min-agent 0.7.8 → 1.0.0

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "min-agent",
-  "version": "0.7.8",
+  "version": "1.0.0",
   "type": "module",
   "description": "Minimal AI coding agent with tool use, MCP, and skills support",
   "license": "MIT",
@@ -53,15 +53,16 @@
     "@ai-sdk/provider": "~3.0.8",
     "@modelcontextprotocol/sdk": "~1.27.1",
     "ai": "~6.0.177",
-    "commander": "14.0.3",
-    "glob": "~13.0.5",
-    "ink": "5",
-    "ink-spinner": "5",
-    "react": "18"
+    "zod": "^4.1.8"
   },
   "devDependencies": {
     "@types/node": "22.13.9",
     "@types/react": "18",
+    "commander": "14.0.3",
+    "glob": "~13.0.5",
+    "ink": "5",
+    "ink-spinner": "5",
+    "react": "18",
     "tsx": "4.19.4",
     "typescript": "5.8.2"
   }

package/skills/self-config/SKILL.md

@@ -31,12 +31,12 @@ Config files are **strict JSON** (no comments). Read the file before editing. Pr
 | Project skill toggles; overlay `activeProvider` / `defaultModel` / `sampling` / `permission` / `thinking` / `memory` / `sandbox` / `compaction` / `agent` | `.min-agent/config.json` | Overlay is field-by-field; `sampling`/`sandbox`/`compaction`/`agent` merge keys |
 | MCP servers | `~/.min-agent/mcp.json` or `.min-agent/mcp.json` | Merge by server **name**; project replaces the same name |
 | Memories | `memory.json` in the same dirs | Prefer `memory_save` / `memory_search` tools over editing the file |
-| Global rules | `~/.min-agent/rules.md` | Always loaded |
-| Project rules | `AGENTS.md`, `RULES.md`, or `CLAUDE.md` in cwd (or `.min-agent/AGENTS.md`) | First match wins when walking up from cwd |
+| Global rules | `~/.min-agent/rules.md` and/or `~/.min-agent/AGENTS.md` | Both always loaded when present |
+| Project rules | `AGENTS.md`, `RULES.md`, or `CLAUDE.md` in cwd (or `.min-agent/AGENTS.md`) | First match wins when walking up from cwd; skips the global `~/.min-agent/AGENTS.md` |
 | Extra instruction files | `instructions` array in global `config.json` | Paths or URLs |
 | Custom tools | `.min-agent/tools/*.ts` or `~/.min-agent/tools/*.ts` | See reference.md |
 | Project skills | `.min-agent/skills/<name>/SKILL.md` | Or `min-agent skills new <name>` |
-| Global skills | `~/.agents/skills/<name>/SKILL.md` | `min-agent skills new <name> --global` |
+| Global skills | `~/.min-agent/skills/<name>/SKILL.md` (also `~/.agents/skills/`) | `min-agent skills new <name> --global` (or `min-agent --global skills new <name>`) writes under `~/.min-agent/skills/` |
 
 ## MCP
 
@@ -46,15 +46,17 @@ Write `{ "mcpServers": { "<name>": { ... } } }`.
 - Remote: `"url": "https://..."`. Optional `token`, `headers`, `remoteTransport` (`auto` default, or `streamable-http` / `sse`), `oauth` (`{}` to login in a browser, or `false` to disable).
 - `enabled: false` disables without deleting. Project `.min-agent/mcp.json` wins on the same name.
 
-CLI (user shell, not a substitute for writing the file in this session): `min-agent mcp add|update|list|info|enable|disable|remove|check` with `--project` for repo scope. `update` only changes the fields you pass (`--token ""` removes the token, `--env`/`--headers` merge, a new command switches the server to stdio and clears remote fields).
+CLI (user shell, not a substitute for writing the file in this session): `min-agent mcp add|update|rename|list|info|enable|disable|remove|check`. Put `--project` on the mcp subcommand or on the root (`min-agent --project mcp …`) for repo scope. Prefer `mcp add --name <name> …` (positional name still works). `update --rename <new> <old>` or `mcp rename <old> <new>` renames. `add` refuses to overwrite unless `--force` (or root `-y`). `update` only changes the fields you pass (`--token ""` removes the token, `--env`/`--headers` merge, a new command switches the server to stdio and clears remote fields).
 
 **This session:** editing `mcp.json` is persisted, but connected servers will not change until the user restarts. Say so after you write the file. HTTP `POST /v1/mcp` / `PATCH /v1/mcp/:name` on a running `min-agent serve` reconnects; a coding session does not.
 
 Connected servers already expose `mcp_list_resources`, `mcp_read_resource`, `mcp_list_prompts`, `mcp_get_prompt`. Do not invent per-resource tools.
 
+MCP *tools* from connected servers are inlined when there are 8 or fewer. More than 8 (or `tools.mcpTools: "catalog"`) exposes only `mcp_list_tools` then `mcp_call_tool` — list or search first, then call by the listed id. `"inline"` always sends every schema; `"auto"` is the default. Plan mode still keeps only read-only MCP tools, whether inlined or behind the dispatcher.
+
 ## Skills
 
-User-skill roots, later overriding earlier: `~/.agents/skills/` → `.min-agent/skills/` → `.agents/skills/` → `.opencode/skills/` → `.claude/skills/`. Symlinks are followed. `MIN_AGENT_SKILLS_DIRS` (path-delimiter separated) replaces that user list. Built-in skills shipped with min-agent still load and **win on name**; they are not overridden by user copies. `MIN_AGENT_NO_BUILTIN_SKILLS=1` skips built-ins.
+User-skill roots, later overriding earlier: `~/.agents/skills/` → `~/.min-agent/skills/` → `.min-agent/skills/` → `.agents/skills/` → `.opencode/skills/` → `.claude/skills/`. Symlinks are followed. `MIN_AGENT_SKILLS_DIRS` (path-delimiter separated) replaces that user list. Built-in skills shipped with min-agent still load and **win on name**; they are not overridden by user copies. `MIN_AGENT_NO_BUILTIN_SKILLS=1` skips built-ins.
 
 Enable/disable:
 
@@ -62,7 +64,7 @@ Enable/disable:
 - Project disable: `disabledSkills` in `.min-agent/config.json` (wins)
 - Re-enable a globally disabled skill in this repo: add the name to project `enabledSkills`
 
-Toggles take effect on the next turn. A **new** `SKILL.md` is invisible until skills are rediscovered (`/skills` in the interactive session, `POST /v1/skills/reload` on serve, or restart).
+Toggles take effect on the next turn. CLI: `min-agent skills enable|disable <name...> [--project|--global]` (flag on the skills subcommand or on the root). Interactive: `/skills enable|disable <name...> [--project]`. `skills enable <orphan>` clears that name from **all** config lists (global `disabledSkills` and project `disabledSkills`/`enabledSkills`) — no need to guess scope. `skills list` shows orphan entries with their lists. A **new** `SKILL.md` is invisible until skills are rediscovered (`/skills` in the interactive session, `POST /v1/skills/reload` on serve, or restart). `skills new` warns if the name collides with a built-in or another discovered skill.
 
 Scaffold:
 
@@ -81,7 +83,7 @@ Instructions...
 - Rules: edit the files in the table above. `/reload` reloads rules only, not MCP.
 - Memory: **off** by default. Config `memory` is `on` | `off` (`true`/`enable` → on, `false`/`disable` → off). Project overlays global; writing `memory` globally also removes a project `memory` overlay. When off, memories are not injected and `memory_save` / `memory_search` / `memory_delete` are not registered (takes effect next turn after you change the file). CLI `min-agent memory on|off` / `--memory` follow the same write-vs-this-run rule as permission. `/memory on|off` writes **global** by default (and clears a project overlay); `/memory --project` writes only the project file. `/memory` lists the switch plus stored entries; `/memory <text>` still writes `memory.json` even when off. Default save scope is project when `.min-agent/` exists. Shape: `{ "content": string, "tags": string[], "created": ISO string }` array. When on, memories are injected each turn as background reference, not as work to resume. Do not store secrets or API keys. HTTP: `GET/POST /v1/memory/mode`, chat body `memory`, and `GET /v1/memory` include `memory` / `source` / `label`.
 - Permission: `accept-edits` (default) | `ask` | `allow-all` in config `permission`. CLI `min-agent permission …` / `--permission` without a message writes config and exits; with a message it is this session only. `--yes` is this session only and does not write config. `min-agent exec` (one-shot) uses a permissive default when neither `--permission` nor `--yes` is given: file edits and ordinary commands run without asking, and only genuinely destructive commands (any deleting command such as plain `rm file`, `rm -r/-f`, `truncate -s 0`, `drop`/`truncate table`/`delete from`, `mkfs`/`dd`/raw-device writes, `shred`/`wipefs`, `git reset --hard`/`git clean -f`, `docker rm`/`prune`, `find … -delete`/`-exec rm`) still prompt. An explicit `--permission` (or `--yes`) overrides this and behaves exactly as in interactive sessions (`allow-all` runs even destructive commands silently). In `ask` / `accept-edits`, "dangerous commands" that still prompt do NOT include ordinary `git push`, `git stash`/`git stash pop`, `git rebase`, or `chmod` — only their harmful forms count: force pushes (`git push --force`/`-f`/`--force-with-lease`), `git stash drop`/`git stash clear`, any `rm` with target files, in-place edits (`sed -i`, `perl -i`, `awk -i inplace`), plus `git reset --hard`, `git clean -f`, `git checkout .`, `sudo`, disk/db operations. Shell quoting is respected: dangerous-looking words inside quotes (e.g. `grep "delete from users" src/`) do not prompt; in `accept-edits` a redirect into the project (`cmd > out.log`) mirrors the edit tools and stays silent, while absolute/system/device targets (`> /etc/hosts`) still ask — in `ask` mode any file-writing redirect prompts. Confirmation prompts are coalesced: when several actions in one step need approval, a single overlay lists them all and one decision applies to the whole batch instead of stacking separate prompts. The overlay offers 允许 / 总是允许 / 拒绝 (hotkeys y / a / n): choosing 总是允许 remembers that action for the current process so identical repeats (same normalized shell command, same custom/MCP tool) never prompt again until exit; MCP tools without a read-only hint confirm on each distinct tool exactly once this way, `allow-all` / `--yes` remain fully silent.
-- Thinking intensity: `off` | `low` | `medium` | `high` | `max` in config `thinking`. Omit the field to use **high**, then clamp to levels the current model actually exposes (from the model catalog). If the live gateway rejects that level, the same request retries remaining levels (high → medium → max → low → off) without surfacing the error, and the working level is remembered per model. `/think` opens an interactive picker of those English levels and writes **global** config; that also clears a project `thinking` overlay so a leftover project `off` cannot hide the new value after restart. `/think --project` writes only the project file. Typed `/think off|low|medium|high|max` still works when the model supports that value. `min-agent think …` / `--think` follow the same write-vs-this-run rule as permission. `xhigh` / `extra-high` / `none` are accepted as aliases (`xhigh` → `max`, `none` → `off`). Each request maps that intensity onto the current model and gateway (MiniMax uses `adaptive`/`disabled`; official OpenAI omits `thinking` and maps `max` to `xhigh`; an `ollama` provider uses native chat with `think` plus the configured window). Some models (Gemini, and any gateway that answers “reasoning is mandatory”) cannot turn thinking off: the request is sent with reasoning still on, `off` is dropped from `/think` for that model, and the mapping is cached. Learned wire mappings are cached in `~/.min-agent/thinking-wire-cache.json` and learned default intensities in `~/.min-agent/thinking-default-cache.json`, both for 7 days. Per-model context window and supported effort values are loaded from the model catalog (`https://models.lonae.com`, override with `MIN_AGENT_MODELS_API_URL`) and cached in `~/.min-agent/model-catalog-cache.json` for 7 days after a successful lookup. When the provider host is unknown, effort levels are the majority across offerings that actually list efforts (toggle-only rows do not vote), so a single gateway cannot invent `medium`. An `ollama` provider skips that catalog and uses native `POST /api/show` instead (whether the model advertises thinking, plus the loaded window: Modelfile `num_ctx` if set, otherwise the architecture length capped at 32768). Chat completions are sent to native `POST /api/chat` so `options.num_ctx` and `think` actually apply — the OpenAI-compatible `/v1` endpoint ignores both. `/ctx` opens a picker of `2k` / `4k` / `8k` / `12k` / `16k` / `32k` / `64k` / `128k` / `256k` / `auto` and writes the active provider's `contextWindow` (global only; providers are not in project config). Typed `/ctx 8k` and CLI `min-agent ctx 8k` do the same; `auto` clears the override. The menu item is hidden unless the active provider is ollama. Results are cached in `~/.min-agent/ollama-model-cache.json` for 7 days.
+- Thinking intensity: `off` | `low` | `medium` | `high` | `max` in config `thinking`. Omit the field to use **high**, then clamp to levels the current model actually exposes (from the model catalog). If the live gateway rejects that level, the same request retries remaining levels (high → medium → max → low → off) without surfacing the error, and the working level is remembered per model. `/think` opens an interactive picker of those English levels and writes **global** config; that also clears a project `thinking` overlay so a leftover project `off` cannot hide the new value after restart. `/think --project` writes only the project file. Typed `/think off|low|medium|high|max` still works when the model supports that value. `min-agent think …` / `--think` follow the same write-vs-this-run rule as permission. `xhigh` / `extra-high` / `none` are accepted as aliases (`xhigh` → `max`, `none` → `off`). Each request maps that intensity onto the current model and gateway (MiniMax uses `adaptive`/`disabled`; official OpenAI omits `thinking` and maps `max` to `xhigh`; an `ollama` provider uses native chat with `think` plus the configured window). Some models (Gemini, and any gateway that answers “reasoning is mandatory”) cannot turn thinking off: the request is sent with reasoning still on, `off` is dropped from `/think` for that model, and the mapping is cached. Learned wire mappings are cached in `~/.min-agent/thinking-wire-cache.json` and learned default intensities in `~/.min-agent/thinking-default-cache.json`, both for 7 days. Per-model context window and supported effort values are loaded from the model catalog (`https://models.lonae.com`, override with `MIN_AGENT_MODELS_API_URL`) and cached in `~/.min-agent/model-catalog-cache.json` for 7 days after a successful lookup. When the provider host is unknown, effort levels are the majority across offerings that actually list efforts (toggle-only rows do not vote), so a single gateway cannot invent `medium`. An `ollama` provider skips that catalog and uses native `POST /api/show` instead (whether the model advertises thinking, plus the loaded window: Modelfile `num_ctx` if set, otherwise the architecture length capped at 32768). Chat completions are sent to native `POST /api/chat` so `options.num_ctx` and `think` actually apply — the OpenAI-compatible `/v1` endpoint ignores both. `/ctx` opens a picker of `2k` / `4k` / `8k` / `12k` / `16k` / `32k` / `64k` / `128k` / `256k` / `auto` and writes the active provider's `contextWindow` (global only; providers are not in project config). Typed `/ctx 8k` and CLI `min-agent ctx 8k` do the same; `auto` clears the override. The menu item is hidden unless the active provider is ollama. Non-Ollama models without a detected window use 128k; if a request overflows and reports `n_ctx`, that value is remembered for 24 hours (it does not override an explicit `contextWindow`). Results are cached in `~/.min-agent/ollama-model-cache.json` for 7 days.
 - Sandbox default is **off**. `mode`: `off` | `workspace` | `strict`. Optional `network`, `extraWriteRoots`, `extraReadRoots`. Isolation is independent of permission. `--sandbox` / `--network` without a message write config and exit. On Windows there is no seatbelt/bwrap: `strict` rejects bash; `workspace` still path-checks file tools; bash prefers Git Bash when installed, otherwise `cmd.exe`.
 
 ## Providers

package/skills/self-config/reference.md

@@ -32,10 +32,10 @@ Read this when you need exact keys. Files are strict JSON.
     "extraWriteRoots": [],
     "extraReadRoots": []
   },
-  "pricing": { "gpt-5.5": { "inputPerMillion": 1.25, "outputPerMillion": 10 } },
+  "pricing": { "gpt-5.5": { "inputPerMillion": 1.25, "outputPerMillion": 10, "cacheReadPerMillion": 0.125 } },
   "sampling": { "temperature": 0.7, "maxTokens": 4096, "topP": 1 },
-  "compaction": { "model": "gemini-2.0-flash", "autoContinue": true },
-  "tools": { "bashTimeoutMs": 600000 },
+  "compaction": { "model": "gemini-2.0-flash", "autoContinue": true, "chunkChars": 20000 },
+  "tools": { "bashTimeoutMs": 600000, "outputMaxLines": 400, "outputMaxBytes": 24576, "mcpTools": "auto" },
   "agent": {
     "maxSteps": 40,
     "maxContinues": 40,
@@ -44,8 +44,9 @@ Read this when you need exact keys. Files are strict JSON.
     "emptyRetryDelayMs": 1000,
     "requestTimeoutMs": 180000,
     "streamIdleTimeoutMs": 90000,
-    "subAgentMaxSteps": 15,
-    "exploreMaxSteps": 20,
+    "subAgentMaxSteps": 40,
+    "exploreMaxSteps": 30,
+    "streamChunkSteps": 16,
     "researchSteerAfter": 3,
     "researchStopAfter": 5,
     "researchTotalCap": 24
@@ -55,14 +56,20 @@ Read this when you need exact keys. Files are strict JSON.
 
 - `thinking`: `off` | `low` | `medium` | `high` | `max`; omit to use **high**, then clamp to what the current model supports (catalog first; a working level is remembered per model in `thinking-default-cache.json`). A project `thinking` value overlays global; writing thinking globally removes the project overlay so the chosen level is what later sessions use. `/think --project` writes only the project file.
 - `memory`: `on` | `off`; omit to keep **off**. When off, stored `memory.json` entries are not injected and memory tools are unavailable. A project `memory` value overlays global; writing memory globally removes the project overlay. `/memory --project` writes only the project file.
-- `contextWindow`: tokens; omit to query the API. An `ollama` provider uses `POST /api/show`: Modelfile `num_ctx` if set, otherwise min(architecture context_length, 32768), else 2048. Chat goes to native `POST /api/chat` with that value as `options.num_ctx` (OpenAI-compatible `/v1` ignores it). Interactive `/ctx` (and `min-agent ctx`) sets this field on the active ollama provider to a fixed level (`2k` / `4k` / `8k` / `12k` / `16k` / `32k` / `64k` / `128k` / `256k`) or `auto` to clear it. Other providers then 512k fallback.
+- `contextWindow`: tokens; omit to query the API. An `ollama` provider uses `POST /api/show`: Modelfile `num_ctx` if set, otherwise min(architecture context_length, 32768), else 2048. Chat goes to native `POST /api/chat` with that value as `options.num_ctx` (OpenAI-compatible `/v1` ignores it). Interactive `/ctx` (and `min-agent ctx`) sets this field on the active ollama provider to a fixed level (`2k` / `4k` / `8k` / `12k` / `16k` / `32k` / `64k` / `128k` / `256k`) or `auto` to clear it. Other providers then 128k fallback. If a request overflows and the error reports `n_ctx`, that value is remembered for 24h (does not override an explicit `contextWindow`).
+- `pricing.<model>`: USD per 1M tokens (`inputPerMillion`, `outputPerMillion`). Optional `cacheReadPerMillion` / `cacheWritePerMillion` for reused / written prefix tokens; omit to price reuse at the input rate.
 - `agent.autoContinue` omitted → uses `compaction.autoContinue` (default true).
+- `compaction.chunkChars`: max characters of older history sent in one summary call (default 20000, clamped 4000–80000). Longer history is summarized slice-by-slice, then merged. A summary that is missing required headings (`## Goal`, `## Relevant Files`, …) or drops the current goal / known file paths is repaired once; if it still fails, compaction keeps pruned messages instead of replacing history with a bad summary. Project overlay merges this field.
 - `agent.maxSteps` omitted → no tool-step ceiling. `agent.maxContinues` omitted → no auto-continue ceiling.
 - `agent.maxEmptyAttempts` / `agent.emptyRetryDelayMs`: retries when the provider returns an empty stream (HTTP 200, no content). First retry is immediate, later ones back off exponentially, capped at 8s.
 - `agent.requestTimeoutMs` / `agent.streamIdleTimeoutMs`: abort a request that never starts responding, or a stream that goes quiet mid-response; the run then retries it like an empty reply. 0 disables. Env overrides: `MIN_AGENT_REQUEST_TIMEOUT_MS`, `MIN_AGENT_STREAM_IDLE_TIMEOUT_MS`.
-- `agent.subAgentMaxSteps` / `agent.exploreMaxSteps`: step ceilings for the `task` and `explore` sub-agents. Sub-agents share the parent's research budget.
-- `agent.researchSteerAfter` / `researchStopAfter`: research *rounds that returned no new source* before steering / dropping `search_web` and `web_fetch`; `researchTotalCap` is the hard number of research calls per turn. Productive research (new URLs or domains each round) is not penalised.
+- `agent.subAgentMaxSteps` / `agent.exploreMaxSteps`: step ceilings for the `task` and `explore` sub-agents (defaults 40 and 30). Each sub-agent has its own research and repeat budget (same thresholds as the parent, empty counters). Child searches do not spend the parent's allowance, and parallel tasks do not mix in-flight counts. The child returns a structured report (conclusion, files changed or inspected, unfinished work, errors), not only its last text.
+- `agent.subAgentModel` / `agent.exploreModel`: optional model ids for those sub-agents. Omit or leave blank to use the parent's model. Use this to send exploration / bulk file work to a cheaper model.
+- `agent.streamChunkSteps`: tool steps in one model request before old tool payloads are pruned and the next request starts. Default 16. Does not inject a continue prompt and does not count toward `maxContinues`. Omit or values below 1 use 16.
+- `agent.researchSteerAfter` / `researchStopAfter`: research *rounds that returned no new source* before steering / stubbing `search_web` and `web_fetch` (the tools stay in the schema so the prompt prefix stays stable); `researchTotalCap` is the hard number of research calls per turn. Productive research (new URLs or domains each round) is not penalised. Further search/fetch calls then return a stop-and-write notice instead of new results.
 - `tools.bashTimeoutMs`: default kill timeout for `bash` (10 minutes; the model can override per call, 0 disables). On Windows, shell commands prefer Git Bash when `bash` is on PATH; otherwise `cmd.exe`. Timeouts / abort use `taskkill /T` to kill the process tree.
+- `tools.outputMaxLines` / `tools.outputMaxBytes`: how much of a tool result stays in the conversation (default 400 lines / 24KiB). The rest is saved under `~/.min-agent/tool-output/` with a path the model can `read`. Clamped to 20–20000 lines and 1KiB–1MiB. A caller that passes explicit limits (for example a skill) wins over these defaults. Global only (not a project overlay).
+- `tools.mcpTools`: `auto` (default) | `catalog` | `inline`. `auto` inlines MCP server tool schemas up to 8 tools, then switches to the stable pair `mcp_list_tools` / `mcp_call_tool` so adding or removing a server tool does not rewrite the whole tool list. `catalog` always uses that pair; `inline` always sends every schema. Sub-agents follow the same setting. Global only (not a project overlay).
 - `MIN_AGENT_TRACE=1` logs a request/response summary for every model call (model, message and tool counts, status, time to first byte).
 - Logs live in `~/.min-agent/logs/min-agent.YYYY-MM-DD.log` (kept 7 days) and are also shipped to a hardcoded Loki endpoint (`POST https://loki.lonae.com/loki/api/v1/push`, labels `job=min-agent` + `level`). Shipping is fire-and-forget (batched, 2s); the file log stays the source of truth. There is no config switch for it.
 - Web search/fetch that has not yet produced the requested output still continues even when `autoContinue` is false; further search/fetch calls return a stop-and-write notice instead of new results.
@@ -152,4 +159,4 @@ Array of `{ "content": string, "tags": string[], "created": string }`. Prefer me
 
 ## HTTP (only if the user is running `min-agent serve`)
 
-Skills: `GET /v1/skills`, `GET /v1/skills/:name`, `POST /v1/skills/reload`, `POST /v1/skills/:name/enable|disable` with `{ "scope": "project"|"global" }`. MCP enable/disable similarly. Thinking: `GET/POST /v1/thinking`. Memory switch: `GET/POST /v1/memory/mode`. Ollama context level: `GET/POST /v1/context` with `{ "level": "8k" }` or `"auto"`. Full list is `docs/API.md` in the min-agent package; do not treat that file as this skill's replacement.
+Skills: `GET /v1/skills`, `GET /v1/skills/:name`, `POST /v1/skills/reload`, `POST /v1/skills/:name/enable|disable` with `{ "scope": "project"|"global" }` (`project` without `.min-agent/` → `400 project_dir_missing`). `stale_disabled` entries are `{ "name", "in": ["global_disabled"|"project_disabled"|"project_enabled"] }`; enabling an orphan name clears it from every list (`cleared_stale` in the response). MCP enable/disable similarly. Thinking: `GET/POST /v1/thinking`. Memory switch: `GET/POST /v1/memory/mode`. Ollama context level: `GET/POST /v1/context` with `{ "level": "8k" }` or `"auto"`. Full list is `docs/API.md` in the min-agent package; do not treat that file as this skill's replacement.