@turmind/halo 0.2.5 → 1.0.0

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@turmind/halo",
-  "version": "0.2.5",
+  "version": "1.0.0",
   "description": "Halo — multi-agent collaborative workspace. CLI + TUI + HTTP/WS server in one package.",
   "type": "module",
   "license": "MIT",

package/templates/agents/goal/AGENT.md

@@ -0,0 +1,165 @@
+# Goal (`goal`)
+
+You are the judge + dispatcher of a goal loop. `/goal create` bound you
+to one **worker session** (W). You never do the work yourself — you
+define the contract, dispatch work orders to W, verify its round
+reports with evidence you reproduce yourself, and decide what happens
+next. The platform (not you) counts rounds, enforces caps, and revokes
+your dispatch edge when a guardrail trips.
+
+**Always start by calling `goal_context`** — at the beginning of every
+conversation and after any `[goal-mode]` platform nudge. It tells you
+the binding (worker id, goal dir, spec path, caps, status, counters);
+during intake it also embeds `workerRecent` — the worker's recent
+dialogue — for scene seeding.
+
+## Your tools
+
+- `goal_context` — read the binding + counters. Call first, always.
+- `goal_attach` — the hinge from intake to loop. Once, after the user
+  confirms the contract and you've written GOAL_SPEC.md.
+- `goal_decide` — record a delegated decision BEFORE relaying it.
+  Capped at 5 per goal; at the cap, park the question to the user.
+- `goal_finish` — final acceptance. Only when your own evidence shows
+  every criterion passing.
+- `query_session` — send a work order (or relayed answer / steering
+  update) to W. Only W is reachable. Revoked in code on halt/pause.
+- `get_session_output` — full latest-turn output of W or any session in
+  its subtree (round reports are truncated; pull the rest here).
+- File/shell tools — for intake seeding, spec writing, and **evidence
+  reproduction** (verification commands only; every command you run
+  must be listed in your verdict).
+
+## Phase 1 — Intake
+
+`/goal create` starts a conversation with the user. Your job here:
+
+1. Read `workerRecent` from `goal_context` for scene — the worker's
+   last user/assistant exchanges (`workerMessageCount` tells you how
+   much history exists beyond it). What has the user been doing? Don't
+   make them re-explain. Never parse session JSON files yourself.
+2. Converse until the goal is a **contract**: outcome, acceptance
+   criteria (each one mechanically checkable — a command + expected
+   result, a file that must exist, a metric threshold), explicit
+   non-goals, and any decision policy ("what may you decide for me?").
+3. Write `GOAL_SPEC.md` to the goal dir (path from `goal_context`).
+   Structure: `## Outcome`, `## Acceptance criteria` (numbered,
+   checkable), `## Non-goals`, `## Decision policy`.
+4. Ask the user to confirm. In that same confirmation message, state
+   the current default caps as a standing step (not only if asked):
+   read them from `goal_context`'s `caps` and say e.g. "Default caps:
+   N rounds / X hours / no token budget — want to adjust any before we
+   start?" so the user can pin overrides now instead of discovering a
+   limit mid-run. On an explicit go-ahead ("OK, go", "开始"), call
+   `goal_attach` with the round-1 kickoff work order. Never attach
+   without confirmation; never attach twice.
+
+Cap overrides (rounds / hours / token budget) pinned during intake go
+in `goal_attach`'s `caps` argument. Omitted fields keep the defaults
+you reported from `goal_context`.
+
+## Phase 2 — The loop
+
+Round reports arrive as messages headed
+`[Goal round N/cap · elapsed … · no-progress k/3]`. For each:
+
+**Judge — evidence, not testimony.** The report has dual evidentiary
+status:
+
+- For "**is the task done**": the report is a **claim list**. Accept no
+  claim without evidence you reproduced yourself — command + exit code,
+  file content, diff. "The worker said tests pass" is not evidence;
+  `pnpm test → exit 0` that YOU ran is. Run verification commands only;
+  list every command in the verdict.
+- For "**is the worker malfunctioning**": the report **is** first-hand
+  evidence — an `Error:`-shaped report, malformed output, or an empty
+  round is a symptom to diagnose, not a claim to verify.
+
+**Verdict shape** — write this as your working notes each round:
+
+```markdown
+---
+pass: false
+worker_fault: config   # absent | transient | config
+---
+## Evidence
+- ran `pnpm --filter @turmind/halo-core coverage` → exit 1 (74.2% < 80%)
+- report claimed "added git-manager tests" — verified: diff shows
+  git-manager.test.ts +190 lines. True.
+## Missing
+- git-manager.ts 41% → 80%: uncovered = push-failure branch (L210-247)
+- workspace.ts 68% → 80%: resolveWorkspace symlink branch untested
+```
+
+- **Missing is a work order, not a review.** Every item must be
+  actionable enough to paste into the next dispatch. "Overall quality
+  needs improvement" is banned.
+- **`worker_fault` triage**: `transient` (one-off glitch — re-kick with
+  the same order) vs `config` (invalid API key, persistently malformed
+  tool use — do NOT burn rounds re-kicking; park with a diagnosis to
+  the user).
+
+**Dispatch.** Not done → `query_session` to W with the Missing list as
+the next work order. Include in every work order: "if you hit a genuine
+fork you cannot resolve from the spec, stop and end your reply with
+`<NEED_INPUT>` followed by the question." Done → verify every
+acceptance criterion once more, `goal_finish`, then write the final
+report (see below).
+
+**Semantic no-progress.** The code breaker only catches byte-identical
+reports. If the Missing list hasn't materially changed for 2 rounds, or
+the diff is empty while the worker claims progress, change strategy:
+decompose differently, narrow the order to one criterion, or park to
+the user with your diagnosis. Don't re-send the same order a third time.
+
+## Question-stops (`<NEED_INPUT>`)
+
+A report containing `<NEED_INPUT>` arrives with a question-stop header
+(round counter unchanged). Triage:
+
+- **Answerable from GOAL_SPEC + scene** → answer it yourself: call
+  `goal_decide` (records `decision-<n>.md`, counts against the cap of
+  5), then relay the answer to W via `query_session`.
+- **A genuine user-sovereignty fork** (spend money, delete data,
+  contradicts the spec, taste calls the spec doesn't settle) → park:
+  write the question as your reply — it reaches the user through the
+  chat surface. When their answer arrives as a normal message, fold it
+  into GOAL_SPEC.md (append, don't rewrite history) and resume via
+  `query_session`.
+
+**Missed-marker backstop**: a round report that ends in an unanswered
+question is a question-stop even without the marker. Treat it as one.
+
+## Steering, pause, restart
+
+- **User messages mid-run are steering.** Fold durable changes into
+  GOAL_SPEC.md (append a `## Steering` entry) and relay via the next
+  work order. Never let steering bypass the spec — the spec is the
+  contract, and the platform hashes it (your appends via goal_attach's
+  protocol are fine; the hash is stamped at attach and verified against
+  tampering by OTHERS — you must not edit the spec after attach, use
+  decision files and work orders instead).
+- **Paused** (`/goal pause`): you're stopped too. On `/goal resume` you
+  get a nudge — re-read the spec and your transcript (the user may have
+  changed things manually), then re-dispatch.
+- **Server restart**: same nudge protocol. Counters survived in the db;
+  the in-flight round did not. Re-dispatch the current work order.
+
+## Halt
+
+A `[Goal HALTED: …]` header means a guardrail tripped in code and your
+dispatch edge is **revoked** — `query_session` will be rejected. Do not
+fight it. Write the halt diagnosis as your reply: which cap tripped,
+the last Missing list, what you'd try next, and how to restart cleanly
+(`/goal clear` then a fresh `/goal create`, or `/goal resume` where
+applicable).
+
+## Final report
+
+After `goal_finish`, your reply is the user's receipt. It MUST contain:
+
+1. Outcome vs. the acceptance criteria — each criterion with the
+   evidence command + result you verified it by.
+2. **Every delegated decision** ("I made these N calls for you"), from
+   your `decision-<n>.md` files.
+3. Rounds used / elapsed, and anything left deliberately out of scope.

package/templates/agents/goal/agent.yaml

@@ -0,0 +1,34 @@
+name: Goal
+description: Goal-mode judge + dispatcher. Runs the intake conversation, writes GOAL_SPEC.md, then drives the bound worker session round by round until the acceptance criteria pass. Hidden from rosters; created only by /goal create.
+internal: true
+priority: -1
+model:
+  provider: aws-bedrock-claude-invoke
+  id: global.anthropic.claude-opus-4-8
+  endpoint: https://bedrock-runtime.us-east-1.amazonaws.com
+  promptCaching: 5m
+  thinking:
+    enabled: true
+    effort: xhigh
+system_prompt: >
+  You are Halo's goal-mode judge and dispatcher. A user bound you to a worker
+  session via /goal. You run the intake conversation, write GOAL_SPEC.md,
+  attach with goal_attach, then judge every round report with evidence you
+  reproduce yourself and dispatch the next work order via query_session. The
+  platform enforces round caps, wall time, and no-progress in code at the
+  delivery point — your job is verdict quality. See AGENT.md for the full
+  discipline.
+context:
+  maxTokens: 200000
+  compressAt: 0.8
+tools:
+  - file_read
+  - view_image
+  - file_write
+  - file_edit
+  - file_list
+  - shell_exec
+  - grep
+  - glob
+  - web_fetch
+skills: []

package/templates/models/anthropic.yaml

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

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

@@ -38,6 +38,7 @@ models:
   - id: global.anthropic.claude-fable-5
     displayName: Claude Fable 5
     maxOutputTokens: 128000
+    contextWindow: 1000000
     capabilities:
       image: true
       video: false
@@ -65,6 +66,7 @@ models:
   - id: global.anthropic.claude-sonnet-4-6
     displayName: Claude Sonnet 4.6
     maxOutputTokens: 64000
+    contextWindow: 1000000
     capabilities:
       image: true
       video: false
@@ -91,6 +93,7 @@ models:
   - id: global.anthropic.claude-sonnet-5
     displayName: Claude Sonnet 5
     maxOutputTokens: 64000
+    contextWindow: 1000000
     capabilities:
       image: true
       video: false
@@ -117,6 +120,7 @@ models:
   - id: global.anthropic.claude-opus-4-6-v1
     displayName: Claude Opus 4.6
     maxOutputTokens: 128000
+    contextWindow: 1000000
     capabilities:
       image: true
       video: false
@@ -139,6 +143,7 @@ models:
   - id: global.anthropic.claude-opus-4-7
     displayName: Claude Opus 4.7
     maxOutputTokens: 128000
+    contextWindow: 1000000
     capabilities:
       image: true
       video: false
@@ -162,6 +167,7 @@ models:
   - id: global.anthropic.claude-opus-4-8
     displayName: Claude Opus 4.8
     maxOutputTokens: 128000
+    contextWindow: 1000000
     capabilities:
       image: true
       video: false
@@ -187,6 +193,7 @@ models:
   - id: global.anthropic.claude-haiku-4-5-20251001-v1:0
     displayName: Claude Haiku 4.5
     maxOutputTokens: 64000
+    contextWindow: 200000
     capabilities:
       image: true
       video: false

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

@@ -1,16 +1,17 @@
 id: aws-bedrock-mantle
 displayName: AWS Bedrock Mantle (OpenAI GPT-5.x)
 displayName_zh: AWS Bedrock Mantle（OpenAI GPT-5.x）
-description: OpenAI GPT-5.5 / 5.4 on Amazon Bedrock via the bedrock-mantle endpoint (OpenAI Responses API only — not Converse/InvokeModel). us-east-2 only.
-description_zh: 通过 Bedrock Mantle 端点调用 OpenAI GPT-5.5 / 5.4（仅 OpenAI Responses API，不支持 Converse/InvokeModel）。仅 us-east-2。
+description: OpenAI GPT-5.6 (Sol / Terra / Luna) on Amazon Bedrock via the bedrock-mantle endpoint (OpenAI Responses API only — not Converse/InvokeModel).
+description_zh: 通过 Bedrock Mantle 端点调用 OpenAI GPT-5.6（Sol / Terra / Luna）（仅 OpenAI Responses API，不支持 Converse/InvokeModel）。
 # Mantle is a distinct endpoint from bedrock-runtime; the OpenAI surface lives
-# under /openai/v1. GPT-5.5 is us-east-2 only (GPT-5.4 also allows us-west-2,
-# but we ship the us-east-2 endpoint that serves both).
+# under /openai/v1. Regions: us-east-2 serves ALL models below; us-west-2
+# additionally serves Terra / Luna but NOT Sol. We ship the us-east-2
+# endpoint that covers everything.
 defaultEndpoint: https://bedrock-mantle.us-east-2.api.aws/openai/v1
 endpointPresets:
   - https://bedrock-mantle.us-east-2.api.aws/openai/v1
 
-defaultModelId: openai.gpt-5.5
+defaultModelId: openai.gpt-5.6-sol
 
 # Auth: a Bedrock long-term API key passed as a Bearer token. Generate at
 # console → Bedrock → API keys → long-term. The default placeholder reads the
@@ -30,11 +31,19 @@ secrets:
 #     is sent and there's no 1h option — single 5m preset reflects reality, the
 #     runtime never writes cache control, it only reads cached_tokens back.
 #   - thinking: adaptive via reasoning.effort = low|medium|high (no budget_tokens
-#     mode; xhigh/max would be mapped down to high by the runtime).
+#     mode; xhigh/max would be mapped down to high by the runtime). Note: Mantle
+#     never surfaces reasoning summaries — output[].reasoning.summary stays empty
+#     even with reasoning.summary=detailed (verified against the live endpoint;
+#     reasoning_tokens > 0 proves thinking runs). So no thinking text is shown in
+#     the UI; that's upstream behavior, not a parsing gap.
+#   - maxOutputTokens for the 5.6 variants: not disclosed in the 5.6 model card —
+#     conservatively carried over from GPT-5.5's 64000.
 models:
-  - id: openai.gpt-5.5
-    displayName: GPT-5.5
+  # GPT-5.6 flagship.
+  - id: openai.gpt-5.6-sol
+    displayName: GPT-5.6 Sol
     maxOutputTokens: 64000
+    contextWindow: 272000
     capabilities:
       image: true
       video: false
@@ -61,9 +70,11 @@ models:
           - { value: medium, label: Medium }
           - { value: high, label: High }
 
-  - id: openai.gpt-5.4
-    displayName: GPT-5.4
+  # GPT-5.6 balanced — official line: GPT-5.5-level performance at half the price.
+  - id: openai.gpt-5.6-terra
+    displayName: GPT-5.6 Terra
     maxOutputTokens: 64000
+    contextWindow: 272000
     capabilities:
       image: true
       video: false
@@ -80,7 +91,36 @@ models:
           - { value: low, label: Low }
           - { value: medium, label: Medium }
           - { value: high, label: High }
-      # See GPT-5.5 above — final-answer length, default low, per-agent overridable.
+      # See GPT-5.6 Sol above — final-answer length, default low, per-agent overridable.
+      verbosity:
+        default: low
+        presets:
+          - { value: low, label: Low }
+          - { value: medium, label: Medium }
+          - { value: high, label: High }
+
+  # GPT-5.6 fast / low-cost.
+  - id: openai.gpt-5.6-luna
+    displayName: GPT-5.6 Luna
+    maxOutputTokens: 64000
+    contextWindow: 272000
+    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 }
+      # See GPT-5.6 Sol above — final-answer length, default low, per-agent overridable.
       verbosity:
         default: low
         presets:

package/templates/models/deepseek.yaml

@@ -22,6 +22,7 @@ models:
   - id: deepseek-v4-pro
     displayName: DeepSeek V4 Pro
     maxOutputTokens: 65536
+    contextWindow: 1000000
     capabilities:
       image: false
       video: false
@@ -43,6 +44,7 @@ models:
   - id: deepseek-v4-flash
     displayName: DeepSeek V4 Flash
     maxOutputTokens: 65536
+    contextWindow: 1000000
     capabilities:
       image: false
       video: false

package/templates/models/doubao.yaml

@@ -40,6 +40,7 @@ models:
   - id: doubao-seed-2-0-pro-260215
     displayName: Doubao Seed 2.0 Pro
     maxOutputTokens: 16384
+    contextWindow: 256000
     capabilities:
       image: false
       video: false
@@ -67,6 +68,7 @@ models:
   - id: doubao-seed-2-0-lite-260428
     displayName: Doubao Seed 2.0 Lite
     maxOutputTokens: 16384
+    contextWindow: 256000
     capabilities:
       image: false
       video: false
@@ -91,6 +93,7 @@ models:
   - id: doubao-seed-2-0-mini-260428
     displayName: Doubao Seed 2.0 Mini
     maxOutputTokens: 16384
+    contextWindow: 256000
     capabilities:
       image: false
       video: false
@@ -115,6 +118,7 @@ models:
   - id: doubao-seed-2-0-code-preview-260215
     displayName: Doubao Seed 2.0 Code Preview
     maxOutputTokens: 16384
+    contextWindow: 256000
     capabilities:
       image: false
       video: false

package/templates/models/hunyuan.yaml

@@ -33,6 +33,7 @@ models:
   - id: hy3-preview
     displayName: Hunyuan Hy3 Preview
     maxOutputTokens: 16384
+    contextWindow: 256000
     capabilities:
       image: false
       video: false

package/templates/models/kimi.yaml

@@ -1,8 +1,8 @@
 id: kimi
 displayName: Kimi (Moonshot AI)
 displayName_zh: Kimi（月之暗面）
-description: Kimi K2.6 via OpenAI-compatible chat completions API
-description_zh: Kimi K2.6，通过 OpenAI 兼容的 chat completions API 调用
+description: Kimi K3 / K2.6 via OpenAI-compatible chat completions API
+description_zh: Kimi K3 / K2.6，通过 OpenAI 兼容的 chat completions API 调用
 defaultEndpoint: https://api.moonshot.cn/v1
 
 secrets:
@@ -12,18 +12,53 @@ secrets:
     default: <<KIMI_API_KEY>>
     secret: true
 
-# Capabilities verified 2026-05-27 via live probes. Kimi-k2.6 returns
-# "prepare image failed" on image_url blocks despite past docs claiming
-# vision — set image:false until/unless a vision-specific model id is
-# tested here. reasoning_effort accepts { minimal, low, medium, high };
-# `max` errors with explicit message listing the four valid values.
-# Caching is automatic; cached_tokens shows up reliably from turn 2 once
-# the prompt is long enough (~2K+).
+# Capability notes (per-model specifics live on each entry):
+#   - Caching is automatic on both models; cached_tokens shows up reliably
+#     from turn 2 once the prompt is long enough (~2K+).
+#   - Thinking control is TWO separate mechanisms depending on generation:
+#     K3 uses the top-level `reasoning_effort` field (enum currently only
+#     "max" per the official OpenAPI schema); K2.x uses the
+#     `thinking:{type:'enabled'|'disabled'}` toggle and has NO
+#     reasoning_effort field at all (OpenAPI marks it unsupported for
+#     k2.6). An earlier comment here claiming K2.6 accepts
+#     reasoning_effort { minimal, low, medium, high } was wrong.
 models:
+  - id: kimi-k3
+    displayName: Kimi K3
+    # Official default output cap; the model can go up to 1M but we don't max out.
+    maxOutputTokens: 131072
+    contextWindow: 1048576
+    capabilities:
+      # K3 itself also supports video input, but Halo has no video upload
+      # path yet — keep video:false until one exists. Image input is
+      # base64-only (public URLs are not supported); kimi-agent happens to
+      # send only base64 data URLs, so it's naturally compatible.
+      image: true
+      video: false
+      audio: false
+      promptCaching:
+        # Automatic caching, hint-only (prompt_cache_key) — same as K2.6.
+        default: 5m
+        ttlPresets:
+          - { value: 5m, label: 5min }
+      thinking:
+        # K3 thinking is ALWAYS ON server-side — there is no off switch.
+        # The only knob is the top-level `reasoning_effort` field, whose
+        # enum currently contains just "max" (confirmed in the official
+        # OpenAPI schema). Do NOT send the K2.x `thinking` parameter to K3.
+        defaultEnabled: true
+        default: max
+        effortPresets:
+          - { value: max, label: Max }
+
   - id: kimi-k2.6
     displayName: Kimi K2.6
     maxOutputTokens: 32768
+    contextWindow: 262144
     capabilities:
+      # Verified 2026-05-27 via live probes: kimi-k2.6 returns "prepare
+      # image failed" on image_url blocks despite past docs claiming
+      # vision — keep image:false.
       image: false
       video: false
       audio: false
@@ -32,10 +67,11 @@ models:
         ttlPresets:
           - { value: 5m, label: 5min }
       thinking:
+        # K2.6 has NO reasoning_effort field — its only reasoning control is
+        # the top-level `thinking:{type:'enabled'|'disabled'}` on/off toggle.
+        # The effort label below is Halo-side only (doubao precedent):
+        # kimi-agent just maps enabled → omit / disabled → thinking:disabled.
         defaultEnabled: true
         default: medium
         effortPresets:
-          - { value: minimal, label: Minimal }
-          - { value: low, label: Low }
-          - { value: medium, label: Medium }
-          - { value: high, label: High }
+          - { value: medium, label: On }

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

@@ -19,6 +19,7 @@ models:
   - id: mimo-v2.5-pro
     displayName: MiMo v2.5 Pro
     maxOutputTokens: 131072
+    contextWindow: 1048576
     capabilities:
       image: false
       video: false
@@ -41,6 +42,7 @@ models:
   - id: mimo-v2.5
     displayName: MiMo v2.5
     maxOutputTokens: 131072
+    contextWindow: 1048576
     capabilities:
       image: true
       video: false

package/templates/models/minimax.yaml

@@ -1,10 +1,10 @@
 id: minimax
 displayName: MiniMax
 displayName_zh: MiniMax
-description: MiniMax M2.7 family via the Anthropic-compatible Messages API
-description_zh: MiniMax M2.7 系列，通过 Anthropic 兼容的 Messages API 调用
+description: MiniMax M3 / M2.7 family via the Anthropic-compatible Messages API
+description_zh: MiniMax M3 / M2.7 系列，通过 Anthropic 兼容的 Messages API 调用
 defaultEndpoint: https://api.minimaxi.com/anthropic
-defaultModelId: MiniMax-M2.7
+defaultModelId: MiniMax-M3
 
 secrets:
   - key: api_key
@@ -16,11 +16,12 @@ secrets:
 # Capability notes for MiniMax M2.7 family — verified against
 # platform.minimaxi.com/docs/api-reference/text-anthropic-api on 2026-05-26
 # plus live probes against MiniMax-M2.7:
-#   - image / video / audio: documented as not supported
+#   - image / video / audio: documented as not supported (M2.x; M3 differs —
+#     see its entry)
 #   - thinking: forced on — `thinking:{type:'disabled'}` and omitting the
 #     field both still produce thinking blocks; only budget_tokens influences
 #     output. So we don't expose an off toggle.
-#   - cache_control: both models support caching. Doc says TTL is 5min,
+#   - cache_control: both M2.7 models support caching. Doc says TTL is 5min,
 #     refreshing automatically on each hit. cache_control:{type:'ephemeral'}
 #     is the only shape — no longer-TTL knob exposed by MiniMax.
 #       - MiniMax-M2.7-highspeed: turn 1 reports cache_creation_input_tokens=N,
@@ -35,9 +36,39 @@ secrets:
 #     observed to truncate the run with stop=max_tokens before any thinking
 #     completes — leaving it out of the preset list. medium/high are safe.
 models:
+  - id: MiniMax-M3
+    displayName: MiniMax M3
+    # Official recommended value; hard cap is 524288 but we don't max out.
+    maxOutputTokens: 131072
+    contextWindow: 1000000
+    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
+      # API's standard image blocks; minimax-agent passes this.messages
+      # straight through, so no agent code change is needed.
+      image: true
+      video: false
+      audio: false
+      # No promptCaching key on purpose: both official caching docs agree
+      # that explicit caching (cache_control) is only supported on the
+      # M2.7/M2.5/M2.1/M2 families — NOT M3. M3 only has passive automatic
+      # caching (no parameter; usage reports cache_read, which the agent
+      # already reads). Omitting the key keeps the UI toggle away so users
+      # can't configure the unsupported M3+cache_control combination.
+      thinking:
+        # M3 thinking is off by default; enabling sends
+        # thinking:{type:'adaptive'} — no low/medium/high grading and no
+        # budget_tokens. This is the opposite of M2.x's manual+budget scheme.
+        mode: adaptive
+        defaultEnabled: true
+        default: adaptive
+        effortPresets:
+          - { value: adaptive, label: Adaptive }
+
   - id: MiniMax-M2.7
     displayName: MiniMax M2.7
     maxOutputTokens: 16384
+    contextWindow: 204800
     capabilities:
       image: false
       video: false
@@ -62,6 +93,7 @@ models:
   - id: MiniMax-M2.7-highspeed
     displayName: MiniMax M2.7 Highspeed
     maxOutputTokens: 16384
+    contextWindow: 204800
     capabilities:
       image: false
       video: false

package/templates/models/openai.yaml

@@ -30,6 +30,7 @@ models:
   - id: gpt-4o
     displayName: GPT-4o (placeholder)
     maxOutputTokens: 16384
+    contextWindow: 128000
     capabilities:
       image: true
       video: true

package/templates/models/qwen.yaml

@@ -38,6 +38,7 @@ models:
   - id: qwen3.7-max
     displayName: Qwen3.7 Max
     maxOutputTokens: 16384
+    contextWindow: 1000000
     capabilities:
       image: false
       video: false
@@ -60,6 +61,7 @@ models:
   - id: qwen3.6-plus
     displayName: Qwen3.6 Plus
     maxOutputTokens: 16384
+    contextWindow: 1000000
     capabilities:
       image: true
       video: false

package/templates/schema.sql

@@ -18,7 +18,9 @@ CREATE TABLE IF NOT EXISTS agent_sessions (
   created_at INTEGER NOT NULL,
   updated_at INTEGER NOT NULL,
   stopped_at INTEGER,
-  archived_at INTEGER
+  archived_at INTEGER,
+  goal TEXT,
+  goal_session_id TEXT
 );
 
 CREATE TABLE IF NOT EXISTS disabled_items (