durar-ai 2026.4.4

package/docs/tools/lobster.md

@@ -0,0 +1,348 @@
+---
+title: Lobster
+summary: "Typed workflow runtime for Durar with resumable approval gates."
+read_when:
+  - You want deterministic multi-step workflows with explicit approvals
+  - You need to resume a workflow without re-running earlier steps
+---
+
+# Lobster
+
+Lobster is a workflow shell that lets Durar run multi-step tool sequences as a single, deterministic operation with explicit approval checkpoints.
+
+Lobster is one authoring layer above detached background work. For flow orchestration above individual tasks, see [Task Flow](/automation/taskflow) (`Durar tasks flow`). For the task activity ledger, see [`Durar tasks`](/automation/tasks).
+
+## Hook
+
+Your assistant can build the tools that manage itself. Ask for a workflow, and 30 minutes later you have a CLI plus pipelines that run as one call. Lobster is the missing piece: deterministic pipelines, explicit approvals, and resumable state.
+
+## Why
+
+Today, complex workflows require many back-and-forth tool calls. Each call costs tokens, and the LLM has to orchestrate every step. Lobster moves that orchestration into a typed runtime:
+
+- **One call instead of many**: Durar runs one Lobster tool call and gets a structured result.
+- **Approvals built in**: Side effects (send email, post comment) halt the workflow until explicitly approved.
+- **Resumable**: Halted workflows return a token; approve and resume without re-running everything.
+
+## Why a DSL instead of plain programs?
+
+Lobster is intentionally small. The goal is not "a new language," it's a predictable, AI-friendly pipeline spec with first-class approvals and resume tokens.
+
+- **Approve/resume is built in**: A normal program can prompt a human, but it can’t _pause and resume_ with a durable token without you inventing that runtime yourself.
+- **Determinism + auditability**: Pipelines are data, so they’re easy to log, diff, replay, and review.
+- **Constrained surface for AI**: A tiny grammar + JSON piping reduces “creative” code paths and makes validation realistic.
+- **Safety policy baked in**: Timeouts, output caps, sandbox checks, and allowlists are enforced by the runtime, not each script.
+- **Still programmable**: Each step can call any CLI or script. If you want JS/TS, generate `.lobster` files from code.
+
+## How it works
+
+Durar launches the local `lobster` CLI in **tool mode** and parses a JSON envelope from stdout.
+If the pipeline pauses for approval, the tool returns a `resumeToken` so you can continue later.
+
+## Pattern: small CLI + JSON pipes + approvals
+
+Build tiny commands that speak JSON, then chain them into a single Lobster call. (Example command names below — swap in your own.)
+
+```bash
+inbox list --json
+inbox categorize --json
+inbox apply --json
+```
+
+```json
+{
+  "action": "run",
+  "pipeline": "exec --json --shell 'inbox list --json' | exec --stdin json --shell 'inbox categorize --json' | exec --stdin json --shell 'inbox apply --json' | approve --preview-from-stdin --limit 5 --prompt 'Apply changes?'",
+  "timeoutMs": 30000
+}
+```
+
+If the pipeline requests approval, resume with the token:
+
+```json
+{
+  "action": "resume",
+  "token": "<resumeToken>",
+  "approve": true
+}
+```
+
+AI triggers the workflow; Lobster executes the steps. Approval gates keep side effects explicit and auditable.
+
+Example: map input items into tool calls:
+
+```bash
+gog.gmail.search --query 'newer_than:1d' \
+  | Durar.invoke --tool message --action send --each --item-key message --args-json '{"provider":"telegram","to":"..."}'
+```
+
+## JSON-only LLM steps (llm-task)
+
+For workflows that need a **structured LLM step**, enable the optional
+`llm-task` plugin tool and call it from Lobster. This keeps the workflow
+deterministic while still letting you classify/summarize/draft with a model.
+
+Enable the tool:
+
+```json
+{
+  "plugins": {
+    "entries": {
+      "llm-task": { "enabled": true }
+    }
+  },
+  "agents": {
+    "list": [
+      {
+        "id": "main",
+        "tools": { "allow": ["llm-task"] }
+      }
+    ]
+  }
+}
+```
+
+Use it in a pipeline:
+
+```lobster
+Durar.invoke --tool llm-task --action json --args-json '{
+  "prompt": "Given the input email, return intent and draft.",
+  "thinking": "low",
+  "input": { "subject": "Hello", "body": "Can you help?" },
+  "schema": {
+    "type": "object",
+    "properties": {
+      "intent": { "type": "string" },
+      "draft": { "type": "string" }
+    },
+    "required": ["intent", "draft"],
+    "additionalProperties": false
+  }
+}'
+```
+
+See [LLM Task](/tools/llm-task) for details and configuration options.
+
+## Workflow files (.lobster)
+
+Lobster can run YAML/JSON workflow files with `name`, `args`, `steps`, `env`, `condition`, and `approval` fields. In Durar tool calls, set `pipeline` to the file path.
+
+```yaml
+name: inbox-triage
+args:
+  tag:
+    default: "family"
+steps:
+  - id: collect
+    command: inbox list --json
+  - id: categorize
+    command: inbox categorize --json
+    stdin: $collect.stdout
+  - id: approve
+    command: inbox apply --approve
+    stdin: $categorize.stdout
+    approval: required
+  - id: execute
+    command: inbox apply --execute
+    stdin: $categorize.stdout
+    condition: $approve.approved
+```
+
+Notes:
+
+- `stdin: $step.stdout` and `stdin: $step.json` pass a prior step’s output.
+- `condition` (or `when`) can gate steps on `$step.approved`.
+
+## Install Lobster
+
+Install the Lobster CLI on the **same host** that runs the Durar Gateway (see the [Lobster repo](https://github.com/Durar/lobster)), and ensure `lobster` is on `PATH`.
+
+## Enable the tool
+
+Lobster is an **optional** plugin tool (not enabled by default).
+
+Recommended (additive, safe):
+
+```json
+{
+  "tools": {
+    "alsoAllow": ["lobster"]
+  }
+}
+```
+
+Or per-agent:
+
+```json
+{
+  "agents": {
+    "list": [
+      {
+        "id": "main",
+        "tools": {
+          "alsoAllow": ["lobster"]
+        }
+      }
+    ]
+  }
+}
+```
+
+Avoid using `tools.allow: ["lobster"]` unless you intend to run in restrictive allowlist mode.
+
+Note: allowlists are opt-in for optional plugins. If your allowlist only names
+plugin tools (like `lobster`), Durar keeps core tools enabled. To restrict core
+tools, include the core tools or groups you want in the allowlist too.
+
+## Example: Email triage
+
+Without Lobster:
+
+```
+User: "Check my email and draft replies"
+→ Durar calls gmail.list
+→ LLM summarizes
+→ User: "draft replies to #2 and #5"
+→ LLM drafts
+→ User: "send #2"
+→ Durar calls gmail.send
+(repeat daily, no memory of what was triaged)
+```
+
+With Lobster:
+
+```json
+{
+  "action": "run",
+  "pipeline": "email.triage --limit 20",
+  "timeoutMs": 30000
+}
+```
+
+Returns a JSON envelope (truncated):
+
+```json
+{
+  "ok": true,
+  "status": "needs_approval",
+  "output": [{ "summary": "5 need replies, 2 need action" }],
+  "requiresApproval": {
+    "type": "approval_request",
+    "prompt": "Send 2 draft replies?",
+    "items": [],
+    "resumeToken": "..."
+  }
+}
+```
+
+User approves → resume:
+
+```json
+{
+  "action": "resume",
+  "token": "<resumeToken>",
+  "approve": true
+}
+```
+
+One workflow. Deterministic. Safe.
+
+## Tool parameters
+
+### `run`
+
+Run a pipeline in tool mode.
+
+```json
+{
+  "action": "run",
+  "pipeline": "gog.gmail.search --query 'newer_than:1d' | email.triage",
+  "cwd": "workspace",
+  "timeoutMs": 30000,
+  "maxStdoutBytes": 512000
+}
+```
+
+Run a workflow file with args:
+
+```json
+{
+  "action": "run",
+  "pipeline": "/path/to/inbox-triage.lobster",
+  "argsJson": "{\"tag\":\"family\"}"
+}
+```
+
+### `resume`
+
+Continue a halted workflow after approval.
+
+```json
+{
+  "action": "resume",
+  "token": "<resumeToken>",
+  "approve": true
+}
+```
+
+### Optional inputs
+
+- `cwd`: Relative working directory for the pipeline (must stay within the current process working directory).
+- `timeoutMs`: Kill the subprocess if it exceeds this duration (default: 20000).
+- `maxStdoutBytes`: Kill the subprocess if stdout exceeds this size (default: 512000).
+- `argsJson`: JSON string passed to `lobster run --args-json` (workflow files only).
+
+## Output envelope
+
+Lobster returns a JSON envelope with one of three statuses:
+
+- `ok` → finished successfully
+- `needs_approval` → paused; `requiresApproval.resumeToken` is required to resume
+- `cancelled` → explicitly denied or cancelled
+
+The tool surfaces the envelope in both `content` (pretty JSON) and `details` (raw object).
+
+## Approvals
+
+If `requiresApproval` is present, inspect the prompt and decide:
+
+- `approve: true` → resume and continue side effects
+- `approve: false` → cancel and finalize the workflow
+
+Use `approve --preview-from-stdin --limit N` to attach a JSON preview to approval requests without custom jq/heredoc glue. Resume tokens are now compact: Lobster stores workflow resume state under its state dir and hands back a small token key.
+
+## OpenProse
+
+OpenProse pairs well with Lobster: use `/prose` to orchestrate multi-agent prep, then run a Lobster pipeline for deterministic approvals. If a Prose program needs Lobster, allow the `lobster` tool for sub-agents via `tools.subagents.tools`. See [OpenProse](/prose).
+
+## Safety
+
+- **Local subprocess only** — no network calls from the plugin itself.
+- **No secrets** — Lobster doesn't manage OAuth; it calls Durar tools that do.
+- **Sandbox-aware** — disabled when the tool context is sandboxed.
+- **Hardened** — fixed executable name (`lobster`) on `PATH`; timeouts and output caps enforced.
+
+## Troubleshooting
+
+- **`lobster subprocess timed out`** → increase `timeoutMs`, or split a long pipeline.
+- **`lobster output exceeded maxStdoutBytes`** → raise `maxStdoutBytes` or reduce output size.
+- **`lobster returned invalid JSON`** → ensure the pipeline runs in tool mode and prints only JSON.
+- **`lobster failed (code …)`** → run the same pipeline in a terminal to inspect stderr.
+
+## Learn more
+
+- [Plugins](/tools/plugin)
+- [Plugin tool authoring](/plugins/building-plugins#registering-agent-tools)
+
+## Case study: community workflows
+
+One public example: a “second brain” CLI + Lobster pipelines that manage three Markdown vaults (personal, partner, shared). The CLI emits JSON for stats, inbox listings, and stale scans; Lobster chains those commands into workflows like `weekly-review`, `inbox-triage`, `memory-consolidation`, and `shared-task-sync`, each with approval gates. AI handles judgment (categorization) when available and falls back to deterministic rules when not.
+
+- Thread: [https://x.com/plattenschieber/status/2014508656335770033](https://x.com/plattenschieber/status/2014508656335770033)
+- Repo: [https://github.com/bloomedai/brain-cli](https://github.com/bloomedai/brain-cli)
+
+## Related
+
+- [Automation & Tasks](/automation) — scheduling Lobster workflows
+- [Automation Overview](/automation) — all automation mechanisms
+- [Tools Overview](/tools) — all available agent tools

package/docs/tools/loop-detection.md

@@ -0,0 +1,100 @@
+---
+title: "Tool-loop detection"
+summary: "How to enable and tune guardrails that detect repetitive tool-call loops"
+read_when:
+  - A user reports agents getting stuck repeating tool calls
+  - You need to tune repetitive-call protection
+  - You are editing agent tool/runtime policies
+---
+
+# Tool-loop detection
+
+Durar can keep agents from getting stuck in repeated tool-call patterns.
+The guard is **disabled by default**.
+
+Enable it only where needed, because it can block legitimate repeated calls with strict settings.
+
+## Why this exists
+
+- Detect repetitive sequences that do not make progress.
+- Detect high-frequency no-result loops (same tool, same inputs, repeated errors).
+- Detect specific repeated-call patterns for known polling tools.
+
+## Configuration block
+
+Global defaults:
+
+```json5
+{
+  tools: {
+    loopDetection: {
+      enabled: false,
+      historySize: 30,
+      warningThreshold: 10,
+      criticalThreshold: 20,
+      globalCircuitBreakerThreshold: 30,
+      detectors: {
+        genericRepeat: true,
+        knownPollNoProgress: true,
+        pingPong: true,
+      },
+    },
+  },
+}
+```
+
+Per-agent override (optional):
+
+```json5
+{
+  agents: {
+    list: [
+      {
+        id: "safe-runner",
+        tools: {
+          loopDetection: {
+            enabled: true,
+            warningThreshold: 8,
+            criticalThreshold: 16,
+          },
+        },
+      },
+    ],
+  },
+}
+```
+
+### Field behavior
+
+- `enabled`: Master switch. `false` means no loop detection is performed.
+- `historySize`: number of recent tool calls kept for analysis.
+- `warningThreshold`: threshold before classifying a pattern as warning-only.
+- `criticalThreshold`: threshold for blocking repetitive loop patterns.
+- `globalCircuitBreakerThreshold`: global no-progress breaker threshold.
+- `detectors.genericRepeat`: detects repeated same-tool + same-params patterns.
+- `detectors.knownPollNoProgress`: detects known polling-like patterns with no state change.
+- `detectors.pingPong`: detects alternating ping-pong patterns.
+
+## Recommended setup
+
+- Start with `enabled: true`, defaults unchanged.
+- Keep thresholds ordered as `warningThreshold < criticalThreshold < globalCircuitBreakerThreshold`.
+- If false positives occur:
+  - raise `warningThreshold` and/or `criticalThreshold`
+  - (optionally) raise `globalCircuitBreakerThreshold`
+  - disable only the detector causing issues
+  - reduce `historySize` for less strict historical context
+
+## Logs and expected behavior
+
+When a loop is detected, Durar reports a loop event and blocks or dampens the next tool-cycle depending on severity.
+This protects users from runaway token spend and lockups while preserving normal tool access.
+
+- Prefer warning and temporary suppression first.
+- Escalate only when repeated evidence accumulates.
+
+## Notes
+
+- `tools.loopDetection` is merged with agent-level overrides.
+- Per-agent config fully overrides or extends global values.
+- If no config exists, guardrails stay off.

package/docs/tools/minimax-search.md

@@ -0,0 +1,99 @@
+---
+summary: "MiniMax Search via the Coding Plan search API"
+read_when:
+  - You want to use MiniMax for web_search
+  - You need a MiniMax Coding Plan key
+  - You want MiniMax CN/global search host guidance
+title: "MiniMax Search"
+---
+
+# MiniMax Search
+
+Durar supports MiniMax as a `web_search` provider through the MiniMax
+Coding Plan search API. It returns structured search results with titles, URLs,
+snippets, and related queries.
+
+## Get a Coding Plan key
+
+<Steps>
+  <Step title="Create a key">
+    Create or copy a MiniMax Coding Plan key from
+    [MiniMax Platform](https://platform.minimax.io/user-center/basic-information/interface-key).
+  </Step>
+  <Step title="Store the key">
+    Set `MINIMAX_CODE_PLAN_KEY` in the Gateway environment, or configure via:
+
+    ```bash
+    Durar configure --section web
+    ```
+
+  </Step>
+</Steps>
+
+Durar also accepts `MINIMAX_CODING_API_KEY` as an env alias. `MINIMAX_API_KEY`
+is still read as a compatibility fallback when it already points at a coding-plan token.
+
+## Config
+
+```json5
+{
+  plugins: {
+    entries: {
+      minimax: {
+        config: {
+          webSearch: {
+            apiKey: "sk-cp-...", // optional if MINIMAX_CODE_PLAN_KEY is set
+            region: "global", // or "cn"
+          },
+        },
+      },
+    },
+  },
+  tools: {
+    web: {
+      search: {
+        provider: "minimax",
+      },
+    },
+  },
+}
+```
+
+**Environment alternative:** set `MINIMAX_CODE_PLAN_KEY` in the Gateway environment.
+For a gateway install, put it in `~/.Durar/.env`.
+
+## Region selection
+
+MiniMax Search uses these endpoints:
+
+- Global: `https://api.minimax.io/v1/coding_plan/search`
+- CN: `https://api.minimaxi.com/v1/coding_plan/search`
+
+If `plugins.entries.minimax.config.webSearch.region` is unset, Durar resolves
+the region in this order:
+
+1. `tools.web.search.minimax.region` / plugin-owned `webSearch.region`
+2. `MINIMAX_API_HOST`
+3. `models.providers.minimax.baseUrl`
+4. `models.providers.minimax-portal.baseUrl`
+
+That means CN onboarding or `MINIMAX_API_HOST=https://api.minimaxi.com/...`
+automatically keeps MiniMax Search on the CN host too.
+
+Even when you authenticated MiniMax through the OAuth `minimax-portal` path,
+web search still registers as provider id `minimax`; the OAuth provider base URL
+is only used as a region hint for CN/global host selection.
+
+## Supported parameters
+
+MiniMax Search supports:
+
+- `query`
+- `count` (Durar trims the returned result list to the requested count)
+
+Provider-specific filters are not currently supported.
+
+## Related
+
+- [Web Search overview](/tools/web) -- all providers and auto-detection
+- [MiniMax](/providers/minimax) -- model, image, speech, and auth setup