@turmind/halo 1.5.1-alpha → 1.5.3-alpha

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@turmind/halo",
-  "version": "1.5.1-alpha",
+  "version": "1.5.3-alpha",
   "description": "Halo — multi-agent collaborative workspace. CLI + TUI + HTTP/WS server in one package.",
   "type": "module",
   "license": "MIT",

package/templates/agents/__apply_agent__/agent.yaml

@@ -17,8 +17,8 @@ system_prompt: >
   those are the wrapper's jobs. See AGENT.md for the full procedure. You
   are invoked by an evo wrapper process — never by a user directly.
 context:
-  maxTokens: 200000
-  compressAt: 0.8
+  maxTokens: 272000
+  compressAt: 0.9
 tools:
   - file_read
   - file_write

package/templates/agents/__evo_agent__/agent.yaml

@@ -19,8 +19,8 @@ system_prompt: >
   not by you. See AGENT.md for the full procedure. You are invoked by an evo
   wrapper process — never by a user directly.
 context:
-  maxTokens: 200000
-  compressAt: 0.8
+  maxTokens: 272000
+  compressAt: 0.9
 tools:
   - file_read
   - view_image

package/templates/agents/__score__/agent.yaml

@@ -18,8 +18,8 @@ system_prompt: >
   for the full procedure. You are invoked by an evo wrapper process — never
   by a user directly.
 context:
-  maxTokens: 200000
-  compressAt: 0.8
+  maxTokens: 272000
+  compressAt: 0.9
 tools:
   - file_read
   - file_write

package/templates/agents/deep-executor/agent.yaml

@@ -9,8 +9,8 @@ model:
     enabled: true
     effort: max
 context:
-  maxTokens: 200000
-  compressAt: 0.8
+  maxTokens: 272000
+  compressAt: 0.9
 priority: 40
 tools:
   - file_read

package/templates/agents/default/agent.yaml

@@ -15,8 +15,8 @@ model:
     enabled: true
     effort: xhigh
 context:
-  maxTokens: 200000
-  compressAt: 0.8
+  maxTokens: 272000
+  compressAt: 0.9
 priority: 99
 tools:
   - file_edit

package/templates/agents/executor/agent.yaml

@@ -12,8 +12,8 @@ model:
     enabled: true
     effort: xhigh
 context:
-  maxTokens: 200000
-  compressAt: 0.8
+  maxTokens: 272000
+  compressAt: 0.9
 priority: 50
 tools:
   - file_read

package/templates/agents/goal/agent.yaml

@@ -19,8 +19,8 @@ system_prompt: >
   delivery point — your job is verdict quality. See AGENT.md for the full
   discipline.
 context:
-  maxTokens: 200000
-  compressAt: 0.8
+  maxTokens: 272000
+  compressAt: 0.9
 tools:
   - file_read
   - view_image

package/templates/models/anthropic.yaml

@@ -27,7 +27,7 @@ models:
   - id: claude-sonnet-4-6
     displayName: Claude Sonnet 4.6 (placeholder)
     maxOutputTokens: 64000
-    contextWindow: 1000000
+    contextWindow: 272000
     capabilities:
       image: true
       video: false

package/templates/models/aws-bedrock-claude-invoke.yaml

@@ -38,7 +38,7 @@ models:
   - id: global.anthropic.claude-fable-5-1
     displayName: Claude Fable 5.1
     maxOutputTokens: 128000
-    contextWindow: 1000000
+    contextWindow: 272000
     capabilities:
       image: true
       video: false
@@ -66,7 +66,7 @@ models:
   - id: global.anthropic.claude-fable-5
     displayName: Claude Fable 5
     maxOutputTokens: 128000
-    contextWindow: 1000000
+    contextWindow: 272000
     capabilities:
       image: true
       video: false
@@ -94,7 +94,7 @@ models:
   - id: global.anthropic.claude-sonnet-4-6
     displayName: Claude Sonnet 4.6
     maxOutputTokens: 64000
-    contextWindow: 1000000
+    contextWindow: 272000
     capabilities:
       image: true
       video: false
@@ -121,7 +121,7 @@ models:
   - id: global.anthropic.claude-sonnet-5
     displayName: Claude Sonnet 5
     maxOutputTokens: 64000
-    contextWindow: 1000000
+    contextWindow: 272000
     capabilities:
       image: true
       video: false
@@ -148,7 +148,7 @@ models:
   - id: global.anthropic.claude-sonnet-5-5
     displayName: Claude Sonnet 5.5
     maxOutputTokens: 64000
-    contextWindow: 1000000
+    contextWindow: 272000
     capabilities:
       image: true
       video: false
@@ -179,7 +179,7 @@ models:
   - id: global.anthropic.claude-opus-4-6-v1
     displayName: Claude Opus 4.6
     maxOutputTokens: 128000
-    contextWindow: 1000000
+    contextWindow: 272000
     capabilities:
       image: true
       video: false
@@ -202,7 +202,7 @@ models:
   - id: global.anthropic.claude-opus-4-7
     displayName: Claude Opus 4.7
     maxOutputTokens: 128000
-    contextWindow: 1000000
+    contextWindow: 272000
     capabilities:
       image: true
       video: false
@@ -226,7 +226,7 @@ models:
   - id: global.anthropic.claude-opus-4-8
     displayName: Claude Opus 4.8
     maxOutputTokens: 128000
-    contextWindow: 1000000
+    contextWindow: 272000
     capabilities:
       image: true
       video: false
@@ -252,7 +252,7 @@ models:
   - id: global.anthropic.claude-opus-5
     displayName: Claude Opus 5
     maxOutputTokens: 128000
-    contextWindow: 1000000
+    contextWindow: 272000
     capabilities:
       image: true
       video: false
@@ -279,7 +279,7 @@ models:
   - id: global.anthropic.claude-opus-5-5
     displayName: Claude Opus 5.5
     maxOutputTokens: 128000
-    contextWindow: 1000000
+    contextWindow: 272000
     capabilities:
       image: true
       video: false

package/templates/models/aws-bedrock-mantle.yaml

@@ -75,7 +75,7 @@ models:
   - id: openai.gpt-6-astra
     displayName: GPT-6 Astra
     maxOutputTokens: 128000
-    contextWindow: 1050000
+    contextWindow: 272000
     capabilities:
       image: true
       video: false
@@ -113,7 +113,7 @@ models:
   - id: openai.gpt-6-sol
     displayName: GPT-6 Sol
     maxOutputTokens: 128000
-    contextWindow: 1050000
+    contextWindow: 272000
     capabilities:
       image: true
       video: false
@@ -147,7 +147,7 @@ models:
   - id: openai.gpt-6.1-sol
     displayName: GPT-6.1 Sol
     maxOutputTokens: 128000
-    contextWindow: 1050000
+    contextWindow: 272000
     capabilities:
       image: true
       video: false
@@ -176,7 +176,7 @@ models:
   - id: openai.gpt-6-luna
     displayName: GPT-6 Luna
     maxOutputTokens: 128000
-    contextWindow: 1050000
+    contextWindow: 272000
     capabilities:
       image: true
       video: false

package/templates/models/aws-bedrock-openai.yaml

@@ -68,7 +68,7 @@ models:
   - id: global.openai.gpt-6-sol
     displayName: GPT-6 Sol
     maxOutputTokens: 128000
-    contextWindow: 1050000
+    contextWindow: 272000
     capabilities:
       image: true
       video: false
@@ -102,7 +102,7 @@ models:
   - id: global.openai.gpt-6.1-sol
     displayName: GPT-6.1 Sol
     maxOutputTokens: 128000
-    contextWindow: 1050000
+    contextWindow: 272000
     capabilities:
       image: true
       video: false
@@ -131,7 +131,7 @@ models:
   - id: global.openai.gpt-6-astra
     displayName: GPT-6 Astra
     maxOutputTokens: 128000
-    contextWindow: 1050000
+    contextWindow: 272000
     capabilities:
       image: true
       video: false
@@ -160,7 +160,7 @@ models:
   - id: global.openai.gpt-6-luna
     displayName: GPT-6 Luna
     maxOutputTokens: 128000
-    contextWindow: 1050000
+    contextWindow: 272000
     capabilities:
       image: true
       video: false
@@ -285,7 +285,7 @@ models:
   - id: global.xai.grok-4.6
     displayName: Grok 4.6
     maxOutputTokens: 64000
-    contextWindow: 500000
+    contextWindow: 272000
     capabilities:
       image: true
       video: false
@@ -322,7 +322,7 @@ models:
   - id: global.moonshotai.kimi-k3
     displayName: Kimi K3
     maxOutputTokens: 131072
-    contextWindow: 1048576
+    contextWindow: 272000
     capabilities:
       image: true
       video: false

package/templates/models/deepseek.yaml

@@ -27,7 +27,7 @@ models:
   - id: deepseek-v4-pro
     displayName: DeepSeek V4 Pro
     maxOutputTokens: 65536
-    contextWindow: 1000000
+    contextWindow: 272000
     capabilities:
       image: false
       video: false
@@ -53,7 +53,7 @@ models:
   - id: deepseek-flash
     displayName: DeepSeek Flash
     maxOutputTokens: 65536
-    contextWindow: 1000000
+    contextWindow: 272000
     capabilities:
       image: true
       video: false

package/templates/models/kimi.yaml

@@ -27,7 +27,7 @@ models:
     displayName: Kimi K3
     # Official default output cap; the model can go up to 1M but we don't max out.
     maxOutputTokens: 131072
-    contextWindow: 1048576
+    contextWindow: 272000
     capabilities:
       # K3 itself also supports video input, but Halo has no video upload
       # path yet — keep video:false until one exists. Image input is

package/templates/models/mimo-token-plan-china.yaml

@@ -37,7 +37,7 @@ models:
   - id: mimo-v2.6-pro
     displayName: MiMo v2.6 Pro
     maxOutputTokens: 131072
-    contextWindow: 1048576
+    contextWindow: 272000
     capabilities:
       image: true
       video: false
@@ -60,7 +60,7 @@ models:
   - id: mimo-v2.6-flash
     displayName: MiMo v2.6 Flash
     maxOutputTokens: 131072
-    contextWindow: 1048576
+    contextWindow: 272000
     capabilities:
       image: true
       video: false
@@ -84,7 +84,7 @@ models:
   - id: mimo-v2.5-pro
     displayName: MiMo v2.5 Pro (deprecated 2026-10-21)
     maxOutputTokens: 131072
-    contextWindow: 1048576
+    contextWindow: 272000
     capabilities:
       image: false
       video: false
@@ -107,7 +107,7 @@ models:
   - id: mimo-v2.5
     displayName: MiMo v2.5 (deprecated 2026-10-21)
     maxOutputTokens: 131072
-    contextWindow: 1048576
+    contextWindow: 272000
     capabilities:
       image: true
       video: false

package/templates/models/minimax.yaml

@@ -40,7 +40,7 @@ models:
     displayName: MiniMax M3
     # Official recommended value; hard cap is 524288 but we don't max out.
     maxOutputTokens: 131072
-    contextWindow: 1000000
+    contextWindow: 272000
     capabilities:
       # M3 also supports video input, but Halo has no video upload path yet —
       # keep video:false until one exists. Image rides the Anthropic-compatible

package/templates/models/qwen.yaml

@@ -38,7 +38,7 @@ models:
   - id: qwen3.7-max
     displayName: Qwen3.7 Max
     maxOutputTokens: 16384
-    contextWindow: 1000000
+    contextWindow: 272000
     capabilities:
       image: false
       video: false
@@ -61,7 +61,7 @@ models:
   - id: qwen3.6-plus
     displayName: Qwen3.6 Plus
     maxOutputTokens: 16384
-    contextWindow: 1000000
+    contextWindow: 272000
     capabilities:
       image: true
       video: false

package/templates/models/zhipu.yaml

@@ -35,7 +35,7 @@ models:
   - id: glm-5.3
     displayName: GLM-5.3
     maxOutputTokens: 131072
-    contextWindow: 1000000
+    contextWindow: 272000
     capabilities:
       image: false
       video: false
@@ -56,7 +56,7 @@ models:
   - id: glm-5.3-flash
     displayName: GLM-5.3 Flash
     maxOutputTokens: 131072
-    contextWindow: 1000000
+    contextWindow: 272000
     capabilities:
       image: true
       video: false
@@ -76,7 +76,7 @@ models:
   - id: glm-5.3-flashx
     displayName: GLM-5.3 FlashX
     maxOutputTokens: 131072
-    contextWindow: 1000000
+    contextWindow: 272000
     capabilities:
       image: true
       video: false
@@ -96,7 +96,7 @@ models:
   - id: glm-5.2
     displayName: GLM-5.2
     maxOutputTokens: 131072
-    contextWindow: 1000000
+    contextWindow: 272000
     capabilities:
       image: false
       video: false

package/templates/prompts/root/DELEGATION.md

@@ -21,9 +21,10 @@ talk to you while you're busy, and fills your context with intermediate output.
 
 ### What tends to fit doing it yourself
 
-- It leans on this conversation: decisions the user made, options they ruled
-  out, constraints they mentioned in passing — and restating those in a brief
-  would take about as long as the work.
+- It is light work that leans on this conversation: decisions the user made,
+  options they ruled out, constraints they mentioned in passing — and
+  restating those in a brief would take about as long as the work. For
+  anything heavier, the context goes into the brief.
 - It is small and its cause is already known.
 - The user is waiting on an answer, or wants to steer step by step.
 - Your next step depends on the result and there is nothing else to do meanwhile.
@@ -48,6 +49,14 @@ brief leaves out, it explores or guesses.
   options, and constraints. Quote the user's wording when it matters.
 - **What you already know**: paths, line numbers, snippets, the pattern to copy.
   Findings, not pointers.
+- **Settled questions, not open branches**: anything the work depends on that
+  this conversation has already established — by you or by an earlier worker's
+  report — goes in as a stated fact or decision. "If you find X, pick the
+  simplest" hands over an investigation; settle it before sending. This is
+  about questions of intent and choice — ask the user when unsure. Technical
+  unknowns (where the change goes, how many call sites) are part of the task:
+  locate the entry point and stop, don't pre-investigate on the worker's
+  behalf.
 - **Boundaries**: files not to touch; whether commit / push / build / deploy is
   allowed; in-flight work it must not disturb.
 - **Verification**: the exact commands, scoped to what changed.
@@ -56,3 +65,11 @@ brief leaves out, it explores or guesses.
 One task per session. When fanning out, partition by file so parallel sessions
 don't edit the same ones. For a follow-up or a fix from the same worker,
 `query_session` continues it with its context intact.
+
+### Context a worker already gathered
+
+A worker's session keeps everything it read. Ask for the reusable part in its
+report — files and line ranges it touched, the call sites it traced, facts
+that differed from the brief — and when later work needs the same ground,
+`query_session` that worker (to answer, or to take the follow-up) before
+re-reading it yourself or briefing a fresh worker to rediscover it.

package/templates/skills/agent/SKILL.md

@@ -66,8 +66,8 @@ model:
 system_prompt: >
   <Brief one-paragraph role statement. Keep behavioral detail in AGENT.md.>
 context:
-  maxTokens: 200000
-  compressAt: 0.8       # auto-compact at 80%
+  maxTokens: 272000
+  compressAt: 0.9       # auto-compact at 90%
 tools:
   - file_read
   - file_write