenated text content; non-text content
        * items are summarized by type. Throws when the server flags isError.
        */
-      async callTool(name, args, timeoutMs2 = DEFAULT_REQUEST_TIMEOUT_MS) {
+      async callTool(name, args, timeoutMs2) {
         const res = await this.request(
           "tools/call",
           { name, arguments: args },
-          timeoutMs2
+          timeoutMs2 ?? this.defaultTimeoutMs
         );
         const text = (res.content ?? []).map(
           (c) => c.type === "text" && typeof c.text === "string" ? c.text : `[${c.type ?? "unknown"} content]`
@@ -55272,20 +55602,39 @@ var init_mcpClient = __esm({
           throw new Error(text || `tool "${name}" reported an error`);
         return text;
       }
-      /** Terminate the server process and reject all in-flight requests. */
+      /** Terminate the server / session and reject all in-flight requests. */
       close() {
         this.closed = true;
         this.failAll(new Error(`[mcp:${this.serverName}] client closed`));
+        if (this.transport) {
+          this.transport.close();
+          this.transport = null;
+          return;
+        }
         this.child?.kill();
         this.child = null;
       }
-      // ── JSON-RPC plumbing ────────────────────────────────────────────────
-      request(method, params, timeoutMs2 = DEFAULT_REQUEST_TIMEOUT_MS) {
+      // ── JSON-RPC plumbing (shared by both transports) ────────────────────
+      request(method, params, timeoutMs2) {
+        const effective = timeoutMs2 ?? this.defaultTimeoutMs;
+        if (!this.serial || this.recovering)
+          return this.dispatch(method, params, effective);
+        const run = this.queueTail.then(
+          () => this.dispatch(method, params, effective),
+          () => this.dispatch(method, params, effective)
+        );
+        this.queueTail = run.then(
+          () => void 0,
+          () => void 0
+        );
+        return run;
+      }
+      dispatch(method, params, timeoutMs2) {
+        const transport = this.transport;
         const child = this.child;
-        if (!child)
+        if (!transport && !child)
           return Promise.reject(new Error(`[mcp:${this.serverName}] not started`));
         const id3 = this.nextId++;
-        const payload = JSON.stringify({ jsonrpc: "2.0", id: id3, method, params });
         return new Promise((resolve9, reject) => {
           const timer = setTimeout(() => {
             this.pending.delete(id3);
@@ -55296,16 +55645,25 @@ var init_mcpClient = __esm({
             );
           }, timeoutMs2);
           this.pending.set(id3, { resolve: resolve9, reject, timer });
-          child.stdin.write(payload + "\n", (err) => {
-            if (err) {
-              clearTimeout(timer);
-              this.pending.delete(id3);
-              reject(err);
-            }
-          });
+          if (transport) {
+            void transport.send({ id: id3, method, params }, timeoutMs2);
+          } else {
+            const payload = JSON.stringify({ jsonrpc: "2.0", id: id3, method, params });
+            child.stdin.write(payload + "\n", (err) => {
+              if (err) {
+                clearTimeout(timer);
+                this.pending.delete(id3);
+                reject(err);
+              }
+            });
+          }
         });
       }
       notify(method, params) {
+        if (this.transport) {
+          void this.transport.send({ method, params }, this.defaultTimeoutMs);
+          return;
+        }
         this.child?.stdin.write(
           JSON.stringify({ jsonrpc: "2.0", method, params }) + "\n"
         );
@@ -55317,26 +55675,29 @@ var init_mcpClient = __esm({
           const line = this.stdoutBuffer.slice(0, nl).trim();
           this.stdoutBuffer = this.stdoutBuffer.slice(nl + 1);
           if (!line) continue;
-          let msg;
           try {
-            msg = JSON.parse(line);
+            this.handleMessage(JSON.parse(line));
           } catch {
             continue;
           }
-          if (typeof msg.id !== "number") continue;
-          const pending = this.pending.get(msg.id);
-          if (!pending) continue;
-          this.pending.delete(msg.id);
-          clearTimeout(pending.timer);
-          if (msg.error) {
-            pending.reject(
-              new Error(
-                `[mcp:${this.serverName}] ${msg.error.message ?? "JSON-RPC error"} (code ${msg.error.code ?? "?"})`
-              )
-            );
-          } else {
-            pending.resolve(msg.result);
-          }
+        }
+      }
+      /** Route one parsed JSON-RPC message to its pending request (both transports). */
+      handleMessage(msg) {
+        const m = msg;
+        if (typeof m.id !== "number") return;
+        const pending = this.pending.get(m.id);
+        if (!pending) return;
+        this.pending.delete(m.id);
+        clearTimeout(pending.timer);
+        if (m.error) {
+          pending.reject(
+            new Error(
+              `[mcp:${this.serverName}] ${m.error.message ?? "JSON-RPC error"} (code ${m.error.code ?? "?"})`
+            )
+          );
+        } else {
+          pending.resolve(m.result);
         }
       }
       failAll(err) {
@@ -55371,11 +55732,17 @@ function readFile6(path91) {
     const parsed = JSON.parse(readFileSync29(path91, "utf8"));
     const out = {};
     for (const [name, cfg] of Object.entries(parsed.mcpServers ?? {})) {
-      if (!cfg || typeof cfg.command !== "string" || !cfg.command.trim()) continue;
+      const hasCommand = !!cfg && typeof cfg.command === "string" && !!cfg.command.trim();
+      const hasUrl = !!cfg && typeof cfg.url === "string" && /^https?:\/\//i.test(cfg.url);
+      if (!cfg || !hasCommand && !hasUrl) continue;
       out[name] = {
-        command: cfg.command.trim(),
+        command: hasCommand ? cfg.command.trim() : void 0,
         args: Array.isArray(cfg.args) ? cfg.args.map(String) : void 0,
         env: cfg.env && typeof cfg.env === "object" ? cfg.env : void 0,
+        type: hasUrl && !hasCommand ? "http" : cfg.type === "http" ? "http" : "stdio",
+        url: hasUrl ? cfg.url.trim() : void 0,
+        timeoutMs: typeof cfg.timeoutMs === "number" && cfg.timeoutMs > 0 ? cfg.timeoutMs : void 0,
+        serial: typeof cfg.serial === "boolean" ? cfg.serial : void 0,
         enabled: cfg.enabled !== false
       };
     }
@@ -55417,8 +55784,13 @@ function upsertMcpServer(opts) {
       error: "Invalid server name (use letters, digits, _ -)"
     };
   }
-  if (!opts.config.command?.trim()) {
-    return { ok: false, error: "command is required" };
+  const hasCommand = !!opts.config.command?.trim();
+  const hasUrl = typeof opts.config.url === "string" && /^https?:\/\//i.test(opts.config.url);
+  if (!hasCommand && !hasUrl) {
+    return {
+      ok: false,
+      error: "either command (stdio) or url (http) is required"
+    };
   }
   let path91;
   if (opts.scope === "user") {
@@ -55435,9 +55807,13 @@ function upsertMcpServer(opts) {
   }
   const current = readFile6(path91);
   current[name] = {
-    command: opts.config.command.trim(),
+    command: hasCommand ? opts.config.command.trim() : void 0,
     args: opts.config.args,
     env: opts.config.env,
+    type: hasUrl ? "http" : opts.config.type === "http" ? "http" : "stdio",
+    url: hasUrl ? opts.config.url.trim() : void 0,
+    timeoutMs: opts.config.timeoutMs,
+    serial: opts.config.serial,
     enabled: opts.config.enabled !== false
   };
   writeFile2(path91, current);
@@ -55512,9 +55888,35 @@ function buildQwenMmPreset() {
     ]
   };
 }
+function buildUnrealPreset() {
+  const url2 = process.env.UNREAL_MCP_URL?.trim() || "http://127.0.0.1:8000/mcp";
+  return {
+    id: "unreal-mcp",
+    servers: {
+      "unreal-mcp": {
+        type: "http",
+        url: url2,
+        // Editor tool calls (asset scans, builds, PIE) can be slow.
+        timeoutMs: 12e4,
+        // Epic guidance: never overlap calls on the editor's game thread.
+        serial: true,
+        enabled: true
+      }
+    },
+    notes: [
+      "Unreal Engine 5.8+ \u2014 MCP server embedded in the editor (Experimental feature).",
+      "Editor: enable the 'Model Context Protocol' plugin, then Edit \u2192 Project Settings \u2192 Plugins \u2192 MCP Server.",
+      `Endpoint: ${url2} (override with UNREAL_MCP_URL; port/path configurable in the editor).`,
+      "Tool Search ON by default: tools surface as list_toolsets / describe_toolset / call_tool.",
+      "Requests are serialized and time out after 120s (editor tools can be slow).",
+      "Start the editor before OR after zelari \u2014 unreachable servers are retried on each turn.",
+      "Kill switch: disable the unreal-mcp server in mcp.json or ZELARI_MCP=0"
+    ]
+  };
+}
 function listMcpPresetIds() {
   return Object.keys(PRESETS).filter(
-    (k) => k === "cua" || k === "composio" || k === "qwen-mm-plugins"
+    (k) => k === "cua" || k === "composio" || k === "qwen-mm-plugins" || k === "unreal-mcp"
   );
 }
 function getMcpPreset(id3) {
@@ -55583,7 +55985,10 @@ var init_mcpPresets = __esm({
       "cua-driver": () => CUA_DRIVER_PRESET,
       composio: buildComposioPreset,
       "qwen-mm-plugins": buildQwenMmPreset,
-      "qwen-mm": buildQwenMmPreset
+      "qwen-mm": buildQwenMmPreset,
+      unreal: buildUnrealPreset,
+      "unreal-mcp": buildUnrealPreset,
+      unrealEditor: buildUnrealPreset
     };
   }
 });
@@ -55613,7 +56018,9 @@ function readMcpConfig(projectRoot = process.cwd(), opts) {
     try {
       const parsed = JSON.parse(readFileSync30(p3, "utf8"));
       for (const [name, cfg] of Object.entries(parsed.mcpServers ?? {})) {
-        if (!cfg || typeof cfg.command !== "string" || cfg.command.length === 0) continue;
+        const hasCommand = !!cfg && typeof cfg.command === "string" && cfg.command.length > 0;
+        const hasUrl = !!cfg && typeof cfg.url === "string" && /^https?:\/\//i.test(cfg.url);
+        if (!cfg || !hasCommand && !hasUrl) continue;
         merged[name] = cfg;
       }
     } catch {
@@ -55654,16 +56061,52 @@ async function ensureLoaded(projectRoot) {
       }
     } catch (err) {
       client.close();
-      state2.warnings.push(`[mcp:${name}] disabled: ${err instanceof Error ? err.message : String(err)}`);
+      const msg = err instanceof Error ? err.message : String(err);
+      if (isHttpConfig(cfg)) {
+        state2.pendingHttp.push({ name, cfg });
+        state2.warnings.push(
+          `[mcp:${name}] not reachable yet (${msg}) \u2014 will retry each turn`
+        );
+      } else {
+        state2.warnings.push(`[mcp:${name}] disabled: ${msg}`);
+      }
     }
   }
 }
+function isHttpConfig(cfg) {
+  return cfg.type === "http" || !!cfg.url && !cfg.command;
+}
+async function retryPendingHttp() {
+  if (state2.pendingHttp.length === 0) return;
+  const remaining = [];
+  for (const p3 of state2.pendingHttp) {
+    const client = new McpClient(p3.name, p3.cfg);
+    try {
+      await client.start();
+      const tools = await client.listTools();
+      state2.clients.push(client);
+      for (const info of tools) {
+        state2.tools.push({
+          registryName: sanitizeToolName(`mcp_${p3.name}_${info.name}`),
+          serverName: p3.name,
+          info,
+          client
+        });
+      }
+    } catch {
+      client.close();
+      remaining.push(p3);
+    }
+  }
+  state2.pendingHttp = remaining;
+}
 function sanitizeToolName(raw) {
   return raw.replace(/[^a-zA-Z0-9_-]/g, "_").slice(0, 64);
 }
 async function registerMcpTools(registry4, projectRoot = process.cwd(), opts) {
   if (process.env["ZELARI_MCP"] === "0") return { registered: [], warnings: [] };
   await ensureLoaded(projectRoot);
+  await retryPendingHttp();
   const skipCuaForCouncil = opts?.councilMode === true && !isCuaAllowedForCouncil();
   const registered = [];
   for (const t of state2.tools) {
@@ -55701,6 +56144,7 @@ function closeMcpClients() {
   state2.tools = [];
   state2.loaded = false;
   state2.warnings = [];
+  state2.pendingHttp = [];
 }
 function _resetMcpForTests() {
   closeMcpClients();
@@ -55714,7 +56158,7 @@ var init_mcpManager = __esm({
     init_mcpClient();
     init_mcpPresets();
     init_folderTrust();
-    state2 = { loaded: false, tools: [], warnings: [], clients: [] };
+    state2 = { loaded: false, tools: [], warnings: [], clients: [], pendingHttp: [] };
   }
 });
 
@@ -66217,11 +66661,14 @@ async function runHeadlessCouncil(opts, provider, model, providerStream, extras)
     const { composeProjectContext: composeProjectContext2 } = await Promise.resolve().then(() => (init_composeContext(), composeContext_exports));
     const { loadDurableContext: loadDurableContext2 } = await Promise.resolve().then(() => (init_loadDurableContext(), loadDurableContext_exports));
     const durableState = await loadDurableContext2(cwd);
+    const { resolveRetrievalPolicy: resolveRetrievalPolicy2 } = await Promise.resolve().then(() => (init_retrievalPolicy(), retrievalPolicy_exports));
+    const retrieval = resolveRetrievalPolicy2(councilContext.budget.occupancy);
     const memoryContext = nativeMemory ? (await nativeMemory.buildContext({
       text: effectiveTask,
       useGraph: true,
-      maxChars: 2e3,
-      maxMemories: 8
+      maxChars: retrieval.maxChars,
+      maxMemories: retrieval.maxMemories,
+      ...retrieval.weights ? { weights: retrieval.weights } : {}
     })).text : "";
     const composed = composeProjectContext2({
       mode: "council",
@@ -67412,6 +67859,7 @@ var init_skillConfigIo = __esm({
       "@zelari/core/skills/builtin/planning",
       "@zelari/core/skills/builtin/refactoring",
       "@zelari/core/skills/builtin/review",
+      "@zelari/core/skills/builtin/unrealEditor",
       "@zelari/core/skills/builtin/testing",
       "@zelari/core/skills/builtin/schema-loop",
       "@zelari/core/skills/builtin/computer-use-cua",
@@ -73439,21 +73887,27 @@ function createPermissionAskHandler(opts) {
     const catLabel = cats.length === 1 ? cats[0] : cats.join("+") || "action";
     const title = `Allow tool "${req.toolName}"?`;
     const detail = req.reason + (cats.length ? ` [${cats.join(", ")}]` : "");
+    const claimLines = (req.claims ?? []).map((c) => `\xB7 ${c.summary}`);
+    const note = req.policyNote ? `
+${req.policyNote}` : "";
+    const claimsBlock = claimLines.length ? `
+Claims:
+${claimLines.join("\n")}` : "";
     appendSystem2?.(
       `[permission] ${title}
-${detail}
+${detail}${note}${claimsBlock}
 \u2192 Allow once \xB7 Always (tool) \xB7 Always (${catLabel}) \xB7 Deny`,
       Date.now()
     );
     let settled = false;
     const askTimeoutMs = askUserTimeoutMs();
     let cancelAskTimeout = () => void 0;
-    const finish2 = (ok, note) => {
+    const finish2 = (ok, note2) => {
       if (settled) return;
       settled = true;
       cancelAskTimeout();
       setPicker2(null);
-      if (note) appendSystem2?.(note, Date.now());
+      if (note2) appendSystem2?.(note2, Date.now());
       resolve9(ok);
     };
     cancelAskTimeout = armPickerTimeout(
@@ -74587,6 +75041,7 @@ _(${req.context})_` : "") + "\n\u2192 scegli dalla lista (il membro riprende dop
   let councilMemory;
   let councilMemoryAutoWrite = false;
   let nativeMemoryContext = "";
+  let councilRetrievalBand = "low";
   try {
     const memoryFactory = await Promise.resolve().then(() => (init_serviceFactory(), serviceFactory_exports));
     if (memoryFactory.isMemoryV2Enabled()) {
@@ -74601,11 +75056,17 @@ _(${req.context})_` : "") + "\n\u2192 scegli dalla lista (il membro riprende dop
       });
       councilMemoryAutoWrite = memoryFactory.isMemoryAutoWriteEnabled();
       if (!overrides.ragContext) {
+        const { resolveRetrievalPolicy: resolveRetrievalPolicy2 } = await Promise.resolve().then(() => (init_retrievalPolicy(), retrievalPolicy_exports));
+        const retrieval = resolveRetrievalPolicy2(
+          councilContext.budget.occupancy
+        );
+        councilRetrievalBand = retrieval.band;
         nativeMemoryContext = (await councilMemory.buildContext({
           text: effectiveText,
           useGraph: true,
-          maxChars: 2e3,
-          maxMemories: 8
+          maxChars: retrieval.maxChars,
+          maxMemories: retrieval.maxMemories,
+          ...retrieval.weights ? { weights: retrieval.weights } : {}
         })).text;
       }
     }
@@ -80285,7 +80746,7 @@ function pickRootComponent() {
   }
   if (argv.includes("--help") || argv.includes("-h")) {
     console.log(
-      "zelari-code \u2014 AI Council coding agent CLI.\n\nUsage: zelari-code [options]\n\nOptions:\n  --version, -v       Print version and exit\n  --help, -h          Print this help and exit\n  --doctor            Diagnose install health (shim, bundle, PATH, deps,\n                      node/git/bash in the agent shell)\n  --inspect [--json]  Unified project inspection (config, skills, MCP,\n                      hooks, plugins, AGENTS.md, trust status)\n  --trust [path]      Trust the cwd (or path) so project MCP + hooks load\n  --fix-path          Add the npm global prefix to the user PATH\n                      (Windows only; fixes 'command not found' after install)\n  --fix-budget        Set recommended ZELARI_MAX_TOOL_LOOP_HARD=180,\n                      ZELARI_MAX_TOOL_LOOP_ITERATIONS=60, ZELARI_CONTEXT_LIMIT=400000\n                      at User scope (prevents the agent stopping mid-task)\n  --permission-mcp <socket>  MCP stdio server for external agent permission prompts\n                      (spawned by claude --permission-prompt-tool)\n  --memory-mcp       Optional MCP stdio server for project memory\n    --cwd <path>      Trusted project root (requires ZELARI_MEMORY_MCP=1)\n    --client-id <id>  Stable local owner identity for private memories\n  --memory-json <json>  Read-only project-memory bridge for Zelari Desktop\n  --skip-checks       Skip the boot-time prerequisite check\n                      (alias for ZELARI_SKIP_PREFLIGHT=1)\n  --no-wizard         Skip the first-run wizard\n  --reset-config      Re-run the wizard (clears provider.json on commit)\n  --headless          Run a single task without mounting the TUI\n    --task <text>       Task prompt (required in headless mode)\n    --output json|plain Output format (default: json)\n    --mode kraken|council|zelari  Dispatch mode (default: kraken; agent=alias)\n    --council          Alias for --mode council\n    --phase plan|build  Work phase (default: build)\n    --provider <id>    Provider override (default: active)\n    --model <id>       Model override (default: provider default)\n    --profile <id>     Capability profile (minimal/v1|kraken/v1|council/v1|mission/v1)\n    --resume <id>      Continue a 2.0 spine session\n    --export-session <path>  Write zelari-session-export/1 after the run\n    --strict-done      Evidence-based BUILD completion gate\n  --session-export <id>  Print a portable 2.0 session export (no LLM)\n  serve               Companion host for Android/remote clients (Tailscale)\n    --bind <ip>       Listen address (default: 127.0.0.1; use Tailscale IP)\n    --port <n>        Port (default: 7421)\n    --token <secret>  Bearer token (default: ~/.zelari-code/companion.token)\n    --project <path>  Allowlisted project root (repeatable)\n    --save-projects   Persist --project list to companion.json\n  --print-config      Print provider/model config as JSON (no secrets)\n  --plugins-status    JSON status of optional plugins (Playwright, eslint, \u2026)\n  --plugins-install <id>  Install plugin (playwright also fetches Chromium)\n    --cwd <path>       Workspace for -D installs (default: process.cwd())\n  --set-config        Persist provider/model/endpoint\n    --provider <id>    Set active provider\n    --model <id>       Set model for that provider\n    --endpoint <url>   Custom OpenAI-compatible base URL\n    --endpoint-clear   Remove custom endpoint override\n    --thinking <spec>  Thinking effort (auto|off|low|medium|high|budget:N)\n  --set-key           Store an API key (never printed back)\n    --provider <id>    Provider id (required)\n    --key <secret>     API key (required)\n  --login-oauth       Start subscription OAuth (grok, chatgpt, anthropic)\n    --provider <id>    grok | chatgpt | anthropic\n    --code <paste>    Anthropic magic-link code (CODE#STATE)\n    --no-browser      Do not open the system browser\n  --refresh-oauth     Force-refresh an OAuth access token\n    --provider <id>    grok | chatgpt | anthropic\n  --logout-oauth      Clear stored OAuth credentials\n    --provider <id>    grok | chatgpt | anthropic\n  --discover-models   Refresh model list for a provider\n    --provider <id>    Provider (default: active)\n  --print-mcp         Print MCP server config (user + project)\n    --cwd <path>      Project root for .zelari/mcp.json\n  --set-mcp           Add/update an MCP server entry\n    --name <id>       Server name (required)\n    --command <bin>   Executable (required)\n    --args <json>     JSON array of args (optional)\n    --scope user|project  Default: user\n    --enabled true|false  Default: true\n    --cwd <path>      Required when scope=project\n  --set-mcp-preset    Install a named MCP preset (e.g. cua)\n    --preset cua      Cua Driver desktop computer-use (MCP)\n    --scope user|project  Default: user\n    --cwd <path>      Required when scope=project\n  --remove-mcp        Remove an MCP server entry\n    --name <id> --scope user|project [--cwd <path>]\n  --print-skills      Print skills (builtin + user + project)\n    --cwd <path>      Project root for .zelari/skills\n  --set-skill         Add/update a SKILL.md skill\n    --name <id>       Skill id (required)\n    --description <t> One-line description (required)\n    --body <text>     Markdown instructions (required)\n    --category <c>    plan|refactor|debug|review|test|docs|ops|git|db|maint\n    --tools <csv>     Comma-separated required tools (optional)\n    --cost <l>        low|medium|high (default: medium)\n    --scope user|project  Default: user\n    --cwd <path>      Required when scope=project\n  --remove-skill      Remove a user/project SKILL.md\n    --name <id> --scope user|project [--cwd <path>]\n  --generate-skill-from-url  Fetch URL + draft skill via model (JSON)\n    --url <https://...>  Required\n    --provider <id>   Override active provider\n    --model <name>    Override model for the selected provider\n  --print-ssh-targets Print SSH deploy/monitor targets\n  --set-ssh-target    Upsert target (--json '{...}' or flags)\n  --remove-ssh-target --id <id>\n  --test-ssh-target   --id <id>  (BatchMode ssh true)\n  --print-ssh-pubkey  --path <private-or-.pub>  (display public key)\n\nEnvironment:\n  ZELARI_NO_WIZARD=1    Skip the first-run wizard\n  ZELARI_SKIP_PREFLIGHT=1  Skip the boot prerequisite check\n  ZELARI_NO_PLUGIN_PROMPT=1  Skip the boot plugin-install prompt\n  ZELARI_MEMORY_MCP=1  Enable the optional external memory MCP server\n  ANATHEMA_DEV=1        Disable background update check + preflight\n"
+      "zelari-code \u2014 AI Council coding agent CLI.\n\nUsage: zelari-code [options]\n\nOptions:\n  --version, -v       Print version and exit\n  --help, -h          Print this help and exit\n  --doctor            Diagnose install health (shim, bundle, PATH, deps,\n                      node/git/bash in the agent shell)\n  --inspect [--json]  Unified project inspection (config, skills, MCP,\n                      hooks, plugins, AGENTS.md, trust status)\n  --trust [path]      Trust the cwd (or path) so project MCP + hooks load\n  --fix-path          Add the npm global prefix to the user PATH\n                      (Windows only; fixes 'command not found' after install)\n  --fix-budget        Set recommended ZELARI_MAX_TOOL_LOOP_HARD=180,\n                      ZELARI_MAX_TOOL_LOOP_ITERATIONS=60, ZELARI_CONTEXT_LIMIT=400000\n                      at User scope (prevents the agent stopping mid-task)\n  --permission-mcp <socket>  MCP stdio server for external agent permission prompts\n                      (spawned by claude --permission-prompt-tool)\n  --memory-mcp       Optional MCP stdio server for project memory\n    --cwd <path>      Trusted project root (requires ZELARI_MEMORY_MCP=1)\n    --client-id <id>  Stable local owner identity for private memories\n  --memory-json <json>  Read-only project-memory bridge for Zelari Desktop\n  --skip-checks       Skip the boot-time prerequisite check\n                      (alias for ZELARI_SKIP_PREFLIGHT=1)\n  --no-wizard         Skip the first-run wizard\n  --reset-config      Re-run the wizard (clears provider.json on commit)\n  --headless          Run a single task without mounting the TUI\n    --task <text>       Task prompt (required in headless mode)\n    --output json|plain Output format (default: json)\n    --mode kraken|council|zelari  Dispatch mode (default: kraken; agent=alias)\n    --council          Alias for --mode council\n    --phase plan|build  Work phase (default: build)\n    --provider <id>    Provider override (default: active)\n    --model <id>       Model override (default: provider default)\n    --profile <id>     Capability profile (minimal/v1|kraken/v1|council/v1|mission/v1)\n    --resume <id>      Continue a 2.0 spine session\n    --export-session <path>  Write zelari-session-export/1 after the run\n    --strict-done      Evidence-based BUILD completion gate\n  --session-export <id>  Print a portable 2.0 session export (no LLM)\n  serve               Companion host for Android/remote clients (Tailscale)\n    --bind <ip>       Listen address (default: 127.0.0.1; use Tailscale IP)\n    --port <n>        Port (default: 7421)\n    --token <secret>  Bearer token (default: ~/.zelari-code/companion.token)\n    --project <path>  Allowlisted project root (repeatable)\n    --save-projects   Persist --project list to companion.json\n  --print-config      Print provider/model config as JSON (no secrets)\n  --plugins-status    JSON status of optional plugins (Playwright, eslint, \u2026)\n  --plugins-install <id>  Install plugin (playwright also fetches Chromium)\n    --cwd <path>       Workspace for -D installs (default: process.cwd())\n  --set-config        Persist provider/model/endpoint\n    --provider <id>    Set active provider\n    --model <id>       Set model for that provider\n    --endpoint <url>   Custom OpenAI-compatible base URL\n    --endpoint-clear   Remove custom endpoint override\n    --thinking <spec>  Thinking effort (auto|off|low|medium|high|budget:N)\n  --set-key           Store an API key (never printed back)\n    --provider <id>    Provider id (required)\n    --key <secret>     API key (required)\n  --login-oauth       Start subscription OAuth (grok, chatgpt, anthropic)\n    --provider <id>    grok | chatgpt | anthropic\n    --code <paste>    Anthropic magic-link code (CODE#STATE)\n    --no-browser      Do not open the system browser\n  --refresh-oauth     Force-refresh an OAuth access token\n    --provider <id>    grok | chatgpt | anthropic\n  --logout-oauth      Clear stored OAuth credentials\n    --provider <id>    grok | chatgpt | anthropic\n  --discover-models   Refresh model list for a provider\n    --provider <id>    Provider (default: active)\n  --print-mcp         Print MCP server config (user + project)\n    --cwd <path>      Project root for .zelari/mcp.json\n  --set-mcp           Add/update an MCP server entry\n    --name <id>       Server name (required)\n    --command <bin>   Executable for stdio servers\n    --url <endpoint>  Streamable HTTP endpoint (e.g. UE 5.8 editor)\n    --args <json>     JSON array of args (stdio, optional)\n    --timeout <ms>    Per-server request timeout (http, optional)\n    --scope user|project  Default: user\n    --enabled true|false  Default: true\n    --cwd <path>      Required when scope=project\n  --set-mcp-preset    Install a named MCP preset (e.g. cua)\n    --preset cua      Cua Driver desktop computer-use (MCP)\n    --preset unreal-mcp  Unreal Engine 5.8+ editor (MCP over HTTP)\n    --scope user|project  Default: user\n    --cwd <path>      Required when scope=project\n  --remove-mcp        Remove an MCP server entry\n    --name <id> --scope user|project [--cwd <path>]\n  --print-skills      Print skills (builtin + user + project)\n    --cwd <path>      Project root for .zelari/skills\n  --set-skill         Add/update a SKILL.md skill\n    --name <id>       Skill id (required)\n    --description <t> One-line description (required)\n    --body <text>     Markdown instructions (required)\n    --category <c>    plan|refactor|debug|review|test|docs|ops|git|db|maint\n    --tools <csv>     Comma-separated required tools (optional)\n    --cost <l>        low|medium|high (default: medium)\n    --scope user|project  Default: user\n    --cwd <path>      Required when scope=project\n  --remove-skill      Remove a user/project SKILL.md\n    --name <id> --scope user|project [--cwd <path>]\n  --generate-skill-from-url  Fetch URL + draft skill via model (JSON)\n    --url <https://...>  Required\n    --provider <id>   Override active provider\n    --model <name>    Override model for the selected provider\n  --print-ssh-targets Print SSH deploy/monitor targets\n  --set-ssh-target    Upsert target (--json '{...}' or flags)\n  --remove-ssh-target --id <id>\n  --test-ssh-target   --id <id>  (BatchMode ssh true)\n  --print-ssh-pubkey  --path <private-or-.pub>  (display public key)\n\nEnvironment:\n  ZELARI_NO_WIZARD=1    Skip the first-run wizard\n  ZELARI_SKIP_PREFLIGHT=1  Skip the boot prerequisite check\n  ZELARI_NO_PLUGIN_PROMPT=1  Skip the boot plugin-install prompt\n  ZELARI_MEMORY_MCP=1  Enable the optional external memory MCP server\n  ANATHEMA_DEV=1        Disable background update check + preflight\n"
     );
     process.exit(0);
   }
@@ -80311,6 +80772,8 @@ function pickRootComponent() {
       };
       const name = get("--name");
       const command = get("--command");
+      const url2 = get("--url");
+      const timeoutRaw = get("--timeout");
       const scopeRaw = get("--scope") ?? "user";
       const scope = scopeRaw === "project" ? "project" : "user";
       const cwd = get("--cwd") ?? process.cwd();
@@ -80323,14 +80786,22 @@ function pickRootComponent() {
         if (!Array.isArray(parsed)) throw new Error("--args must be a JSON array");
         args = parsed.map(String);
       }
-      if (!name || !command) {
-        throw new Error("--name and --command are required");
+      const timeoutMs2 = timeoutRaw !== void 0 ? Number(timeoutRaw) : void 0;
+      if (timeoutMs2 !== void 0 && (!Number.isFinite(timeoutMs2) || timeoutMs2 <= 0)) {
+        throw new Error("--timeout must be a positive number of milliseconds");
+      }
+      if (!name) throw new Error("--name is required");
+      if (!command && !url2) {
+        throw new Error("either --command (stdio) or --url (http) is required");
+      }
+      if (url2 && !/^https?:\/\//i.test(url2)) {
+        throw new Error("--url must be an http(s) endpoint");
       }
       const result = upsertMcpServer({
         scope,
         name,
         projectRoot: cwd,
-        config: { command, args, enabled }
+        config: url2 ? { type: "http", url: url2, timeoutMs: timeoutMs2, serial: true, enabled } : { command, args, enabled }
       });
       if (!result.ok) throw new Error(result.error);
       console.log(JSON.stringify({ ok: true, path: result.path, name, scope }));