maestro-agent-sdk 0.1.27 → 0.1.28

package/dist/index.js

@@ -11,6 +11,7 @@ export const MAESTRO_UPSTREAM_SNAPSHOT = "v0.13.0 (2026-05-07)";
 export { AIAgent } from "./core/agent.js";
 export { runConversation } from "./core/loop.js";
 export { ACTIVE_TASK_TEMPLATE, wrapCompactedSummary } from "./memory/active-task-template.js";
+export { resolveAuxModel } from "./memory/aux-model-map.js";
 // ─── Memory / compression ────────────────────────────────────────────────────
 export { compressIfNeeded } from "./memory/compressor.js";
 export { hashToolContent } from "./memory/hash.js";
@@ -20,14 +21,14 @@ export { onShutdown, runShutdown } from "./platform/lifecycle.js";
 // ─── Host integration points (dependency injection) ──────────────────────────
 export { setLogger } from "./platform/logger.js";
 export { setMcpResolver } from "./platform/mcp-config.js";
-export { applySkillAllowlist, DEFAULT_MAX_ITERATIONS, isAbortError, iterationBudgetLine, MAESTRO_DEFAULT_SKILL_KEY, maestroProvider, providerForModel, resolveSkillsDir, wrapUpOverlayLine, } from "./provider.js";
+export { applySkillAllowlist, DEFAULT_MAX_ITERATIONS, isAbortError, isTimeoutError, iterationBudgetLine, MAESTRO_DEFAULT_SKILL_KEY, maestroProvider, providerForModel, resolveSkillsDir, wrapUpOverlayLine, } from "./provider.js";
 export { AnthropicProvider, applyThinkingBudget, buildCacheableMessages, buildCacheableSystem, buildCacheableTools, detectThinkingKeyword, effortToPersonaPrompt, effortToThinkingBudget, isWrapUpZone, thinkingBudgetForTurn, } from "./providers/anthropic.js";
-export { DeepseekProvider, effortForDeepseek, translateMessagesToOpenAI, translateToolsToOpenAI, } from "./providers/deepseek.js";
 // ─── Codex Responses API (ChatGPT OAuth) ─────────────────────────────────────
 export { CodexResponsesProvider, effortForCodex, } from "./providers/codex.js";
 export { accessTokenExpiresAt, accessTokenIsExpiring, CodexAuthError, cloudflareHeaders as codexCloudflareHeaders, codexAuthPath, decodeJwtClaims, extractAccountId, readCodexAuth, refreshAccessToken, resolveAccessToken, writeRefreshedTokens, } from "./providers/codex-auth.js";
-export { parseCodexStream, } from "./providers/codex-stream.js";
+export { parseCodexStream } from "./providers/codex-stream.js";
 export { translateMessagesToResponses, translateMessageToResponsesItems, translateToolsToResponses, } from "./providers/codex-translators.js";
+export { DeepseekProvider, effortForDeepseek, translateMessagesToOpenAI, translateToolsToOpenAI, } from "./providers/deepseek.js";
 // ─── Maestro registry + top-level provider entry point ───────────────────────
 export { maestroRegistry } from "./registry.js";
 // ─── Session store ───────────────────────────────────────────────────────────

package/dist/index.js.map

@@ -1 +1 @@
-{"version":3,"file":"index.js","sourceRoot":"","sources":["../src/index.ts"],"names":[],"mappings":"AAAA;;;;;;GAMG;AAEH,OAAO,EAAE,mBAAmB,EAAE,MAAM,oBAAoB,CAAC;AACzD,MAAM,CAAC,MAAM,yBAAyB,GAAG,sBAA+B,CAAC;AAUzE,gFAAgF;AAChF,OAAO,EAAE,OAAO,EAAsB,MAAM,cAAc,CAAC;AAC3D,OAAO,EAAE,eAAe,EAAE,MAAM,aAAa,CAAC;AAI9C,OAAO,EAAE,oBAAoB,EAAE,oBAAoB,EAAE,MAAM,+BAA+B,CAAC;AAC3F,gFAAgF;AAChF,OAAO,EAAE,gBAAgB,EAAE,MAAM,qBAAqB,CAAC;AACvD,OAAO,EAAE,eAAe,EAAE,MAAM,eAAe,CAAC;AAChD,OAAO,EAAE,mBAAmB,EAAE,MAAM,mBAAmB,CAAC;AACxD,OAAO,EAAE,cAAc,EAAE,MAAM,yBAAyB,CAAC;AACzD,OAAO,EAAE,UAAU,EAAE,WAAW,EAAE,MAAM,sBAAsB,CAAC;AAC/D,gFAAgF;AAChF,OAAO,EAA2B,SAAS,EAAE,MAAM,mBAAmB,CAAC;AACvE,OAAO,EAAuC,cAAc,EAAE,MAAM,uBAAuB,CAAC;AAC5F,OAAO,EACL,mBAAmB,EACnB,sBAAsB,EACtB,YAAY,EACZ,mBAAmB,EACnB,yBAAyB,EACzB,eAAe,EACf,gBAAgB,EAChB,gBAAgB,EAChB,iBAAiB,GAClB,MAAM,YAAY,CAAC;AACpB,OAAO,EACL,iBAAiB,EACjB,mBAAmB,EACnB,sBAAsB,EACtB,oBAAoB,EACpB,mBAAmB,EACnB,qBAAqB,EACrB,qBAAqB,EACrB,sBAAsB,EACtB,YAAY,EACZ,qBAAqB,GACtB,MAAM,uBAAuB,CAAC;AAe/B,OAAO,EACL,gBAAgB,EAChB,iBAAiB,EACjB,yBAAyB,EACzB,sBAAsB,GACvB,MAAM,sBAAsB,CAAC;AAC9B,gFAAgF;AAChF,OAAO,EACL,sBAAsB,EACtB,cAAc,GACf,MAAM,mBAAmB,CAAC;AAC3B,OAAO,EACL,oBAAoB,EACpB,qBAAqB,EAErB,cAAc,EACd,iBAAiB,IAAI,sBAAsB,EAC3C,aAAa,EACb,eAAe,EACf,gBAAgB,EAChB,aAAa,EACb,kBAAkB,EAClB,kBAAkB,EAClB,oBAAoB,GACrB,MAAM,wBAAwB,CAAC;AAChC,OAAO,EACL,gBAAgB,GACjB,MAAM,0BAA0B,CAAC;AAClC,OAAO,EAIL,4BAA4B,EAC5B,gCAAgC,EAChC,yBAAyB,GAC1B,MAAM,+BAA+B,CAAC;AACvC,gFAAgF;AAChF,OAAO,EAAE,eAAe,EAAE,MAAM,YAAY,CAAC;AAC7C,gFAAgF;AAChF,OAAO,EACL,2BAA2B,EAC3B,8BAA8B,EAC9B,oBAAoB,EAGpB,aAAa,EACb,sBAAsB,EAEtB,kBAAkB,GACnB,MAAM,iBAAiB,CAAC;AACzB,OAAO,EAAE,YAAY,EAAE,MAAM,kBAAkB,CAAC;AAChD,OAAO,EAAE,gBAAgB,EAAE,MAAM,wBAAwB,CAAC;AAC1D,gFAAgF;AAChF,OAAO,EAAE,eAAe,EAAE,gBAAgB,EAAmB,MAAM,iBAAiB,CAAC;AACrF,OAAO,EAAE,QAAQ,EAAE,SAAS,EAAsB,MAAM,gBAAgB,CAAC;AACzE,gFAAgF;AAChF,OAAO,EAAE,aAAa,EAAE,YAAY,EAAmC,MAAM,eAAe,CAAC;AAE7F,OAAO,EAA2B,qBAAqB,EAAE,MAAM,yBAAyB,CAAC;AACzF,OAAO,EAAE,eAAe,EAAE,MAAM,uBAAuB,CAAC;AACxD,gFAAgF;AAChF,OAAO,EAAE,QAAQ,EAAE,cAAc,EAAE,MAAM,sBAAsB,CAAC;AAChE,OAAO,EAEL,4BAA4B,EAC5B,oBAAoB,EACpB,kBAAkB,GACnB,MAAM,iCAAiC,CAAC;AACzC,OAAO,EAAE,cAAc,EAAE,MAAM,sBAAsB,CAAC;AACtD,OAAO,EAAE,WAAW,EAAE,QAAQ,EAAE,MAAM,sBAAsB,CAAC;AAC7D,OAAO,EAAE,QAAQ,EAAE,MAAM,sBAAsB,CAAC;AAChD,OAAO,EAAE,cAAc,EAAE,MAAM,sBAAsB,CAAC;AACtD,OAAO,EAAE,mBAAmB,EAAE,MAAM,4BAA4B,CAAC;AACjE,OAAO,EAAE,oBAAoB,EAAE,MAAM,6BAA6B,CAAC;AACnE,OAAO,EACL,oBAAoB,EACpB,iBAAiB,EACjB,kBAAkB,EAClB,oBAAoB,GACrB,MAAM,uBAAuB,CAAC;AAC/B,OAAO,EAAE,YAAY,EAAE,MAAM,2BAA2B,CAAC;AACzD,OAAO,EAAE,eAAe,EAAE,MAAM,uBAAuB,CAAC;AACxD,OAAO,EAAE,oBAAoB,EAAE,mBAAmB,EAAE,MAAM,oBAAoB,CAAC;AAC/E,gFAAgF;AAChF,OAAO,EAUL,YAAY,GACb,MAAM,kBAAkB,CAAC;AAc1B,OAAO,EACL,cAAc,EACd,WAAW,EACX,qBAAqB,EACrB,gBAAgB,GACjB,MAAM,SAAS,CAAC"}
+{"version":3,"file":"index.js","sourceRoot":"","sources":["../src/index.ts"],"names":[],"mappings":"AAAA;;;;;;GAMG;AAEH,OAAO,EAAE,mBAAmB,EAAE,MAAM,oBAAoB,CAAC;AACzD,MAAM,CAAC,MAAM,yBAAyB,GAAG,sBAA+B,CAAC;AAUzE,gFAAgF;AAChF,OAAO,EAAE,OAAO,EAAsB,MAAM,cAAc,CAAC;AAC3D,OAAO,EAAE,eAAe,EAAE,MAAM,aAAa,CAAC;AAS9C,OAAO,EAAE,oBAAoB,EAAE,oBAAoB,EAAE,MAAM,+BAA+B,CAAC;AAC3F,OAAO,EAAE,eAAe,EAAE,MAAM,wBAAwB,CAAC;AACzD,gFAAgF;AAChF,OAAO,EAAE,gBAAgB,EAAE,MAAM,qBAAqB,CAAC;AACvD,OAAO,EAAE,eAAe,EAAE,MAAM,eAAe,CAAC;AAChD,OAAO,EAAE,mBAAmB,EAAE,MAAM,mBAAmB,CAAC;AACxD,OAAO,EAAE,cAAc,EAAE,MAAM,yBAAyB,CAAC;AACzD,OAAO,EAAE,UAAU,EAAE,WAAW,EAAE,MAAM,sBAAsB,CAAC;AAC/D,gFAAgF;AAChF,OAAO,EAA2B,SAAS,EAAE,MAAM,mBAAmB,CAAC;AACvE,OAAO,EAAuC,cAAc,EAAE,MAAM,uBAAuB,CAAC;AAC5F,OAAO,EACL,mBAAmB,EACnB,sBAAsB,EACtB,YAAY,EACZ,cAAc,EACd,mBAAmB,EACnB,yBAAyB,EACzB,eAAe,EACf,gBAAgB,EAChB,gBAAgB,EAChB,iBAAiB,GAClB,MAAM,YAAY,CAAC;AACpB,OAAO,EACL,iBAAiB,EACjB,mBAAmB,EACnB,sBAAsB,EACtB,oBAAoB,EACpB,mBAAmB,EACnB,qBAAqB,EACrB,qBAAqB,EACrB,sBAAsB,EACtB,YAAY,EACZ,qBAAqB,GACtB,MAAM,uBAAuB,CAAC;AAe/B,gFAAgF;AAChF,OAAO,EACL,sBAAsB,EACtB,cAAc,GACf,MAAM,mBAAmB,CAAC;AAC3B,OAAO,EACL,oBAAoB,EACpB,qBAAqB,EACrB,cAAc,EAEd,iBAAiB,IAAI,sBAAsB,EAC3C,aAAa,EACb,eAAe,EACf,gBAAgB,EAChB,aAAa,EACb,kBAAkB,EAClB,kBAAkB,EAClB,oBAAoB,GACrB,MAAM,wBAAwB,CAAC;AAChC,OAAO,EAAE,gBAAgB,EAAE,MAAM,0BAA0B,CAAC;AAC5D,OAAO,EAIL,4BAA4B,EAC5B,gCAAgC,EAChC,yBAAyB,GAC1B,MAAM,+BAA+B,CAAC;AACvC,OAAO,EACL,gBAAgB,EAChB,iBAAiB,EACjB,yBAAyB,EACzB,sBAAsB,GACvB,MAAM,sBAAsB,CAAC;AAC9B,gFAAgF;AAChF,OAAO,EAAE,eAAe,EAAE,MAAM,YAAY,CAAC;AAC7C,gFAAgF;AAChF,OAAO,EACL,2BAA2B,EAC3B,8BAA8B,EAC9B,oBAAoB,EAGpB,aAAa,EACb,sBAAsB,EAEtB,kBAAkB,GACnB,MAAM,iBAAiB,CAAC;AACzB,OAAO,EAAE,YAAY,EAAE,MAAM,kBAAkB,CAAC;AAChD,OAAO,EAAE,gBAAgB,EAAE,MAAM,wBAAwB,CAAC;AAC1D,gFAAgF;AAChF,OAAO,EAAE,eAAe,EAAE,gBAAgB,EAAmB,MAAM,iBAAiB,CAAC;AACrF,OAAO,EAAE,QAAQ,EAAE,SAAS,EAAsB,MAAM,gBAAgB,CAAC;AACzE,gFAAgF;AAChF,OAAO,EAAE,aAAa,EAAE,YAAY,EAAmC,MAAM,eAAe,CAAC;AAE7F,OAAO,EAA2B,qBAAqB,EAAE,MAAM,yBAAyB,CAAC;AACzF,OAAO,EAAE,eAAe,EAAE,MAAM,uBAAuB,CAAC;AACxD,gFAAgF;AAChF,OAAO,EAAE,QAAQ,EAAE,cAAc,EAAE,MAAM,sBAAsB,CAAC;AAChE,OAAO,EAEL,4BAA4B,EAC5B,oBAAoB,EACpB,kBAAkB,GACnB,MAAM,iCAAiC,CAAC;AACzC,OAAO,EAAE,cAAc,EAAE,MAAM,sBAAsB,CAAC;AACtD,OAAO,EAAE,WAAW,EAAE,QAAQ,EAAE,MAAM,sBAAsB,CAAC;AAC7D,OAAO,EAAE,QAAQ,EAAE,MAAM,sBAAsB,CAAC;AAChD,OAAO,EAAE,cAAc,EAAE,MAAM,sBAAsB,CAAC;AACtD,OAAO,EAAE,mBAAmB,EAAE,MAAM,4BAA4B,CAAC;AACjE,OAAO,EAAE,oBAAoB,EAAE,MAAM,6BAA6B,CAAC;AACnE,OAAO,EACL,oBAAoB,EACpB,iBAAiB,EACjB,kBAAkB,EAClB,oBAAoB,GACrB,MAAM,uBAAuB,CAAC;AAC/B,OAAO,EAAE,YAAY,EAAE,MAAM,2BAA2B,CAAC;AACzD,OAAO,EAAE,eAAe,EAAE,MAAM,uBAAuB,CAAC;AACxD,OAAO,EAAE,oBAAoB,EAAE,mBAAmB,EAAE,MAAM,oBAAoB,CAAC;AAC/E,gFAAgF;AAChF,OAAO,EAUL,YAAY,GACb,MAAM,kBAAkB,CAAC;AAc1B,OAAO,EACL,cAAc,EACd,WAAW,EACX,qBAAqB,EACrB,gBAAgB,GACjB,MAAM,SAAS,CAAC"}

package/dist/memory/active-task-template.d.ts

@@ -12,16 +12,19 @@
  *     pending**. Without an explicit "Pending" section the summary
  *     degenerates into a recap; the model loses track of the actual task
  *     it should be working on next.
- *   - The headers (Active Task / Goal / Pending / Files / Recent context)
- *     match the structure upstream uses, calibrated against many real
- *     compaction events on long sessions.
+ *   - The headers are calibrated against many real compaction events on
+ *     long sessions: they preserve the current task, durable constraints,
+ *     decisions already made, pending work, next steps, relevant files, and
+ *     recent tool context. v0.1.28 expands the schema from five to eight
+ *     headers; expect one prompt-cache miss at upgrade time, then keep the
+ *     new header order stable for future compactions.
  *
  * Upstream reference: `hermes-agent/agent/context_compressor.py`
  * (look for the "ACTIVE_TASK_SUMMARY_TEMPLATE" / "compression_system_prompt"
  * constants). We keep the schema verbatim so summaries between agents stay
  * mutually intelligible if the topic later switches via set_agent.
  */
-export declare const ACTIVE_TASK_TEMPLATE = "You are compressing a long agent conversation so the main agent can continue\nwithout losing context. Produce a single concise summary using EXACTLY these\nsection headers, in this order:\n\n## Active Task\nOne sentence: what is the agent currently working on?\n\n## Goal\nOne or two sentences: the user's overall objective in this session.\n\n## Pending\nBulleted list of unresolved items, decisions to make, or work explicitly\ndeferred. Use \"(blocked: <reason>)\" when applicable.\n\n## Files\nBulleted list of `absolute/paths` touched or referenced (read, written,\ninspected). Skip if none.\n\n## Recent context\n3\u20135 bullets capturing the most recent tool calls + their salient outputs.\nPrefer specifics (paths, line numbers, exit codes, key values) over generic\nrecaps. Skip details that have no bearing on the next step.\n\nRULES:\n- Output ONLY the five sections above, with no preamble or postscript.\n- Do NOT echo the user's words verbatim \u2014 paraphrase tightly.\n- Do NOT invent file paths or facts not present in the transcript.\n- Keep the entire summary under 1500 words.";
+export declare const ACTIVE_TASK_TEMPLATE = "You are compressing a long agent conversation so the main agent can continue\nwithout losing context. Produce a single concise summary using EXACTLY these\nsection headers, in this order:\n\n## Active Task\nOne sentence: what is the agent currently working on?\n\n## Goal\nOne or two sentences: the user's overall objective in this session.\n\n## Constraints\nBulleted list of durable requirements, user preferences, technical limits, or\nprocess rules that should continue to govern the work. Skip if none.\n\n## Key Decisions\nBulleted list of decisions already made that should not be reopened unless the\nuser asks. Include the rationale when it is short and important. Skip if none.\n\n## Pending\nBulleted list of unresolved items, decisions to make, or work explicitly\ndeferred. Use \"(blocked: <reason>)\" when applicable.\n\n## Next Steps\nBulleted list of the concrete next actions the main agent should take after\ncompaction, in likely execution order. Skip if none.\n\n## Files\nBulleted list of `absolute/paths` touched or referenced (read, written,\ninspected). Skip if none.\n\n## Recent context\n3\u20135 bullets capturing the most recent tool calls + their salient outputs.\nPrefer specifics (paths, line numbers, exit codes, key values) over generic\nrecaps. Skip details that have no bearing on the next step.\n\nRULES:\n- Output ONLY the eight sections above, with no preamble or postscript.\n- Do NOT echo the user's words verbatim \u2014 paraphrase tightly.\n- Do NOT invent file paths or facts not present in the transcript.\n- Keep the entire summary under 1500 words.";
 /** Header line for the summary user message that the main loop sees in
  *  place of the compressed history. Surrounded by visible markers so the
  *  main model recognizes this as a system-injected compaction, not normal

package/dist/memory/active-task-template.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"active-task-template.d.ts","sourceRoot":"","sources":["../../src/memory/active-task-template.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;GAsBG;AAEH,eAAO,MAAM,oBAAoB,+kCA2BW,CAAC;AAE7C;;;;mEAImE;AACnE,eAAO,MAAM,qBAAqB,wBAAwB,CAAC;AAC3D,eAAO,MAAM,sBAAsB,yBAAyB,CAAC;AAE7D,8EAA8E;AAC9E,wBAAgB,oBAAoB,CAAC,OAAO,EAAE,MAAM,GAAG,MAAM,CAE5D"}
+{"version":3,"file":"active-task-template.d.ts","sourceRoot":"","sources":["../../src/memory/active-task-template.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;GAyBG;AAEH,eAAO,MAAM,oBAAoB,+jDAuCW,CAAC;AAE7C;;;;mEAImE;AACnE,eAAO,MAAM,qBAAqB,wBAAwB,CAAC;AAC3D,eAAO,MAAM,sBAAsB,yBAAyB,CAAC;AAE7D,8EAA8E;AAC9E,wBAAgB,oBAAoB,CAAC,OAAO,EAAE,MAAM,GAAG,MAAM,CAE5D"}

package/dist/memory/active-task-template.js

@@ -12,9 +12,12 @@
  *     pending**. Without an explicit "Pending" section the summary
  *     degenerates into a recap; the model loses track of the actual task
  *     it should be working on next.
- *   - The headers (Active Task / Goal / Pending / Files / Recent context)
- *     match the structure upstream uses, calibrated against many real
- *     compaction events on long sessions.
+ *   - The headers are calibrated against many real compaction events on
+ *     long sessions: they preserve the current task, durable constraints,
+ *     decisions already made, pending work, next steps, relevant files, and
+ *     recent tool context. v0.1.28 expands the schema from five to eight
+ *     headers; expect one prompt-cache miss at upgrade time, then keep the
+ *     new header order stable for future compactions.
  *
  * Upstream reference: `hermes-agent/agent/context_compressor.py`
  * (look for the "ACTIVE_TASK_SUMMARY_TEMPLATE" / "compression_system_prompt"
@@ -31,10 +34,22 @@ One sentence: what is the agent currently working on?
 ## Goal
 One or two sentences: the user's overall objective in this session.
 
+## Constraints
+Bulleted list of durable requirements, user preferences, technical limits, or
+process rules that should continue to govern the work. Skip if none.
+
+## Key Decisions
+Bulleted list of decisions already made that should not be reopened unless the
+user asks. Include the rationale when it is short and important. Skip if none.
+
 ## Pending
 Bulleted list of unresolved items, decisions to make, or work explicitly
 deferred. Use "(blocked: <reason>)" when applicable.
 
+## Next Steps
+Bulleted list of the concrete next actions the main agent should take after
+compaction, in likely execution order. Skip if none.
+
 ## Files
 Bulleted list of \`absolute/paths\` touched or referenced (read, written,
 inspected). Skip if none.
@@ -45,7 +60,7 @@ Prefer specifics (paths, line numbers, exit codes, key values) over generic
 recaps. Skip details that have no bearing on the next step.
 
 RULES:
-- Output ONLY the five sections above, with no preamble or postscript.
+- Output ONLY the eight sections above, with no preamble or postscript.
 - Do NOT echo the user's words verbatim — paraphrase tightly.
 - Do NOT invent file paths or facts not present in the transcript.
 - Keep the entire summary under 1500 words.`;

package/dist/memory/active-task-template.js.map

@@ -1 +1 @@
-{"version":3,"file":"active-task-template.js","sourceRoot":"","sources":["../../src/memory/active-task-template.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;GAsBG;AAEH,MAAM,CAAC,MAAM,oBAAoB,GAAG;;;;;;;;;;;;;;;;;;;;;;;;;;;4CA2BQ,CAAC;AAE7C;;;;mEAImE;AACnE,MAAM,CAAC,MAAM,qBAAqB,GAAG,qBAAqB,CAAC;AAC3D,MAAM,CAAC,MAAM,sBAAsB,GAAG,sBAAsB,CAAC;AAE7D,8EAA8E;AAC9E,MAAM,UAAU,oBAAoB,CAAC,OAAe;IAClD,OAAO,GAAG,qBAAqB,KAAK,OAAO,CAAC,IAAI,EAAE,KAAK,sBAAsB,EAAE,CAAC;AAClF,CAAC"}
+{"version":3,"file":"active-task-template.js","sourceRoot":"","sources":["../../src/memory/active-task-template.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;GAyBG;AAEH,MAAM,CAAC,MAAM,oBAAoB,GAAG;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;4CAuCQ,CAAC;AAE7C;;;;mEAImE;AACnE,MAAM,CAAC,MAAM,qBAAqB,GAAG,qBAAqB,CAAC;AAC3D,MAAM,CAAC,MAAM,sBAAsB,GAAG,sBAAsB,CAAC;AAE7D,8EAA8E;AAC9E,MAAM,UAAU,oBAAoB,CAAC,OAAe;IAClD,OAAO,GAAG,qBAAqB,KAAK,OAAO,CAAC,IAAI,EAAE,KAAK,sBAAsB,EAAE,CAAC;AAClF,CAAC"}

package/dist/memory/aux-model-map.d.ts

@@ -0,0 +1,49 @@
+/**
+ * Map a main model id to the cheapest sibling on the same provider for
+ * background context compaction.
+ *
+ * Why this exists
+ * ---------------
+ * `compressIfNeeded` (memory/compressor.ts) reuses the agent's main model
+ * for its summary call by default. That's fine when the main model is
+ * already a cheap/fast tier (Haiku, deepseek-flash, gpt-5.4-mini), but
+ * pathological for the heavy tiers:
+ *
+ *   - gpt-5.5: a 330K-token compaction call balloons the request body to
+ *     ~1.3 MB and routinely hits undici's 5-minute headersTimeout. The
+ *     v0.1.28 diagnostic logs caught this in production — both the main
+ *     turn *and* the aux LLM compaction call timed out on the same
+ *     response cycle, leaving the loop with a prune-only fallback.
+ *   - claude-opus: each compaction costs 4-8× a Haiku call for the same
+ *     summary quality.
+ *   - deepseek-v4-pro: same shape as gpt-5.5 — the heavy reasoning model
+ *     wasted on a structural summary pass.
+ *
+ * The fix is one indirection: pick the cheapest sibling on the *same
+ * provider* (so we don't have to swap clients) and route compaction
+ * through it. Cross-provider swaps (e.g. gpt-5.5 main + Haiku aux) would
+ * require the host to wire a second provider client, so we stay
+ * intra-provider here and let hosts override via `AIAgentConfig.auxModel`
+ * if they want something fancier.
+ *
+ * Returns the input unchanged when the main model is already the cheapest
+ * sibling on its provider, or when we don't recognize it (foreign slugs
+ * pass through so a host's custom model id doesn't get silently rewritten).
+ */
+/**
+ * Resolve the aux (compaction) model id for a given main model.
+ *
+ *   - Known main model with a cheaper sibling → that sibling.
+ *   - Known main model that *is* the cheapest sibling → itself (no-op,
+ *     callers can still route compaction through it without a second
+ *     network identity).
+ *   - Unknown slug → returns the input unchanged. Lets hosts pin custom
+ *     model ids (proxies, alternate gateways) without us guessing wrong.
+ *
+ * Prefix fallbacks catch model ids that look like a known family but
+ * differ in version suffix (e.g. `claude-sonnet-4-7`, `deepseek-v5-pro`).
+ * They are conservative — only match when the slug clearly belongs to
+ * one of the three supported providers.
+ */
+export declare function resolveAuxModel(mainModel: string): string;
+//# sourceMappingURL=aux-model-map.d.ts.map

package/dist/memory/aux-model-map.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"aux-model-map.d.ts","sourceRoot":"","sources":["../../src/memory/aux-model-map.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA+BG;AAsDH;;;;;;;;;;;;;;GAcG;AACH,wBAAgB,eAAe,CAAC,SAAS,EAAE,MAAM,GAAG,MAAM,CAkBzD"}

package/dist/memory/aux-model-map.js

@@ -0,0 +1,103 @@
+/**
+ * Map a main model id to the cheapest sibling on the same provider for
+ * background context compaction.
+ *
+ * Why this exists
+ * ---------------
+ * `compressIfNeeded` (memory/compressor.ts) reuses the agent's main model
+ * for its summary call by default. That's fine when the main model is
+ * already a cheap/fast tier (Haiku, deepseek-flash, gpt-5.4-mini), but
+ * pathological for the heavy tiers:
+ *
+ *   - gpt-5.5: a 330K-token compaction call balloons the request body to
+ *     ~1.3 MB and routinely hits undici's 5-minute headersTimeout. The
+ *     v0.1.28 diagnostic logs caught this in production — both the main
+ *     turn *and* the aux LLM compaction call timed out on the same
+ *     response cycle, leaving the loop with a prune-only fallback.
+ *   - claude-opus: each compaction costs 4-8× a Haiku call for the same
+ *     summary quality.
+ *   - deepseek-v4-pro: same shape as gpt-5.5 — the heavy reasoning model
+ *     wasted on a structural summary pass.
+ *
+ * The fix is one indirection: pick the cheapest sibling on the *same
+ * provider* (so we don't have to swap clients) and route compaction
+ * through it. Cross-provider swaps (e.g. gpt-5.5 main + Haiku aux) would
+ * require the host to wire a second provider client, so we stay
+ * intra-provider here and let hosts override via `AIAgentConfig.auxModel`
+ * if they want something fancier.
+ *
+ * Returns the input unchanged when the main model is already the cheapest
+ * sibling on its provider, or when we don't recognize it (foreign slugs
+ * pass through so a host's custom model id doesn't get silently rewritten).
+ */
+import { MODEL_CODEX_GPT5_2, MODEL_CODEX_GPT5_3_CODEX, MODEL_CODEX_GPT5_4, MODEL_CODEX_GPT5_4_MINI, MODEL_CODEX_GPT5_5, MODEL_DEEPSEEK_V4_FLASH, MODEL_DEEPSEEK_V4_PRO, MODEL_OPUS, MODEL_SONNET, } from "../platform/config.js";
+const CLAUDE_HAIKU = "claude-haiku-4-5";
+/**
+ * Per-family aux model. Pinned by family, not cheapest-strictly: the host
+ * picks the model it considers the right balance of summary quality and
+ * latency for each provider line.
+ *
+ *   - Claude family → sonnet. Haiku turned out too lossy for production
+ *     summary work; sonnet is the chosen aux tier even when main is opus
+ *     (compaction is bounded so the cost differential is small) or when
+ *     main is haiku (the floor is sonnet either way).
+ *   - DeepSeek family → flash.
+ *   - Codex (gpt-5.x) family → gpt-5.4-mini.
+ *
+ * The function stays total — callers always get a usable model id back —
+ * and the same Provider client handles both main and aux, so no extra
+ * client wiring is required at the host layer.
+ */
+const AUX_MODEL_BY_MAIN = {
+    // Anthropic (Claude). Sonnet on every tier — host preference, not
+    // strictly cheapest. Haiku is intentionally avoided as an aux target
+    // because the summaries it produces dropped too much structural info
+    // for downstream resume turns.
+    [MODEL_OPUS]: MODEL_SONNET,
+    [MODEL_SONNET]: MODEL_SONNET,
+    [CLAUDE_HAIKU]: MODEL_SONNET,
+    // DeepSeek V4
+    [MODEL_DEEPSEEK_V4_PRO]: MODEL_DEEPSEEK_V4_FLASH,
+    [MODEL_DEEPSEEK_V4_FLASH]: MODEL_DEEPSEEK_V4_FLASH,
+    // Codex (gpt-5.x). gpt-5.4-mini is the cheapest tier accepted by the
+    // Responses endpoint, so every heavier slug routes to it.
+    [MODEL_CODEX_GPT5_5]: MODEL_CODEX_GPT5_4_MINI,
+    [MODEL_CODEX_GPT5_4]: MODEL_CODEX_GPT5_4_MINI,
+    [MODEL_CODEX_GPT5_4_MINI]: MODEL_CODEX_GPT5_4_MINI,
+    [MODEL_CODEX_GPT5_3_CODEX]: MODEL_CODEX_GPT5_4_MINI,
+    [MODEL_CODEX_GPT5_2]: MODEL_CODEX_GPT5_4_MINI,
+};
+/**
+ * Resolve the aux (compaction) model id for a given main model.
+ *
+ *   - Known main model with a cheaper sibling → that sibling.
+ *   - Known main model that *is* the cheapest sibling → itself (no-op,
+ *     callers can still route compaction through it without a second
+ *     network identity).
+ *   - Unknown slug → returns the input unchanged. Lets hosts pin custom
+ *     model ids (proxies, alternate gateways) without us guessing wrong.
+ *
+ * Prefix fallbacks catch model ids that look like a known family but
+ * differ in version suffix (e.g. `claude-sonnet-4-7`, `deepseek-v5-pro`).
+ * They are conservative — only match when the slug clearly belongs to
+ * one of the three supported providers.
+ */
+export function resolveAuxModel(mainModel) {
+    const exact = AUX_MODEL_BY_MAIN[mainModel];
+    if (exact !== undefined)
+        return exact;
+    // Prefix-based fallback for forward-compat with new minor versions.
+    // Any claude-* slug routes to sonnet (host preference); any heavy gpt-5.*
+    // routes to mini; any deepseek-pro variant routes to flash.
+    if (mainModel.startsWith("claude-")) {
+        return MODEL_SONNET;
+    }
+    if (mainModel === "deepseek-pro" || (mainModel.startsWith("deepseek-v") && mainModel.includes("pro"))) {
+        return MODEL_DEEPSEEK_V4_FLASH;
+    }
+    if (mainModel.startsWith("gpt-5") && !mainModel.includes("mini")) {
+        return MODEL_CODEX_GPT5_4_MINI;
+    }
+    return mainModel;
+}
+//# sourceMappingURL=aux-model-map.js.map

package/dist/memory/aux-model-map.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"aux-model-map.js","sourceRoot":"","sources":["../../src/memory/aux-model-map.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA+BG;AAEH,OAAO,EACL,kBAAkB,EAClB,wBAAwB,EACxB,kBAAkB,EAClB,uBAAuB,EACvB,kBAAkB,EAClB,uBAAuB,EACvB,qBAAqB,EACrB,UAAU,EACV,YAAY,GACb,MAAM,mBAAmB,CAAC;AAE3B,MAAM,YAAY,GAAG,kBAAkB,CAAC;AAExC;;;;;;;;;;;;;;;GAeG;AACH,MAAM,iBAAiB,GAA2B;IAChD,kEAAkE;IAClE,qEAAqE;IACrE,qEAAqE;IACrE,+BAA+B;IAC/B,CAAC,UAAU,CAAC,EAAE,YAAY;IAC1B,CAAC,YAAY,CAAC,EAAE,YAAY;IAC5B,CAAC,YAAY,CAAC,EAAE,YAAY;IAE5B,cAAc;IACd,CAAC,qBAAqB,CAAC,EAAE,uBAAuB;IAChD,CAAC,uBAAuB,CAAC,EAAE,uBAAuB;IAElD,qEAAqE;IACrE,0DAA0D;IAC1D,CAAC,kBAAkB,CAAC,EAAE,uBAAuB;IAC7C,CAAC,kBAAkB,CAAC,EAAE,uBAAuB;IAC7C,CAAC,uBAAuB,CAAC,EAAE,uBAAuB;IAClD,CAAC,wBAAwB,CAAC,EAAE,uBAAuB;IACnD,CAAC,kBAAkB,CAAC,EAAE,uBAAuB;CAC9C,CAAC;AAEF;;;;;;;;;;;;;;GAcG;AACH,MAAM,UAAU,eAAe,CAAC,SAAiB;IAC/C,MAAM,KAAK,GAAG,iBAAiB,CAAC,SAAS,CAAC,CAAC;IAC3C,IAAI,KAAK,KAAK,SAAS;QAAE,OAAO,KAAK,CAAC;IAEtC,oEAAoE;IACpE,0EAA0E;IAC1E,4DAA4D;IAC5D,IAAI,SAAS,CAAC,UAAU,CAAC,SAAS,CAAC,EAAE,CAAC;QACpC,OAAO,YAAY,CAAC;IACtB,CAAC;IACD,IAAI,SAAS,KAAK,cAAc,IAAI,CAAC,SAAS,CAAC,UAAU,CAAC,YAAY,CAAC,IAAI,SAAS,CAAC,QAAQ,CAAC,KAAK,CAAC,CAAC,EAAE,CAAC;QACtG,OAAO,uBAAuB,CAAC;IACjC,CAAC;IACD,IAAI,SAAS,CAAC,UAAU,CAAC,OAAO,CAAC,IAAI,CAAC,SAAS,CAAC,QAAQ,CAAC,MAAM,CAAC,EAAE,CAAC;QACjE,OAAO,uBAAuB,CAAC;IACjC,CAAC;IAED,OAAO,SAAS,CAAC;AACnB,CAAC"}

package/dist/memory/compressor.d.ts

@@ -1,90 +1,65 @@
 import type { Provider, ProviderMessage } from "../providers/base.js";
 /**
- * Maestro context auto-compaction.
+ * Maestro context auto-compaction (OpenCode-style incremental).
  *
  * When estimated tokens exceed `triggerRatio` × `contextWindow`, dispatch an
- * aux LLM call — by default the agent's own configured model — to summarize
- * the middle slice of the conversation into the Active Task template, then
- * return a new message array shaped as:
+ * aux LLM call to summarize the middle slice of the conversation. The
+ * resulting summary is appended to the canonical `messages` array as a
+ * compaction user + summary assistant pair so it survives persist → resume.
+ *
+ * Wire structure (returned to loop.ts):
  *
  *   [ ...head_protected, { role: "user", content: "<compacted-history>..." },
  *     ...tail_protected ]
  *
- * Head + tail protection rationale:
- *   - Head: the first user prompt + first assistant turn anchor the user's
- *     actual ask. Losing them to summarization makes the next compaction
- *     produce a recap with no goal, which cascades into hallucination.
- *   - Tail: the last few turns are the model's working memory; folding them
- *     into a summary destroys the in-progress reasoning.
- *
- * The middle is what gets compressed. We feed the aux LLM the raw
- * ProviderMessage[] slice (with the Active Task system prompt) and replace
- * it in the returned array with a single user message containing the fenced
- * summary. Anthropic's tool_use/tool_result pairing requirement is
- * preserved by snapping head/tail boundaries to message boundaries (we
- * never split a user/assistant pair).
+ * Compaction blocks in messages (persisted):
  *
- * Fallbacks:
- *   - Already below threshold → return input as-is (no LLM call).
- *   - Aux LLM call fails → fall back to pruneMessages only and log a
- *     warning. The turn proceeds; we never throw because that would break
- *     the user's in-flight conversation over an optimization.
- *   - Compacted result is *larger* than the input (degenerate aux output)
- *     → discard compaction and return pruned-only.
+ *   { role: "user", content: "\x00maestro-compaction\x00" }
+ *   { role: "assistant", content: summaryText }
  *
- * The compactor itself is pure with respect to the caller's array: it
- * builds and returns a new array. Caller's persistence path keeps the
- * canonical (uncompacted) history on disk so resume can replay the full
- * thread if desired.
+ * On re-compaction (same session or resumed), previous compaction blocks
+ * are detected, the last summary extracted, and the aux LLM receives an
+ * *incremental* prompt that asks it to update the anchored summary using
+ * ONLY the conversation delta after the last compaction. This avoids
+ * re-summarizing the entire history from scratch every time.
  *
- * Upstream reference: `hermes-agent/agent/context_compressor.py`
- * — the `should_compress(real_tokens)` → `compress_messages()` path, minus
- * the multi-provider tokenizer plumbing (we estimate; see token-estimate.ts).
+ * Head + tail protection:
+ *   - Head: first user prompt + first assistant turn.
+ *   - Tail: last 3 turns (6 messages) for working memory.
+ *   - Middle: everything else gets summarized.
  */
 export interface CompressOptions {
-    /** Model context window in tokens. Default reads `MAESTRO_CONTEXT_WINDOW`
-     *  env (Sonnet 4.6 default = 200_000). */
     contextWindow?: number;
-    /** Compaction triggers when estimated tokens / window ≥ this ratio.
-     *  Default 0.8 — matches upstream and leaves enough headroom for the
-     *  current turn's prompt + response to fit inside the cap. */
     triggerRatio?: number;
-    /** Number of HEAD messages preserved verbatim. Default 2 (first user
-     *  prompt + first assistant turn). */
     headProtect?: number;
-    /** Number of TAIL messages preserved verbatim. Default 6 (~ last 3 turns
-     *  of user/assistant alternation). */
     tailProtect?: number;
-    /** Aux model id for the summarization call. The agent loop wires the
-     *  agent's own configured model in by default, so callers usually don't
-     *  set this unless they want compaction to run on a different model than
-     *  the main turn. */
     auxModel?: string;
-    /** Inject a different provider for tests. Defaults to a fresh
-     *  `AnthropicProvider.fromEnv()` reuse-of-the-main-provider via DI. */
     auxProvider?: Provider;
-    /** Disable pruning fallback when aux LLM fails. Tests use this to verify
-     *  the fallback path exits cleanly without re-pruning. */
     disablePruneFallback?: boolean;
-    /** Abort signal for the aux summarization request. */
     abortSignal?: AbortSignal;
+    emergencyTargetTokens?: number;
+    onEmergencyTrim?: (notice: string) => void;
 }
 /**
  * Run the auto-compaction pipeline.
  *
  * Steps:
- *   1. Apply `pruneMessages` first — pass 1+2 are cheap and frequently
- *      bring the wire size below the trigger ratio on their own.
- *   2. Re-estimate tokens. If below threshold, return the pruned array.
- *   3. Otherwise slice head + tail and dispatch the aux LLM to summarize
- *      the middle. Reconstruct as [head, summary user message, tail].
- *   4. Anti-thrash: if two consecutive compactions on this array saved
- *      <10%, drop back to prune-only and stop calling the aux LLM until
- *      the caller hands us a different array reference.
+ *   1. Prune (cheap: dedup, age-summary, truncate tool output).
+ *   2. Re-estimate tokens. If below threshold, return pruned.
+ *   3. Snap head/tail boundaries. If no middle, return pruned.
+ *   4. Find previous compaction blocks → extract last summary for
+ *      incremental aux-LLM prompt.
+ *   5. Call aux LLM to summarize the middle.  When a previous summary
+ *      exists, only the *delta* after the last compaction
+ *      (messages[assistantIdx+1 .. tailStart]) is sent to the aux LLM.
+ *   6. After the degenerate-savings guard, persist compaction user +
+ *      summary assistant pair in the canonical `messages` array.
+ *   7. Return wire array: [head, wrapped-summary, tail].
  *
- * Returns a new array — caller is responsible for using the returned slice
- * on the wire and keeping the unmodified canonical history for persistence.
+ * All wire head/tail boundaries are built from a compaction-stripped
+ * view of messages so the wire never leaks internal sentinels.
  */
 export declare function compressIfNeeded(messages: ProviderMessage[], opts?: CompressOptions): Promise<ProviderMessage[]>;
+/** Test-only: reset the compactor anti-thrash WeakMap entry for an array. */
 export declare function __resetCompactorState(messages: ProviderMessage[]): void;
 //# sourceMappingURL=compressor.d.ts.map

package/dist/memory/compressor.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"compressor.d.ts","sourceRoot":"","sources":["../../src/memory/compressor.ts"],"names":[],"mappings":"AAIA,OAAO,KAAK,EAAE,QAAQ,EAAwB,eAAe,EAAE,MAAM,kBAAkB,CAAC;AAExF;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAyCG;AAEH,MAAM,WAAW,eAAe;IAC9B;8CAC0C;IAC1C,aAAa,CAAC,EAAE,MAAM,CAAC;IACvB;;kEAE8D;IAC9D,YAAY,CAAC,EAAE,MAAM,CAAC;IACtB;0CACsC;IACtC,WAAW,CAAC,EAAE,MAAM,CAAC;IACrB;0CACsC;IACtC,WAAW,CAAC,EAAE,MAAM,CAAC;IACrB;;;yBAGqB;IACrB,QAAQ,CAAC,EAAE,MAAM,CAAC;IAClB;2EACuE;IACvE,WAAW,CAAC,EAAE,QAAQ,CAAC;IACvB;8DAC0D;IAC1D,oBAAoB,CAAC,EAAE,OAAO,CAAC;IAC/B,sDAAsD;IACtD,WAAW,CAAC,EAAE,WAAW,CAAC;CAC3B;AA0BD;;;;;;;;;;;;;;;GAeG;AACH,wBAAsB,gBAAgB,CACpC,QAAQ,EAAE,eAAe,EAAE,EAC3B,IAAI,GAAE,eAAoB,GACzB,OAAO,CAAC,eAAe,EAAE,CAAC,CAoJ5B;AAoCD,wBAAgB,qBAAqB,CAAC,QAAQ,EAAE,eAAe,EAAE,GAAG,IAAI,CAEvE"}
+{"version":3,"file":"compressor.d.ts","sourceRoot":"","sources":["../../src/memory/compressor.ts"],"names":[],"mappings":"AAIA,OAAO,KAAK,EAAE,QAAQ,EAAwB,eAAe,EAAE,MAAM,kBAAkB,CAAC;AAExF;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA4BG;AAEH,MAAM,WAAW,eAAe;IAC9B,aAAa,CAAC,EAAE,MAAM,CAAC;IACvB,YAAY,CAAC,EAAE,MAAM,CAAC;IACtB,WAAW,CAAC,EAAE,MAAM,CAAC;IACrB,WAAW,CAAC,EAAE,MAAM,CAAC;IACrB,QAAQ,CAAC,EAAE,MAAM,CAAC;IAClB,WAAW,CAAC,EAAE,QAAQ,CAAC;IACvB,oBAAoB,CAAC,EAAE,OAAO,CAAC;IAC/B,WAAW,CAAC,EAAE,WAAW,CAAC;IAC1B,qBAAqB,CAAC,EAAE,MAAM,CAAC;IAC/B,eAAe,CAAC,EAAE,CAAC,MAAM,EAAE,MAAM,KAAK,IAAI,CAAC;CAC5C;AAiHD;;;;;;;;;;;;;;;;;;GAkBG;AACH,wBAAsB,gBAAgB,CACpC,QAAQ,EAAE,eAAe,EAAE,EAC3B,IAAI,GAAE,eAAoB,GACzB,OAAO,CAAC,eAAe,EAAE,CAAC,CAuK5B;AAwED,6EAA6E;AAC7E,wBAAgB,qBAAqB,CAAC,QAAQ,EAAE,eAAe,EAAE,GAAG,IAAI,CAEvE"}