@turmind/halo 1.4.6 → 1.5.0-alpha

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@turmind/halo",
-  "version": "1.4.6",
+  "version": "1.5.0-alpha",
   "description": "Halo — multi-agent collaborative workspace. CLI + TUI + HTTP/WS server in one package.",
   "type": "module",
   "license": "MIT",
@@ -18,7 +18,7 @@
   ],
   "dependencies": {
     "@aws-crypto/sha256-js": "^5.2.0",
-    "@aws-sdk/client-bedrock-runtime": "^3.1030.0",
+    "@aws-sdk/client-bedrock-runtime": "^3.1142.0",
     "@aws-sdk/credential-provider-node": "^3.972.30",
     "@hono/node-server": "^1.19.17",
     "@larksuiteoapi/node-sdk": "^1.65.0",

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

@@ -13,9 +13,6 @@ context:
   compressAt: 0.8
 priority: 40
 tools:
-  # `draft` matches this agent's draft-critique-revise persona (AGENT.md
-  # "What deep means here") — deliberate, unlike executor which omits it.
-  - draft
   - file_read
   - file_write
   - file_edit

package/templates/agents/default/AGENT.md

@@ -23,23 +23,6 @@ useful. Don't reply off the cuff on questions that deserve care.
   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
 
 When you start a sub-session, one sentence — "Asked X to do Y" — usually

package/templates/agents/default/agent.yaml

@@ -19,7 +19,6 @@ context:
   compressAt: 0.8
 priority: 99
 tools:
-  - draft
   - file_edit
   - file_write
   - file_read

package/templates/agents/executor/agent.yaml

@@ -2,7 +2,7 @@ name: Executor
 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
+  id: global.anthropic.claude-sonnet-5-5
   endpoint: https://bedrock-runtime.us-east-1.amazonaws.com
   promptCaching: 1h
   thinking:

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

@@ -145,6 +145,37 @@ models:
           - { value: xhigh, label: XHigh }
           - { value: max, label: Max }
 
+  - id: global.anthropic.claude-sonnet-5-5
+    displayName: Claude Sonnet 5.5
+    maxOutputTokens: 64000
+    contextWindow: 1000000
+    capabilities:
+      image: true
+      video: false
+      audio: false
+      promptCaching:
+        default: 1h
+        ttlPresets:
+          - { value: 5m, label: 5min }
+          - { value: 1h, label: 1hour }
+      thinking:
+        # Sonnet 5.5 (Bedrock GA 2026-09-28; global profile only, no us.).
+        # Live-probed 2026-09-29 on invoke: invalid-value probe lists
+        # low / medium / high / xhigh / max; max_tokens ceiling 128000
+        # (maxOutputTokens kept at Sonnet 5's 64000); smoke call (xhigh,
+        # tool_use round trip) returned normally. Adaptive thinking
+        # always-on (type disabled / enabled both rejected); image input
+        # verified. Model card: 1M context.
+        mode: adaptive
+        defaultEnabled: true
+        default: high
+        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-opus-4-6-v1
     displayName: Claude Opus 4.6
     maxOutputTokens: 128000

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

@@ -1,16 +1,17 @@
 id: aws-bedrock-mantle
-displayName: AWS Bedrock Mantle (OpenAI GPT-6 / 5.6)
-displayName_zh: AWS Bedrock Mantle（OpenAI GPT-6 / 5.6）
-description: OpenAI GPT-6 (Astra / Sol / Luna) and GPT-5.6 (Sol / Terra / Luna) on Amazon Bedrock via the bedrock-mantle endpoint (OpenAI Responses API only — not Converse/InvokeModel). Region-split — GPT-6 Sol / Luna are on us-east-1 only, Astra on us-west-2 only.
-description_zh: 通过 Bedrock Mantle 端点调用 OpenAI GPT-6（Astra / Sol / Luna）与 GPT-5.6（Sol / Terra / Luna）（仅 OpenAI Responses API，不支持 Converse/InvokeModel）。按区域分布：GPT-6 Sol / Luna 仅 us-east-1，Astra 仅 us-west-2。
+displayName: AWS Bedrock Mantle (OpenAI GPT-6.1 / 6 / 5.6)
+displayName_zh: AWS Bedrock Mantle（OpenAI GPT-6.1 / 6 / 5.6）
+description: OpenAI GPT-6.1 Sol, GPT-6 (Astra / Sol / Luna) and GPT-5.6 (Sol / Terra / Luna) on Amazon Bedrock via the bedrock-mantle endpoint (OpenAI Responses API only — not Converse/InvokeModel). Region-split — GPT-6.1 Sol and GPT-6 Sol / Luna are on us-east-1 only, Astra on us-west-2 only.
+description_zh: 通过 Bedrock Mantle 端点调用 OpenAI GPT-6.1 Sol、GPT-6（Astra / Sol / Luna）与 GPT-5.6（Sol / Terra / Luna）（仅 OpenAI Responses API，不支持 Converse/InvokeModel）。按区域分布：GPT-6.1 Sol 与 GPT-6 Sol / Luna 仅 us-east-1，Astra 仅 us-west-2。
 # Mantle is a distinct endpoint from bedrock-runtime; the OpenAI surface lives
-# under /openai/v1. Region split (live-probed 2026-09-24, 404 on the wrong
-# endpoint):
+# under /openai/v1. Region split (live-probed 2026-09-24, gpt-6.1-sol
+# 2026-09-30; 404 on the wrong endpoint):
 #
 #                us-east-1   us-east-2   us-west-2
 #   gpt-6-astra      -           -           ✓
 #   gpt-6-sol        ✓           -           -
 #   gpt-6-luna       ✓           -           -
+#   gpt-6.1-sol      ✓           -           -
 #   gpt-5.6-sol      ✓           ✓           -
 #   gpt-5.6-terra    ✓           ✓           ✓
 #   gpt-5.6-luna     ✓           ✓           ✓
@@ -19,7 +20,7 @@ description_zh: 通过 Bedrock Mantle 端点调用 OpenAI GPT-6（Astra / Sol /
 # us-east-2 configs (5.6-era) keep working; Astra users pick us-west-2 in the
 # agent form. The endpoint is a per-agent pick.
 #
-# The same six models are also in aws-bedrock-openai.yaml (bedrock-runtime
+# The same seven models are also in aws-bedrock-openai.yaml (bedrock-runtime
 # host, `global.openai.*` ids, no region split) — AWS's recommended endpoint
 # for new agents. Stay here only for Mantle-only features (server-side
 # web_search, background inference), which the web-search skill relies on.
@@ -138,6 +139,40 @@ models:
           - { value: medium, label: Medium }
           - { value: high, label: High }
 
+  # GPT-6.1 Sol (Bedrock GA 2026-09-29) — us-east-1 endpoint ONLY (404 on
+  # us-east-2 / us-west-2, live-probed 2026-09-30). reasoning.effort accepts
+  # low/medium/high/xhigh/max — `none` 400s, unlike GPT-6 Sol, so no "none"
+  # preset; text.verbosity low/medium/high; image input OK; 1,050,000 /
+  # 128,000 per the OpenAI model page.
+  - id: openai.gpt-6.1-sol
+    displayName: GPT-6.1 Sol
+    maxOutputTokens: 128000
+    contextWindow: 1050000
+    capabilities:
+      image: true
+      video: false
+      audio: false
+      promptCaching:
+        default: 5m
+        ttlPresets:
+          - { value: 5m, label: 5min }
+      thinking:
+        mode: adaptive
+        defaultEnabled: true
+        default: medium
+        effortPresets:
+          - { value: low, label: Low }
+          - { value: medium, label: Medium }
+          - { value: high, label: High }
+          - { value: xhigh, label: XHigh }
+          - { value: max, label: Max }
+      verbosity:
+        default: low
+        presets:
+          - { value: low, label: Low }
+          - { value: medium, label: Medium }
+          - { value: high, label: High }
+
   - id: openai.gpt-6-luna
     displayName: GPT-6 Luna
     maxOutputTokens: 128000

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

@@ -1,8 +1,8 @@
 id: aws-bedrock-openai
-displayName: AWS Bedrock OpenAI-compatible (GPT-6 / 5.6 · Grok 4.6 · Kimi K3)
-displayName_zh: AWS Bedrock OpenAI 兼容接口（GPT-6 / 5.6 · Grok 4.6 · Kimi K3）
-description: Frontier models on Amazon Bedrock via the bedrock-runtime endpoint's OpenAI Responses API (/openai/v1) — OpenAI GPT-6 (Astra / Sol / Luna) and GPT-5.6 (Sol / Terra / Luna), xAI Grok 4.6, Moonshot Kimi K3. Cross-Region inference profiles (global.*) only — every region serves every model, no region split. No server-side web_search (that is Mantle-only).
-description_zh: 通过 bedrock-runtime 端点的 OpenAI Responses API（/openai/v1）调用 Bedrock 上的前沿模型——OpenAI GPT-6（Astra / Sol / Luna）与 GPT-5.6（Sol / Terra / Luna）、xAI Grok 4.6、月之暗面 Kimi K3。仅支持跨区域推理 profile（global.*）——每个区域都能调所有模型，无区域分裂。不支持服务端 web_search（那是 Mantle 独有）。
+displayName: AWS Bedrock OpenAI-compatible (GPT-6.1 / 6 / 5.6 · Grok 4.6 · Kimi K3)
+displayName_zh: AWS Bedrock OpenAI 兼容接口（GPT-6.1 / 6 / 5.6 · Grok 4.6 · Kimi K3）
+description: Frontier models on Amazon Bedrock via the bedrock-runtime endpoint's OpenAI Responses API (/openai/v1) — OpenAI GPT-6.1 Sol, GPT-6 (Astra / Sol / Luna) and GPT-5.6 (Sol / Terra / Luna), xAI Grok 4.6, Moonshot Kimi K3. Cross-Region inference profiles (global.*) only — every region serves every model, no region split. No server-side web_search (that is Mantle-only).
+description_zh: 通过 bedrock-runtime 端点的 OpenAI Responses API（/openai/v1）调用 Bedrock 上的前沿模型——OpenAI GPT-6.1 Sol、GPT-6（Astra / Sol / Luna）与 GPT-5.6（Sol / Terra / Luna）、xAI Grok 4.6、月之暗面 Kimi K3。仅支持跨区域推理 profile（global.*）——每个区域都能调所有模型，无区域分裂。不支持服务端 web_search（那是 Mantle 独有）。
 # Why a second Bedrock/OpenAI provider next to aws-bedrock-mantle: this is the
 # bedrock-runtime host (https://bedrock-runtime.<region>.amazonaws.com/openai/v1),
 # which AWS now recommends for new applications; bedrock-mantle is the legacy
@@ -12,10 +12,10 @@ description_zh: 通过 bedrock-runtime 端点的 OpenAI Responses API（/openai/
 #     the bare foundation-model id 400s with "on-demand throughput isn't
 #     supported". Kimi K3 isn't on Mantle at all; Grok 4.6 is Mantle us-west-2
 #     only — here everything is reachable from any region.
-#   - live-probed 2026-09-24 with SigV4: all eight global.* ids below return
-#     200 in us-east-1 / us-west-2 / eu-west-1 / ap-northeast-1 alike, so the
-#     endpoint is purely a latency / data-residency pick — unlike Mantle's
-#     per-region GPT matrix.
+#   - live-probed 2026-09-24 with SigV4: all nine global.* ids below return
+#     200 (gpt-6.1-sol probed 2026-09-30) in us-east-1 / us-west-2 /
+#     eu-west-1 / ap-northeast-1 alike, so the endpoint is purely a latency /
+#     data-residency pick — unlike Mantle's per-region GPT matrix.
 #   - requests are always synchronous and there are NO server-side tools:
 #     `tools:[{type:'web_search'}]` → 400 "web search is not supported for
 #     this request" (verified). MantleAgent never sends those, but the
@@ -51,18 +51,20 @@ secrets:
 #     which MantleAgent doesn't send) — the single 5m preset is a form hint only.
 #   - thinking: reasoning.effort. Turning Thinking off in the form omits the
 #     field (model default) — it does not send `none`. Good thing too: on this
-#     host GPT-6 Astra 400s on `none` while every other GPT accepts it.
+#     host GPT-6 Astra and GPT-6.1 Sol 400 on `none` while every other GPT
+#     accepts it.
 #   - Reasoning summaries are never surfaced (output[].reasoning.summary
 #     empty) — same as GPT on Mantle; reasoning_tokens > 0 proves it ran.
 models:
-  # OpenAI GPT-6 / GPT-5.6 — the same six models as aws-bedrock-mantle.yaml
-  # (see that file for the per-model background: 1,050,000 / 128,000 for GPT-6,
-  # 272,000 / 64,000 for 5.6, 2x input price past 272K). Live-probed
-  # 2026-09-24 on this host with invalid-value requests: every one lists
-  # reasoning.effort none/low/medium/high/xhigh/max in its own validation
-  # error and accepts text.verbosity low/medium/high — identical surface, so
-  # presets are copied verbatim. Pick these over the Mantle entries unless you
-  # need Mantle-only features (server-side web_search, background inference).
+  # OpenAI GPT-6.1 / GPT-6 / GPT-5.6 — the same seven models as
+  # aws-bedrock-mantle.yaml (see that file for the per-model background:
+  # 1,050,000 / 128,000 for GPT-6, 272,000 / 64,000 for 5.6, 2x input price
+  # past 272K). Live-probed 2026-09-24 on this host with invalid-value
+  # requests: every one (except GPT-6.1 Sol, below) lists reasoning.effort
+  # none/low/medium/high/xhigh/max in its own validation error and accepts
+  # text.verbosity low/medium/high — identical surface, so presets are copied
+  # verbatim. Pick these over the Mantle entries unless you need Mantle-only
+  # features (server-side web_search, background inference).
   - id: global.openai.gpt-6-sol
     displayName: GPT-6 Sol
     maxOutputTokens: 128000
@@ -92,6 +94,40 @@ models:
           - { value: medium, label: Medium }
           - { value: high, label: High }
 
+  # GPT-6.1 Sol (Bedrock GA 2026-09-29). Live-probed 2026-09-30 on this host
+  # with SigV4: 200 from all four endpoint presets like the other global.* ids.
+  # reasoning.effort accepts low/medium/high/xhigh/max — `none` 400s, unlike
+  # GPT-6 Sol, so no "none" preset; text.verbosity low/medium/high; image input
+  # OK; 1,050,000 / 128,000 per the OpenAI model page.
+  - id: global.openai.gpt-6.1-sol
+    displayName: GPT-6.1 Sol
+    maxOutputTokens: 128000
+    contextWindow: 1050000
+    capabilities:
+      image: true
+      video: false
+      audio: false
+      promptCaching:
+        default: 5m
+        ttlPresets:
+          - { value: 5m, label: 5min }
+      thinking:
+        mode: adaptive
+        defaultEnabled: true
+        default: medium
+        effortPresets:
+          - { value: low, label: Low }
+          - { value: medium, label: Medium }
+          - { value: high, label: High }
+          - { value: xhigh, label: XHigh }
+          - { value: max, label: Max }
+      verbosity:
+        default: low
+        presets:
+          - { value: low, label: Low }
+          - { value: medium, label: Medium }
+          - { value: high, label: High }
+
   - id: global.openai.gpt-6-astra
     displayName: GPT-6 Astra
     maxOutputTokens: 128000

package/templates/prompts/root/DELEGATION.md

@@ -0,0 +1,58 @@
+## Delegation
+
+This applies when your tools include `start_session`. The agents you can hand
+work to are listed under "Your Team". Routing and size thresholds for this
+workspace are in its INSTRUCTIONS; this section is what holds in any workspace.
+
+A sub-session runs in parallel with your conversation and reports back when it
+finishes. While you run tools, the user's messages can only be answered between
+calls. Both choices cost something: a sub-session you didn't need costs a
+brief to write and a report to read; work you kept costs the user's ability to
+talk to you while you're busy, and fills your context with intermediate output.
+
+### What tends to fit a sub-session
+
+- The piece stands alone: it can be stated as a task with a checkable result.
+- It can run alongside other pieces, or alongside your conversation with the user.
+- Its working output (grep hits, file dumps, build logs) is large compared with
+  the answer you need back.
+- It needs an edit → build/test → fix loop that you don't need to watch.
+- A roster agent is set up for exactly this kind of work.
+
+### 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 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.
+
+### Three checks before `start_session`
+
+1. Could someone with no access to this chat do the task from the brief alone?
+   If the missing part can't be written in a few lines, do it yourself or ask
+   the user.
+2. Does it need more than one edit-and-verify round, or more than a handful of
+   tool calls? If not, doing it yourself is usually cheaper.
+3. Will the user want to talk to you in the meantime? If so, hand off the long
+   part and stay reachable.
+
+### What a brief contains
+
+The worker starts with only the brief and does not ask follow-ups. What the
+brief leaves out, it explores or guesses.
+
+- **Goal and deliverable**: what to produce, where it lives, what "done" means.
+- **Decisions already made in this conversation**: the user's choices, rejected
+  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.
+- **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.
+- **What it need not read or redo.**
+
+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.

package/templates/skills/extension/SKILL.md

@@ -0,0 +1,70 @@
+---
+name: extension
+description: Install / list / remove admin preview extensions (`~/.halo/global/extensions/<id>/`) — static viewers the admin editor loads for file types it has no built-in preview for (e.g. `.glb`). `/extension install <id|zip|url>` pulls the latest `<id>-v*` release from turmind/halo-hub by default. Activate when the user wants to preview a file type the editor can't open, or asks to install / update / remove an extension.
+command: /extension
+requiresAccess: full
+verbs:
+  - { name: install, desc: "Install or upgrade an extension from halo-hub (by id), a local zip path, or a zip URL" }
+  - { name: list,    desc: List installed extensions with version and status }
+  - { name: remove,  desc: Remove an installed extension by id }
+---
+
+# extension
+
+The requested action is **`$1`**; the rest of the line is the payload. With
+natural language ("install the glb viewer"), infer both.
+
+Preview extensions are static bundles the admin editor loads in a sandboxed
+iframe to preview file types it has no built-in viewer for (e.g. `.glb`).
+Installing = putting a validated directory under `~/.halo/global/extensions/<id>/`;
+the server watches that directory and pushes the change to every open admin —
+**no restart, no page reload**.
+
+The helper is `templates/ext.sh` in this skill's directory — the workspace copy
+wins if it exists, else the global one:
+
+- `<workspace>/.halo/skills/extension/templates/ext.sh`
+- `~/.halo/global/skills/extension/templates/ext.sh`
+
+Pick EXT = the first that exists, then:
+
+## install
+
+`bash EXT install <source>` where `<source>` is one of:
+- an extension id (e.g. `glb`) → downloads the newest `glb-v*` release asset
+  (`glb-<version>.zip`) from https://github.com/turmind/halo-hub/releases;
+- a local `.zip` path;
+- an `https://…zip` URL.
+
+The script unpacks to a temp dir, pre-checks `halo-extension.json` (id matches,
+entry exists), then atomically replaces `~/.halo/global/extensions/<id>/`.
+Reinstalling the same id upgrades (or downgrades) it. Relay the script's last
+line to the user (`installed glb 1.0.0` / the error). Never edit files inside an
+installed extension — reinstall instead.
+
+If the download fails with "no release for <id>" the GitHub API may be
+rate-limited (60 unauthenticated requests/hour) or the id may not exist: point
+the user at https://github.com/turmind/halo-hub/releases to download the zip by
+hand, then `install <path>`. Setting `GITHUB_TOKEN` in the environment lifts
+the limit.
+
+## list
+
+`bash EXT list` → one line per extension: `<id>\t<version>\t<name>` or
+`<id>\tERROR\t<reason>` for directories the server would reject. Relay as a
+short `-` list; `(none)` when empty.
+
+## remove
+
+`bash EXT remove <id>` → deletes `~/.halo/global/extensions/<id>/`. Confirm with
+the user first if they didn't name the id explicitly.
+
+## Notes
+
+- Extensions are global (all workspaces on this server). Say so if the user
+  asks "install it for this project".
+- To find what's available, point the user at https://github.com/turmind/halo-hub
+  (the `extensions/` directory); do not scrape it.
+- The script only pre-checks the required fields; the server's scanner is the
+  authority. A package it rejects shows up in `list` (and the admin) as
+  `ERROR` with the reason.

package/templates/skills/extension/templates/ext.sh

@@ -0,0 +1,112 @@
+#!/usr/bin/env bash
+# Install / list / remove admin preview extensions under ~/.halo/global/extensions/.
+# The directory IS the install: the server watches the root and pushes changes
+# to every open admin, so nothing here talks to the server. Validation
+# authority is the server's scanner; this script only pre-checks the required
+# fields so a broken package never replaces a working install.
+#
+#   ext.sh install <id | path/to.zip | https://…zip>
+#   ext.sh list
+#   ext.sh remove <id>
+#
+# Needs curl + unzip + node. No `gh`: releases are read through the public API
+# (GITHUB_TOKEN, if set, is sent to lift the 60 req/h anonymous limit).
+set -euo pipefail
+
+ROOT="$HOME/.halo/global/extensions"
+REPO="turmind/halo-hub"
+ID_RE='^[a-z0-9][a-z0-9_-]{0,63}$'   # keep in sync with server extensions/registry.ts ID_RE
+
+cmd=${1:?usage: ext.sh install <id|zip|url> | list | remove <id>}
+shift || true
+
+# Print `<id> <version> <name>` for the extension dir $1, or fail with the
+# reason. $2 (optional) = the id the directory / request claims.
+precheck() {
+  node -e '
+    const fs = require("fs"), dir = process.argv[1], expect = process.argv[2] || ""
+    const fail = (msg) => { console.log(msg); process.exit(1) }
+    let m
+    try { m = JSON.parse(fs.readFileSync(dir + "/halo-extension.json", "utf8")) } catch (e) { fail("halo-extension.json missing or invalid: " + e.message) }
+    for (const k of ["id", "name", "version", "extensions", "entry"]) if (!m[k]) fail("missing " + k)
+    if (!/^[a-z0-9][a-z0-9_-]{0,63}$/.test(m.id)) fail("bad id: " + m.id)
+    if (expect && m.id !== expect) fail("manifest id \"" + m.id + "\" != \"" + expect + "\"")
+    let isFile = false
+    try { isFile = fs.statSync(dir + "/" + m.entry).isFile() } catch {}
+    if (!isFile) fail("entry not found: " + m.entry)
+    console.log(m.id + " " + m.version + " " + m.name)
+  ' "$1" "${2:-}"
+}
+
+# Echo the path of a local zip for source $1 (path, https URL, or hub id).
+resolve_source() {
+  local src=$1 url
+  if [[ -f $src ]]; then echo "$src"; return; fi
+  if [[ $src =~ ^https?:// ]]; then curl -fsSL "$src" -o "$TMP/pkg.zip"; echo "$TMP/pkg.zip"; return; fi
+  [[ $src =~ $ID_RE ]] || { echo "invalid id or missing file: $src" >&2; exit 1; }
+  # Newest non-draft, non-prerelease release tagged "<id>-v*" (API is newest-first).
+  url=$(curl -fsSL ${GITHUB_TOKEN:+-H "Authorization: Bearer $GITHUB_TOKEN"} \
+        "https://api.github.com/repos/$REPO/releases?per_page=100" \
+    | node -e '
+        let s = ""; process.stdin.on("data", (d) => s += d).on("end", () => {
+          const id = process.argv[1]
+          const r = JSON.parse(s).find((r) => r.tag_name.startsWith(id + "-v") && !r.draft && !r.prerelease)
+          const a = r && r.assets.find((a) => a.name.endsWith(".zip"))
+          if (!a) process.exit(2)
+          console.log(a.browser_download_url)
+        })' "$src") \
+    || { echo "no release for $src in $REPO — download the zip from https://github.com/$REPO/releases and run: install <path>" >&2; exit 1; }
+  curl -fsSL "$url" -o "$TMP/pkg.zip"
+  echo "$TMP/pkg.zip"
+}
+
+case $cmd in
+  install)
+    src=${1:?usage: ext.sh install <id|zip|url>}
+    TMP=$(mktemp -d); trap 'rm -rf "$TMP"' EXIT
+    zip=$(resolve_source "$src")
+    # Same refusals as the server's zip installer: no `..`, no absolute paths,
+    # no symlinks (zipinfo mode column starts with `l`).
+    if unzip -Z1 "$zip" | grep -Eq '(^|/)\.\.(/|$)|^/'; then echo "zip contains ../ or absolute entries" >&2; exit 1; fi
+    if unzip -Z "$zip" | grep -q '^l'; then echo "zip contains symlinks" >&2; exit 1; fi
+    mkdir -p "$TMP/x"
+    unzip -q "$zip" -d "$TMP/x"
+    # `zip -r glb.zip glb/` wraps everything in one directory — descend into it.
+    dir="$TMP/x"
+    if [[ ! -f $dir/halo-extension.json ]]; then
+      only=$(find "$dir" -mindepth 1 -maxdepth 1 | head -2)
+      if [[ $(printf '%s\n' "$only" | wc -l) -eq 1 && -f $only/halo-extension.json ]]; then dir=$only; fi
+    fi
+    expect=""
+    if [[ $src =~ $ID_RE && ! -f $src ]]; then expect=$src; fi
+    if ! out=$(precheck "$dir" "$expect"); then echo "invalid extension: $out" >&2; exit 1; fi
+    read -r id ver _ <<<"$out"
+    mkdir -p "$ROOT"
+    stage="$ROOT/.tmp-$id-$RANDOM$RANDOM"
+    mv "$dir" "$stage"
+    old=""
+    if [[ -d $ROOT/$id ]]; then old="$ROOT/.old-$id-$RANDOM$RANDOM"; mv "$ROOT/$id" "$old"; fi
+    mv "$stage" "$ROOT/$id" || { [[ -n $old ]] && mv "$old" "$ROOT/$id"; rm -rf "$stage"; exit 1; }
+    if [[ -n $old ]]; then rm -rf "$old"; fi
+    echo "installed $id $ver" ;;
+  list)
+    shopt -s nullglob
+    n=0
+    for d in "$ROOT"/*/; do
+      d=${d%/}; id=$(basename "$d"); n=$((n + 1))
+      if out=$(precheck "$d" "$id"); then
+        read -r _ ver name <<<"$out"; printf '%s\t%s\t%s\n' "$id" "$ver" "$name"
+      else
+        printf '%s\tERROR\t%s\n' "$id" "$(printf '%s' "$out" | head -1)"
+      fi
+    done
+    if [[ $n -eq 0 ]]; then echo "(none)"; fi ;;
+  remove)
+    id=${1:?usage: ext.sh remove <id>}
+    [[ $id =~ $ID_RE ]] || { echo "invalid id: $id" >&2; exit 1; }
+    [[ -d $ROOT/$id ]] || { echo "not installed: $id" >&2; exit 1; }
+    rm -rf "$ROOT/$id"
+    echo "removed $id" ;;
+  *)
+    echo "unknown command: $cmd (install | list | remove)" >&2; exit 1 ;;
+esac

package/templates/skills/halo/SKILL.md

@@ -13,6 +13,7 @@ You are running inside **Halo**, a multi-agent collaboration workspace.
 | Location | Scope | Purpose |
 |----------|-------|---------|
 | `~/.halo/global/` | Global (all projects) | User profile, global instructions, global agents & skills, system prompts, bundled docs |
+| `~/.halo/global/extensions/<id>/` | Global (all projects) | Installed admin preview extensions (static viewers for file types the editor can't open, e.g. `.glb`); managed with the `extension` skill (`/extension install|list|remove`) |
 | `<workspace>/.halo/` | Per-project | Project instructions, knowledge index, project agents & skills |
 
 `.halo/` is excluded from `grep` and `glob`. Use `file_read` with the exact path to access files inside.
@@ -97,3 +98,5 @@ Questions outside these rows aren't in the bundled docs (backend internals, REST
 ## Skills
 
 Skills are referenced by id from an agent's `agent.yaml`. The system prompt shows each skill's name + description; the full body is loaded on demand via `activate_skill`.
+
+Admin preview extensions (`.glb` and other file types the editor has no built-in viewer for) are not skills — they are static bundles under `~/.halo/global/extensions/`; the `extension` skill installs / lists / removes them.

package/templates/skills/send-file/SKILL.md

@@ -13,7 +13,7 @@ Multiple `MEDIA:` lines in one reply = multiple attachments.
 
 - Marker must be alone on a line. No leading spaces, no inline text.
 - Path must be absolute and the file must exist. Verify with `file_list` if unsure.
-- The file must live under the current workspace directory or the OS temp dir (`/tmp` on Linux/macOS). Any other path — including another workspace on the same server — is silently dropped: no attachment arrives and you get no error. Copy the file into the workspace (e.g. `.halo/tmp/`) or `/tmp` first, then reference the copy.
+- Where the file may live depends on the channel account's access level. A full-access account may send any readable absolute path. A workspace / readonly account is limited to the current workspace directory or the OS temp dir (`/tmp` on Linux/macOS); any other path — including another workspace on the same server — is not sent, and the user sees a "⚠️ 文件上传失败 / delivery failed" message naming the file. If you are not full-access, copy the file into the workspace (e.g. `.halo/tmp/`) or `/tmp` first, then reference the copy.
 - Write normal text around `MEDIA:` lines for context.
 
 ## Example

package/admin-out/_next/static/chunks/220.1ed70c874686122f.js

@@ -1 +0,0 @@
-"use strict";(self.webpackChunk_N_E=self.webpackChunk_N_E||[]).push([[220],{10912:(e,t,n)=>{function r(e,t){let n=arguments.length>2&&void 0!==arguments[2]?arguments[2]:"font-family:sans-serif;padding:40px;max-width:800px;margin:0 auto",r=window.open("","_blank");r&&(r.document.write("<!DOCTYPE html><html><head><title>".concat(e.replace(/[&<>"']/g,e=>({"&":"&amp;","<":"&lt;",">":"&gt;",'"':"&quot;","'":"&#39;"})[e]),"</title><style>body{").concat(n,"}img{max-width:100%}</style></head><body>").concat(t,"</body></html>")),r.document.close(),r.onload=()=>{r.print()})}n.d(t,{H:()=>r})},37050:(e,t,n)=>{n.d(t,{I:()=>o,n:()=>s});var r=n(41440),l=n(48434),a=n(90302),i=n(17119);function s(e){let{name:t,downloadUrl:n,onOpenAsText:s,extraToolbar:c,loading:d,error:u,children:x}=e,h=(0,i.kj)();return(0,r.jsxs)("div",{className:"flex h-full flex-col bg-[var(--background)]",children:[(0,r.jsxs)("div",{className:"flex h-8 shrink-0 items-center gap-2 border-b border-[var(--border)] bg-[var(--card)] px-3",children:[(0,r.jsx)("span",{className:"truncate text-xs text-[var(--muted-foreground)]",children:t}),(0,r.jsx)("div",{className:"flex-1"}),c,s&&(0,r.jsxs)(o,{onClick:s,title:h("editor.openAsText"),children:[(0,r.jsx)(l.A,{className:"h-3 w-3"}),(0,r.jsx)("span",{children:h("editor.openAsText")})]}),(0,r.jsxs)("a",{href:n,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:h("editor.download"),children:[(0,r.jsx)(a.A,{className:"h-3 w-3"}),(0,r.jsx)("span",{children:h("editor.download")})]})]}),(0,r.jsx)("div",{className:"relative min-h-0 flex-1",children:u?(0,r.jsx)("div",{className:"flex h-full items-center justify-center text-sm text-red-400",children:u}):(0,r.jsxs)(r.Fragment,{children:[x,d&&(0,r.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 o(e){let{onClick:t,title:n,children:l}=e;return(0,r.jsx)("button",{onClick:t,title:n,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:l})}},39250:(e,t,n)=>{n.d(t,{A:()=>r});let r=(0,n(79020).A)("Printer",[["path",{d:"M6 18H4a2 2 0 0 1-2-2v-5a2 2 0 0 1 2-2h16a2 2 0 0 1 2 2v5a2 2 0 0 1-2 2h-2",key:"143wyd"}],["path",{d:"M6 9V3a1 1 0 0 1 1-1h10a1 1 0 0 1 1 1v6",key:"1itne7"}],["rect",{x:"6",y:"14",width:"12",height:"8",rx:"1",key:"1ue0tg"}]])},48434:(e,t,n)=>{n.d(t,{A:()=>r});let r=(0,n(79020).A)("ArrowRightLeft",[["path",{d:"m16 3 4 4-4 4",key:"1x1c3m"}],["path",{d:"M20 7H4",key:"zbl0bi"}],["path",{d:"m8 21-4-4 4-4",key:"h9nckh"}],["path",{d:"M4 17h16",key:"g4d7ey"}]])},62839:(e,t,n)=>{n.r(t),n.d(t,{MediaPreview:()=>m});var r=n(41440),l=n(54604),a=n(79020);let i=(0,a.A)("ZoomOut",[["circle",{cx:"11",cy:"11",r:"8",key:"4ej97u"}],["line",{x1:"21",x2:"16.65",y1:"21",y2:"16.65",key:"13gj7c"}],["line",{x1:"8",x2:"14",y1:"11",y2:"11",key:"durymu"}]]),s=(0,a.A)("ZoomIn",[["circle",{cx:"11",cy:"11",r:"8",key:"4ej97u"}],["line",{x1:"21",x2:"16.65",y1:"21",y2:"16.65",key:"13gj7c"}],["line",{x1:"11",x2:"11",y1:"8",y2:"14",key:"1vmskp"}],["line",{x1:"8",x2:"14",y1:"11",y2:"11",key:"durymu"}]]);var o=n(23715),c=n(39250),d=n(37050),u=n(10912),x=n(27173),h=n(17119);function m(e){let{name:t,viewUrl:n,downloadUrl:a,onOpenAsText:m}=e,f=(0,h.kj)(),g=function(e){var t,n;let r=null!=(n=null==(t=e.split(".").pop())?void 0:t.toLowerCase())?n:"";return x.dj.includes(r)?"image":x.Tv.includes(r)?"video":x._H.includes(r)?"audio":null}(t),p=function(){let e=(0,l.useRef)(null),[t,n]=(0,l.useState)(1),[r,a]=(0,l.useState)({x:0,y:0}),i=(0,l.useRef)(null),[s,o]=(0,l.useState)(!1),c=(0,l.useCallback)(()=>{n(1),a({x:0,y:0})},[]),d=(0,l.useCallback)((e,t,r)=>{n(n=>{let l=Math.min(20,Math.max(.1,n*r)),i=l/n;return a(n=>({x:e-i*(e-n.x),y:t-i*(t-n.y)})),l})},[]),u=(0,l.useCallback)(t=>{let n=e.current;n&&d(n.clientWidth/2,n.clientHeight/2,t)},[d]),x=(0,l.useCallback)(t=>{var n;t.preventDefault();let r=null==(n=e.current)?void 0:n.getBoundingClientRect();r&&d(t.clientX-r.left,t.clientY-r.top,t.deltaY<0?1.15:.8695652173913044)},[d]),h=(0,l.useCallback)(e=>{0===e.button&&(e.preventDefault(),e.target.setPointerCapture(e.pointerId),i.current={startX:e.clientX,startY:e.clientY,origX:r.x,origY:r.y},o(!0))},[r]),m=(0,l.useCallback)(e=>{i.current&&a({x:i.current.origX+(e.clientX-i.current.startX),y:i.current.origY+(e.clientY-i.current.startY)})},[]),f=(0,l.useCallback)(()=>{i.current=null,o(!1)},[]);return(0,l.useEffect)(()=>{let t=e.current;if(!t)return;let n=e=>e.preventDefault();return t.addEventListener("wheel",n,{passive:!1}),()=>t.removeEventListener("wheel",n)},[]),{containerRef:e,scale:t,translate:r,dragging:s,reset:c,zoomCenter:u,onWheel:x,onPointerDown:h,onPointerMove:m,onPointerUp:f}}(),v="image"===g?(0,r.jsxs)(r.Fragment,{children:[(0,r.jsx)(d.I,{onClick:()=>p.zoomCenter(.8695652173913044),title:f("editor.zoomOut"),children:(0,r.jsx)(i,{className:"h-3 w-3"})}),(0,r.jsxs)("span",{className:"min-w-[3ch] text-center text-[10px] text-[var(--muted-foreground)]",children:[Math.round(100*p.scale),"%"]}),(0,r.jsx)(d.I,{onClick:()=>p.zoomCenter(1.15),title:f("editor.zoomIn"),children:(0,r.jsx)(s,{className:"h-3 w-3"})}),(0,r.jsx)(d.I,{onClick:p.reset,title:f("editor.resetZoom"),children:(0,r.jsx)(o.A,{className:"h-3 w-3"})}),(0,r.jsxs)(d.I,{onClick:()=>(0,u.H)(t,'<img src="'.concat(n,'" />'),"margin:0;display:flex;justify-content:center;align-items:center;min-height:100vh;background:#fff"),title:f("editor.print"),children:[(0,r.jsx)(c.A,{className:"h-3 w-3"}),(0,r.jsx)("span",{children:f("editor.print")})]})]}):null;return(0,r.jsxs)(d.n,{name:t,downloadUrl:a,onOpenAsText:m,extraToolbar:v,children:["image"===g&&(0,r.jsx)("div",{ref:p.containerRef,className:"h-full w-full overflow-hidden bg-[#1a1a1a]",style:{cursor:p.dragging?"grabbing":"grab"},onWheel:p.onWheel,onPointerDown:p.onPointerDown,onPointerMove:p.onPointerMove,onPointerUp:p.onPointerUp,onPointerCancel:p.onPointerUp,children:(0,r.jsx)("img",{src:n,alt:t,draggable:!1,style:{transformOrigin:"0 0",transform:"translate(".concat(p.translate.x,"px, ").concat(p.translate.y,"px) scale(").concat(p.scale,")")}})}),"video"===g&&(0,r.jsx)("div",{className:"flex h-full items-center justify-center bg-black p-4",children:(0,r.jsx)("video",{src:n,controls:!0,className:"max-h-full max-w-full",children:(0,r.jsx)("track",{kind:"captions"})})}),"audio"===g&&(0,r.jsx)("div",{className:"flex h-full flex-col items-center justify-center gap-4 bg-[var(--background)]",children:(0,r.jsx)("audio",{src:n,controls:!0,className:"w-full max-w-md"})})]})}}}]);