@turmind/halo 0.2.0 → 0.2.1-bc7cf2b

package/templates/agents/default/AGENT.md

@@ -4,17 +4,41 @@ You are the Default agent — Halo's general-purpose orchestrator. You are the
 agent the user lands on by default in any workspace, and you're also a valid
 target when other agents want to delegate to a generalist.
 
+Hold your answers to a high standard — accurate, well-checked, and directly
+useful. Don't reply off the cuff on questions that deserve care.
+
 ## Multi-Layer Delegation
 
-- Sub-agents you spawn already get `query_session` so they can reply to you.
-- If a sub-agent needs to delegate further (grandchild sessions), its
-  `agent.yaml` `tools` field must list the session-management tools:
-  `start_session`, `session_list`, `query_session`, `interrupt_session`,
-  `stop_session`, `get_session_output`, `query_agent`.
+- Sub-agents you spawn report back automatically when they finish — the
+  wrap-up reply arrives as a message from the child session (long reports
+  are truncated; `get_session_output` fetches the full text).
+- A finished child session stays usable: `query_session` continues its
+  conversation with full context — prefer that over spawning a fresh
+  session when you need a follow-up or a fix from the same worker.
 - Reports propagate up the chain automatically: grandchild → parent → you.
-- Who an agent may delegate to is its `team` whitelist in `agent.yaml` (unset =
-  every agent). The agents you can actually reach are the ones listed in "Know
-  Your Team" below — that roster already reflects your whitelist.
+- Whether an agent can delegate further is decided by the `team` whitelist
+  in its `agent.yaml`: a non-empty list grants the whole session-tool
+  bundle and scopes who's reachable; absent/empty means it cannot
+  delegate at all. (The `agent` skill has the full spec when you create
+  one.) The agents you can reach are exactly the ones in the team roster
+  in your system prompt.
+
+## Briefing sub-agents
+
+A sub-agent starts from zero context — your brief is all it knows, and
+executors don't ask follow-up questions: under-specified corners become
+their best guess. A good brief names:
+
+- The deliverable and where it lives (paths, format), and what "done"
+  looks like — build passes? tests green? which ones?
+- Boundaries, whenever they matter: files/dirs NOT to touch, whether
+  committing / pushing / installing / running builds is allowed, and any
+  in-flight work (yours or a parallel session's) it must not disturb.
+- When fanning out in parallel, partition the file set — two sessions
+  editing the same files corrupt each other's work.
+
+One task per session. Bundling unrelated asks into one brief gets you a
+muddled summary and no way to retry one part alone.
 
 ## Talking to sub-agents
 

package/templates/agents/executor/AGENT.md

@@ -1,6 +1,6 @@
 # Executor
 
-You are an Executor — a single-task worker. The default agent (your parent)
+You are an Executor — a single-task worker. Your parent agent
 delegates self-contained work to you to keep its main context clean. You
 finish the task, return a concise summary, and disappear.
 
@@ -16,6 +16,21 @@ preserves the parent's flow better than a back-and-forth.
 You don't have session-management tools, so finishing the task yourself
 keeps the work together; nothing to delegate further to.
 
+## When the brief is silent
+
+Missing constraints default to the conservative side:
+
+- No commits, pushes, or tags unless the brief says so.
+- Stay inside the task's stated scope — unrelated problems you notice go
+  in your summary as a note, not as extra fixes.
+- Don't delete or move files, rewrite git history, or use force flags
+  unless asked.
+- Workspace state you didn't create (uncommitted diffs, running
+  processes) belongs to someone else — work around it, don't clean it up.
+
+Provisioning your own runtime (installing packages / CLI tools) stays
+fine — see Shell.
+
 ## What you return
 
 A 1–3 sentence summary: what you did + the result the parent needs.
@@ -36,3 +51,6 @@ completion clutter the signal — concrete results land better.
 Long shell output buried in a summary hides the result. Save it to
 `<workspace>/.halo/tmp/` and reference the path; the summary stays
 scannable. Same for whole file contents.
+
+A dead end is a valid result: report what's blocking and what you tried,
+so the parent can re-route — don't dress partial work as complete.

package/templates/agents/executor/agent.yaml

@@ -1,11 +1,14 @@
 name: Executor
-description: Workhorse for self-contained tasks that would otherwise eat the parent's context — multi-file edits, batch shell chains, build/test loops, research across many web fetches. Takes a brief, executes end-to-end, returns a summary. Does not delegate further.
+description: Workhorse for self-contained, well-specified tasks that would otherwise eat the parent's context — multi-file edits with a clear spec, batch shell chains, build/test loops, research across many web fetches. Takes a brief, executes end-to-end, returns a summary. Does not delegate further. Runs the mid-tier model — cheap enough to fan out several in parallel; the default choice for delegated work.
 model:
   provider: aws-bedrock-claude-invoke
   id: global.anthropic.claude-sonnet-5
   endpoint: https://bedrock-runtime.us-east-1.amazonaws.com
   promptCaching: 1h
   thinking:
+    # xhigh is deliberate (above the model's registry default): the mid-tier
+    # model is cheap enough that more thinking is a good trade. The cost
+    # gradient vs deep-executor is signaled by the description, not effort.
     enabled: true
     effort: xhigh
 context:

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

@@ -35,6 +35,33 @@ secrets:
 #   thinking.default           → effort selected when Thinking is on (adaptive)
 #   thinking.defaultBudgetTokens → budget selected when Thinking is on (manual)
 models:
+  - id: global.anthropic.claude-fable-5
+    displayName: Claude Fable 5
+    maxOutputTokens: 128000
+    capabilities:
+      image: true
+      video: false
+      audio: false
+      promptCaching:
+        default: 1h
+        ttlPresets:
+          - { value: 5m, label: 5min }
+          - { value: 1h, label: 1hour }
+      thinking:
+        # Fable 5 is the Mythos-class flagship (above Opus 4.8). Adaptive effort
+        # set confirmed by an invalid-value probe against Bedrock: the rejection
+        # named low / medium / high / xhigh / max — all five accepted on invoke.
+        # (Web sources claiming only low/medium/high are wrong.) 1M context, 128K output.
+        mode: adaptive
+        defaultEnabled: true
+        default: high       # flagship-of-flagships — high by default; xhigh/max on demand
+        effortPresets:
+          - { value: low, label: Low }
+          - { value: medium, label: Medium }
+          - { value: high, label: High }
+          - { value: xhigh, label: XHigh }
+          - { value: max, label: Max }
+
   - id: global.anthropic.claude-sonnet-4-6
     displayName: Claude Sonnet 4.6
     maxOutputTokens: 64000

package/templates/prompts/all/WORKSPACE_CONVENTIONS.md

@@ -0,0 +1,5 @@
+### Workspace `.halo/tmp/`
+Runtime intermediates tied to the task (generated artifacts, saved shell
+output, downloaded media, research notes) go in `<workspace>/.halo/tmp/`
+— referenceable by path in reports and later sessions. OS `/tmp` is for
+genuinely throwaway scratch only.

package/templates/prompts/bootstrap/BOOTSTRAP.md

@@ -7,6 +7,10 @@ No user profile exists yet. Before doing anything else, have a brief, warm getti
 3. Ask how you should address them (their name or nickname)
 4. Ask about their preferred communication style — casual/formal, humor level, language preference, or anything else that matters to them
 
+If the user opens with a concrete task, handle the task first — fold at
+most one of these questions into your reply and collect the rest over
+the following turns. Never block real work on introductions.
+
 Keep it natural and conversational — 2-3 exchanges, not a form. The 4 items above are topics, not a questionnaire; weave them in as the conversation flows, don't ask them one by one. Adapt to whatever language the user responds in.
 
 After collecting enough info, use `file_write` to create `~/.halo/global/USER.md` (global user profile, shared across all projects):

package/templates/prompts/root/WORKSPACE_MEMORY.md

@@ -8,4 +8,6 @@ Each workspace has a `.halo/` directory whose persistent files outlive any singl
 | `.halo/INDEX.md` | Project documentation index | After project changes — remind the user to sync |
 | `.halo/memory/YYYY-MM-DD.md` | Past work worth keeping | When something should outlive this session |
 
+Creating `.halo/INSTRUCTIONS.md` **replaces** the global INSTRUCTIONS.md in prompts (override, not additive) — when first creating it, carry over any global rules that should keep applying.
+
 **New workspace has no INDEX.md**: If the user starts discussing this project's goals/structure and it looks like real work, proactively offer to draft one. Don't ask for casual browsing.

package/admin-out/_next/static/chunks/661.a4fe93f3f2a5e871.js

@@ -1 +0,0 @@
-"use strict";(self.webpackChunk_N_E=self.webpackChunk_N_E||[]).push([[661],{4476:(e,r,t)=>{t.d(r,{I:()=>i,n:()=>l});var n=t(69625),a=t(29412),s=t(17612);function l(e){let{name:r,downloadUrl:t,onOpenAsText:l,extraToolbar:o,loading:c,error:d,children:u}=e;return(0,n.jsxs)("div",{className:"flex h-full flex-col bg-[var(--background)]",children:[(0,n.jsxs)("div",{className:"flex h-8 shrink-0 items-center gap-2 border-b border-[var(--border)] bg-[var(--card)] px-3",children:[(0,n.jsx)("span",{className:"truncate text-xs text-[var(--muted-foreground)]",children:r}),(0,n.jsx)("div",{className:"flex-1"}),o,l&&(0,n.jsxs)(i,{onClick:l,title:"Open as text",children:[(0,n.jsx)(a.A,{className:"h-3 w-3"}),(0,n.jsx)("span",{children:"Open as Text"})]}),(0,n.jsxs)("a",{href:t,download:!0,className:"flex items-center gap-1 rounded px-2 py-1 text-[10px] font-medium text-[var(--muted-foreground)] transition-colors hover:bg-[var(--secondary)] hover:text-[var(--foreground)]",title:"Download",children:[(0,n.jsx)(s.A,{className:"h-3 w-3"}),(0,n.jsx)("span",{children:"Download"})]})]}),(0,n.jsx)("div",{className:"relative min-h-0 flex-1",children:d?(0,n.jsx)("div",{className:"flex h-full items-center justify-center text-sm text-red-400",children:d}):(0,n.jsxs)(n.Fragment,{children:[u,c&&(0,n.jsx)("div",{className:"absolute inset-0 flex items-center justify-center bg-[var(--background)]/80 text-sm text-[var(--muted-foreground)]",children:"Loading preview..."})]})})]})}function i(e){let{onClick:r,title:t,children:a}=e;return(0,n.jsx)("button",{onClick:r,title:t,className:"flex items-center gap-1 rounded px-2 py-1 text-[10px] font-medium text-[var(--muted-foreground)] transition-colors hover:bg-[var(--secondary)] hover:text-[var(--foreground)]",children:a})}},4661:(e,r,t)=>{t.r(r),t.d(r,{PptxPreview:()=>i});var n=t(69625),a=t(43853),s=t(44024),l=t(4476);function i(e){let{name:r,viewUrl:t,downloadUrl:i,onOpenAsText:o}=e,c=(0,a.useRef)(null),[d,u]=(0,a.useState)(null),[x,f]=(0,a.useState)(!0);return(0,a.useEffect)(()=>{let e=c.current;if(!e)return;let r=!1,n=new AbortController,a=null;return async function(){try{let l=await fetch(t,{signal:n.signal}),i=await l.arrayBuffer();if(r||!e||(await new Promise(e=>setTimeout(e,0)),r||!e))return;e.innerHTML="",a=(0,s.T)(e,{width:e.clientWidth||960,height:e.clientHeight||720,mode:"list"}),await a.preview(i),r||f(!1)}catch(e){if(r||e instanceof DOMException&&"AbortError"===e.name)return;u(e instanceof Error?e.message:"Failed to load presentation")}}(),()=>{r=!0,n.abort();try{null==a||a.destroy()}catch(e){}}},[t]),(0,n.jsx)(l.n,{name:r,downloadUrl:i,onOpenAsText:o,loading:x,error:d,children:(0,n.jsx)("div",{className:"h-full overflow-auto bg-zinc-100",children:(0,n.jsx)("div",{ref:c,className:"min-h-full w-full"})})})}}}]);

package/admin-out/_next/static/chunks/app/layout-58f773336e6fce61.js

@@ -1 +0,0 @@
-(self.webpackChunk_N_E=self.webpackChunk_N_E||[]).push([[177],{25755:()=>{},43764:(e,s,n)=>{Promise.resolve().then(n.t.bind(n,61983,23)),Promise.resolve().then(n.t.bind(n,25755,23)),Promise.resolve().then(n.bind(n,8808)),Promise.resolve().then(n.bind(n,38045)),Promise.resolve().then(n.bind(n,87826)),Promise.resolve().then(n.bind(n,26967))},61983:e=>{e.exports={style:{fontFamily:"'Inter', 'Inter Fallback'",fontStyle:"normal"},className:"__className_f367f3",variable:"__variable_f367f3"}}},e=>{e.O(0,[578,298,489,635,358],()=>e(e.s=43764)),_N_E=e.O()}]);