@adaptiqai/openclaw 0.1.4 → 0.1.6

package/README.md

@@ -7,63 +7,174 @@ OpenClaw plugin.
 ## Install
 
 ```bash
-openclaw plugins install @adaptiqai/openclaw
+openclaw plugins install @adaptiqai/openclaw --force --accept-capabilities
 openclaw gateway restart
 ```
 
-Then sign in, in a terminal:
+Then sign in, in a chat with OpenClaw:
+
+```
+/adaptiq signin
+```
+
+or in a terminal:
 
 ```bash
 openclaw models auth login --provider adaptiq
 ```
 
-OpenClaw shows a link and a code; you sign in (or create an account) in your
-browser and approve, and OpenClaw picks it up by itself. Without a browser,
-give a key from the AdaptiqAI console instead:
+Either shows a link and a code; you sign in (or create an account) in your
+browser and approve, and OpenClaw picks it up by itself. In the chat, the link
+is given, never opened: the Gateway may run on another computer than your
+browser. Without a browser, give a key from the AdaptiqAI console instead:
 `openclaw models auth paste-api-key --provider adaptiq`, or `ADAPTIQAI_API_KEY`.
 A setup made by the console's Connect to app keeps working as it is.
 
-Choose a model of the **AdaptiqAI** provider (`openclaw models set adaptiq/<id>`).
-The list is the models in your workspace, each with its own context
-window, output limit and thinking levels.
+The plugin tries every credential it has, in this order: the browser sign-in,
+the key OpenClaw resolved for the provider (a pasted key comes before
+`ADAPTIQAI_API_KEY`, as OpenClaw orders them), then `ADAPTIQAI_API_KEY`. A key
+AdaptiqAI refuses gives way to the next, so an old key never hides a working
+one.
+
+Choose a model of the **AdaptiqAI** provider (`/model adaptiq/<id>`, or
+`openclaw models set adaptiq/<id>`). The list is the models in your
+workspace, each with its own context window, output limit, prices, and
+thinking levels; a run you give no level starts on the model's own default.
+The terminal sign-in starts you on the first model that answers now.
 
 ## What it does
 
 - **A conversation is one AdaptiqAI session.** Every turn carries OpenClaw's
   session id as `x-adaptiq-session-id`, so AdaptiqAI keeps the conversation's
-  memory, and the collections and conversations you attach to it in the
-  console reach the model. `/new` starts a new one.
+  memory, and the collections and conversations you attach to it reach the
+  model. `/new` starts a new one. Switch a conversation to another model and
+  the first reply there ends with a line saying where its memory is, with
+  `/model adaptiq/<id>` to switch back or `/new` to start fresh.
 - **No compaction.** AdaptiqAI keeps the whole conversation in native memory,
   so OpenClaw's compaction is declined for AdaptiqAI models: nothing is sent,
   nothing is removed, and OpenClaw keeps the conversation as it is. Other
   models compact as before.
-- **Helper calls leave nothing.** Titles and other one-off completions run
-  once, marked ephemeral, and never become a conversation. A heartbeat run
-  reads its conversation back, the memory and the knowledge attached to it,
-  and saves nothing; in a conversation AdaptiqAI doesn't store yet it runs
-  once. Cron and memory-flush runs look like turns to a provider; to keep
-  them out as well, let the plugin see each run's trigger:
-
-  ```bash
-  openclaw config set plugins.entries.adaptiq.hooks.allowConversationAccess true
-  ```
-
-  The plugin reads only the trigger.
+- **Helper calls and the runs OpenClaw starts by itself leave nothing.**
+  Titles and other one-off completions run once, marked ephemeral, and never
+  become a conversation. Heartbeat, cron and memory-flush runs read their
+  conversation back, the memory and the knowledge attached to it, and save
+  nothing; in a conversation AdaptiqAI doesn't store yet they run once.
+  OpenClaw 2026.10 marks these runs itself, so this needs no access to
+  conversations. On older OpenClaw, cron and memory-flush runs look like
+  turns unless you let the plugin see each run's trigger
+  (`plugins.entries.adaptiq.hooks.allowConversationAccess`).
+- **Only your own words count as you asking.** The agent saves knowledge only
+  when you asked for it. Prompts OpenClaw writes itself (cron jobs, wake-ups,
+  system events), a subagent's task, and another session's messages never
+  count, whatever they say.
 - **A subagent names the session that started it**, so the console folds it
   under that conversation: as a helper for a one-off task (`mode: "run"`), as
   its own line of work for a persistent one (`mode: "session"`).
 - **Tools for the agent:** `adaptiq_knowledge` (what this conversation uses;
-  search, use or stop a collection), `adaptiq_memory` (how full the memory is;
-  the saved collections; keep a conversation private) and
-  `adaptiq_create_knowledge` (save material as a collection, only when you ask).
+  find words, ids or names in the text of its collections, with the exact
+  lines as document:line; search the library by name, use or stop a
+  collection), `adaptiq_memory` (how full the memory is; the saved
+  collections; keep a conversation private) and `adaptiq_create_knowledge`
+  (save material as a collection, only when you ask). They are always in the
+  model's tool list, also where OpenClaw's tool search hides other plugins'
+  tools.
   Where `commands.ownerAllowFrom` names the owners, only their messages save
   knowledge (in a group, other members' do not), and an agent kept to its
   workspace (a sandbox, or `tools.fs.workspaceOnly`) saves only files inside
   it.
-- **`/adaptiq`** shows the workspace, its models and this conversation.
-  `/adaptiq private` stops saving new messages of this conversation,
-  `/adaptiq remember` starts again, `/adaptiq console` gives the console's link
-  for it, and `/adaptiq logout` signs this computer out.
+- **Each turn is told to look a collection up with find, and which ready
+  collections its conversation reads**, in the system prompt of every request
+  the plugin sends for it. It is the same text until the collections change,
+  and needs no access to conversations. It is prompt injection, which
+  `hooks.allowPromptInjection: false` turns off; a run whose prompt is the
+  identity line alone (prompt mode `none`) gets neither.
+- **Busy answers are waited out.** When AdaptiqAI says it is busy and when to
+  come back (`Retry-After`), the plugin waits and asks again, twice at most
+  and within a minute, before OpenClaw sees a failure.
+- **A full memory is said as the reply.** OpenClaw shows a provider's
+  failure only in its own words ("The AI service is having trouble. Please
+  try again in a moment.") and retries it eight times, which cannot help. So
+  when AdaptiqAI's memory is full, the reply says so, what frees room, and the
+  console's link, never "start a new conversation", which frees nothing.
+- **Memory notices come under the next answer, once.** When idle
+  conversations go to sleep to make room, or the library is full or nearly
+  so, the next answer ends with one line that says so (with the console's
+  link where there is something to do), as DeepSeek Harness shows it for a
+  few seconds. `/adaptiq` repeats it while it holds.
+
+## The AdaptiqAI tab
+
+On OpenClaw 2026.10 the Control UI gets an **AdaptiqAI** tab
+(`http://<gateway>/adaptiq`): your workspace and account, each model with its
+status, context, thinking levels and what runs and waits on it now, memory
+(library and conversations), the conversations that ran on AdaptiqAI since
+OpenClaw started (how much of each is in memory, saving or private, what it
+reads, and what its last turn and all of it read from memory, sent new,
+saved and cost), your collections, and the console's links.
+
+The Control UI shows a plugin's tab in a sandboxed frame that may only read,
+and cannot open pages or send forms. So the tab shows, and the chat acts: each
+change is the `/adaptiq` command it names, and the links are there to copy.
+The page is a route of the Gateway behind its own sign-in.
+
+## `/adaptiq`
+
+| Command | Does |
+|---|---|
+| `/adaptiq` | the workspace, this conversation (its memory and what it processed), memory and models; what needs you (full memory, an empty balance), and the update |
+| `/adaptiq signin` | connect your AdaptiqAI account in the browser |
+| `/adaptiq private` | stop saving new messages of this conversation |
+| `/adaptiq remember` | save this conversation to memory again |
+| `/adaptiq knowledge` | knowledge used in this conversation, and what else you can use |
+| `/adaptiq conversations` | the earlier conversations on this model, to add one with `/adaptiq use` |
+| `/adaptiq use <name>` | use a collection or an earlier conversation here |
+| `/adaptiq stop <name>` | stop using it here (nothing is deleted) |
+| `/adaptiq available <name>` | make a collection built for another model available for this conversation's |
+| `/adaptiq open` | the console's link for this conversation |
+| `/adaptiq update` | whether a newer version is out; `/adaptiq update now` installs it and switches to it |
+| `/adaptiq diagnostics` | a report you can send to support (no secrets) |
+| `/adaptiq signout` | forget the connection on this computer |
+
+Where `commands.ownerAllowFrom` names the owners, only they sign in or out,
+and only they update the plugin.
+
+Replies come with buttons where the channel has them (Telegram, Discord,
+Slack): Add funds, Get access, Sign in again, Private or Remember, Knowledge,
+Use a collection or an earlier conversation, Install an update, Open console.
+A button runs the `/adaptiq` command it names, or opens the console. Where a
+channel has none (the Control UI's chat among them), each is a line under the
+reply, as OpenClaw's own `/status` does.
+
+## Updates
+
+The plugin checks npm once a day. When a newer version is out, `/adaptiq`
+and the Gateway log say so, and `/adaptiq update now` installs it with
+OpenClaw's own command and switches to it: the Gateway loads the new version
+in place, without a restart. The install starts once no AdaptiqAI answer is in
+progress; `/adaptiq` then shows the version running, and `/adaptiq update`
+says if it did not finish. In a terminal:
+
+```bash
+openclaw plugins install @adaptiqai/openclaw@<version> --force --accept-capabilities && openclaw gateway restart
+```
+
+`openclaw plugins update` works too, with the Gateway stopped.
+
+## What OpenClaw does not let a plugin do
+
+- **"Preparing memory".** OpenClaw reads AdaptiqAI's answers itself and gives
+  a provider only each answer's status and headers; the stream's own progress
+  lines never reach a plugin, and its status line takes only OpenClaw's own
+  stages. A memory notice reaches the plugin only when AdaptiqAI sends it as
+  a header.
+- **Picking an earlier conversation after @.** The chat's @ list holds the
+  Gateway's people alone; `/adaptiq conversations` lists them instead.
+- **Acting in the tab.** The tab's frame may only read: changes are the
+  `/adaptiq` commands it names, and files are saved as knowledge by asking the
+  agent (`adaptiq_create_knowledge`) or in the console.
+- **Buttons in the Control UI's chat**, which shows a reply's text only.
+
+An empty balance is said by AdaptiqAI itself, with the link to add funds.
 
 ## Configuration
 
@@ -79,6 +190,9 @@ spawned as sessions of their own are kept in `~/.openclaw/adaptiqai/`
 OpenClaw's auth profile `adaptiq:device` holds only a stand-in: the plugin
 swaps it for the current access token on every run that goes to
 `ADAPTIQAI_API_BASE`. A provider pointed at another address gets no token.
+A sign-in from the chat writes no auth profile, so while one is saved and
+`ADAPTIQAI_API_KEY` is not set, the plugin puts the same stand-in there, in
+the Gateway's own process only; it never replaces a key you set.
 
 ## Development
 
@@ -89,4 +203,4 @@ pnpm --filter @adaptiqai/openclaw run test
 openclaw plugins install --link ./packages/openclaw --force   # try it in your own OpenClaw
 ```
 
-Tested against OpenClaw 2026.9.2.
+Tested against OpenClaw 2026.9.2 and 2026.10.10.