Files
VolcanicAgenticCore/docs/extensions.md

228 lines
14 KiB
Markdown

# Writing extensions
An extension is a class implementing `IExtension` from `Volcanic.AgenticCore.Abstractions`. One instance is
created per `AgentSession`, so instance fields are per conversation. `ActivateAsync` registers things; the
runtime binds actions afterwards.
```csharp
using System.ComponentModel;
using Volcanic.AgenticCore.Abstractions.Extensions;
using Volcanic.AgenticCore.Abstractions.Extensions.Events;
using Volcanic.AgenticCore.Abstractions.Messages;
using Volcanic.AgenticCore.Abstractions.Tools;
public sealed record HelloArgs([property: Description("Name to greet")] string Name);
public sealed class HelloTool : Tool<HelloArgs>
{
public override string Name => "hello";
public override string Description => "Greets someone";
public override string? PromptSnippet => "hello: greet a person by name";
protected override Task<ToolResult> ExecuteAsync(HelloArgs args, ToolInvocation invocation, CancellationToken ct) =>
Task.FromResult(ToolResult.Text($"Hello, {args.Name}!", new { greeted = args.Name }));
}
public sealed class HelloExtension : IExtension
{
public string Name => "hello";
public ValueTask ActivateAsync(IExtensionApi api, CancellationToken ct)
{
api.RegisterTool(new HelloTool());
api.RegisterCommand("hello", new CommandDefinition("Say hello through the model", (args, ctx, ct) =>
{
api.SendUserMessage($"Please greet {args} using the hello tool.");
return Task.CompletedTask;
}));
api.On<ToolCallEvent>(async (e, ctx, ct) =>
{
if (e.ToolName != "bash") return;
var command = e.Input["command"]?.GetValue<string>() ?? "";
if (!command.Contains("rm -rf")) return;
if (!ctx.HasUi) { e.Block("Dangerous command blocked (no UI to confirm)"); return; }
var choice = await ctx.Ui.SelectAsync($"Run `{command}`?", ["Allow", "Block"], null, ct);
if (choice != "Allow") e.Block("Blocked by user");
});
return ValueTask.CompletedTask;
}
}
```
Register it on the host (`o.AddExtension<HelloExtension>()`), pass it per session
(`AgentSessionOptions.ExtraExtensions`), or drop the compiled assembly into an extensions directory.
## The API object
| Member | Purpose |
|---|---|
| `On<TEvent>(handler)` | Subscribe to a lifecycle event. Returns an `IDisposable` to unsubscribe. |
| `RegisterTool(tool)` | Add a tool. First registration of a name wins; an extension tool replaces a built-in of the same name. |
| `RegisterCommand(name, def)` | Add `/name`. Duplicate names get `name:2`, `name:3`. |
| `RegisterProvider(registration)` / `UnregisterProvider(id)` | Add an LLM back end (see [providers](providers.md)). |
| `SendMessage(custom, delivery, triggerTurn)` | Inject a `CustomMessage` (shown to the model as a user message). `Steer`, `FollowUp` or `NextTurn`. |
| `SendUserMessage(text, images, delivery)` | Inject a user message; always triggers a turn. |
| `AppendEntry(customType, data)` | Persist extension state on the session branch; never enters the model context. |
| `SetModelAsync`, `GetThinkingLevel`, `SetThinkingLevel` | Session-scoped model control. |
| `GetActiveTools`, `SetActiveTools`, `GetAllTools`, `GetCommands` | Introspection and tool gating. |
| `GetSessionName`, `SetSessionName` | Display name. |
| `ExecAsync(command, args, options)` | Run a process in the session directory. Its environment comes from the process policy; `ExecOptions.Environment` sets (or, with null, removes) variables on top of the filtered base. |
| `Events` | Bus shared by all extensions in the session (`Emit(channel, data)`, `On(channel, handler)`). |
| `Logger` | An `ILogger` named after the extension. |
| `Processes` | The host's `IProcessSpawnPolicy`: host hooks only, no extension events. Use it where there's no session, e.g. during activation. |
| `Redactor` | `Redact(text)` replaces the host's known secrets with `[REDACTED]`, for text you send to the model outside a tool result. |
Action members throw `ExtensionNotBoundException` when called during `ActivateAsync`: do registration there
and start processes or timers in a `SessionStartEvent` handler. `Processes` and `Redactor` work during activation. After a reload or session replacement the old
API and contexts throw `StaleExtensionContextException`.
## Events
Handlers have the shape `(TEvent e, IExtensionContext ctx, CancellationToken ct) => ValueTask`. Events are
mutable; change them to influence the runtime.
| Event | When | What you can do |
|---|---|---|
| `SessionStartEvent` | After activation, on new/resume/switch | Rebuild state from `ctx.SessionManager.GetBranch()`, start resources |
| `SessionShutdownEvent` | Before dispose or replacement | Clean up (idempotent) |
| `SessionBeforeSwitchEvent` | Before new/switch session | `Cancel()` |
| `SessionInfoChangedEvent` | Name changed | notify |
| `InputEvent` | User text before it becomes a prompt | Change `Text`/`Images`, or `MarkHandled()` to consume it |
| `BeforeAgentStartEvent` | Before a run | Edit `SystemPromptOptions` (base prompt, sections, tools, guidelines), set `SystemPromptOverride`, add `InjectedMessages` |
| `AgentStartEvent` / `AgentEndEvent` / `AgentSettledEvent` | Run boundaries | notify |
| `TurnStartEvent` / `TurnEndEvent` | Turn boundaries | `TurnEndEvent.ContinueRequested = true` forces another model call |
| `ContextEvent` | Before each model call | Edit `Messages` in place (system prompt excluded) |
| `BeforeProviderRequestEvent` | Provider payload built | Replace `Payload` |
| `MessageStartEvent` / `MessageUpdateEvent` / `MessageEndEvent` | Message streaming | `MessageEndEvent.Message` can be replaced (same role) before it is persisted |
| `ToolCallEvent` | Before a tool runs | Patch `Input`, or `Block(reason)`; a throwing handler blocks |
| `ToolResultEvent` | After a tool ran | Edit `Content`, `Details`, `IsError` |
| `ToolExecutionStart/Update/EndEvent` | Tool lifecycle | notify |
| `ProcessSpawnEvent` | Before a process the agent asked for starts | Replace `Context` (command, cwd, environment); a throwing handler fails the spawn |
| `SessionBeforeCompactEvent` | Compaction about to run | `Cancel()` or set `Override` to supply your own summary |
| `SessionCompactEvent` | Compaction entry written | notify |
| `ModelSelectEvent` / `ThinkingLevelSelectEvent` | Settings changed | notify |
Dispatch rules: handlers run in extension load order, then registration order; exceptions are reported as
`extension_error` session events and dispatch continues (except `ProcessSpawnEvent`, where the exception also
fails the spawn); `Cancel`, `Block` and `MarkHandled` stop further handlers.
## The context object
`IExtensionContext` gives handlers and tools: `Cwd`, `SessionId`, `TimeZone` (the session's zone), `HasUi`, `Ui` (select/confirm/input/notify/
status), `SessionManager` (read-only entries and branch), `ModelRegistry` (models and provider-neutral
`CompleteAsync` for nested model calls), `Model`, `ThinkingLevel`, `Signal` (cancelled on abort), `IsIdle()`,
`Abort()`, `HasPendingMessages()`, `GetContextUsage()`, `GetSystemPrompt()`, `Compact()`,
`PrepareProcessAsync(request)` (see [spawning processes](#spawning-processes)), `Channel` (the host's
`IChatChannel`, or null).
Command handlers get `IExtensionCommandContext`, which adds `WaitForIdleAsync`, `NewSessionAsync` and
`SwitchSessionAsync`.
## Tools
Implement `ITool` directly or derive from `Tool<TArgs>`; the JSON Schema is generated from `TArgs`
(records with `[Description]` attributes work well; nullable properties are optional). `DelegateTool<TArgs>`
builds a tool from a lambda.
- Throw to produce an error result. `ToolExecutionException(result)` (in `Volcanic.AgenticCore.Abstractions.Tools`)
returns content and details while marking the result as an error.
- `ToolResult.Content` is what the model sees; `Details` is structured data for hosts.
- `Terminate = true` on every result of a batch ends the run without another model call.
- `ExecutionMode = Sequential` forces the whole batch to run one at a time (the built-in `edit`/`write` do this).
- `PromptSnippet` puts one line in the system prompt's tools section; `PromptGuidelines` adds bullets.
- Report progress through `invocation.Report(partial)`; hosts see it as `tool_execution_update`.
## Spawning processes
A process started on the agent's behalf must not inherit the host's environment as it is: that's how
`TELEGRAM_BOT_TOKEN` or `OPENAI_API_KEY` ends up in tool output. Build its environment through the process policy
(`Volcanic.AgenticCore.Abstractions.Processes`):
```csharp
var spawn = await ctx.PrepareProcessAsync(new ProcessSpawnRequest
{
Purpose = ProcessPurpose.ExtensionExec,
Command = "gh",
Arguments = ["issue", "list"],
Cwd = ctx.Cwd,
Environment = new Dictionary<string, string?> { ["GH_TOKEN"] = myToken },
Origin = "my-extension",
}, ct);
var psi = new ProcessStartInfo(spawn.Command) { UseShellExecute = false };
foreach (var a in spawn.Arguments) psi.ArgumentList.Add(a);
spawn.ApplyTo(psi); // WorkingDirectory = spawn.Cwd, Environment cleared, then spawn.Environment copied in
```
- In a session, call `ctx.PrepareProcessAsync`. It fills `Scope`, `SessionId` and (unless you set it)
`TimeZoneId` from the session, runs the policy, then emits `ProcessSpawnEvent` to extensions in load order.
`TimeZoneId` is filled only when the session or host configured a zone; otherwise the child inherits the host's `TZ`.
Without a session (during activation, or from a process-wide service), call `api.Processes.Prepare(request)`:
only the host hooks run.
- The returned `ProcessSpawnContext.Environment` is the **complete** environment. `spawn.ApplyTo(psi)` clears
`ProcessStartInfo.Environment` and copies it in (and sets the working directory); don't merge it with the inherited one.
- The policy starts from the host's environment, removes what the deny-list (`*_API_KEY`, `*_TOKEN`, `*_SECRET`,
`*_PASSWORD`, their `Section__Key` forms, and a few names), `Allow` and `SecretValues` reject, then applies
`request.Environment`, then sets `TZ` from `TimeZoneId`, then runs the host hooks.
- **Variables you pass explicitly are applied after the deny-list and never filtered.** A process that needs a
token gets it only when you pass it, and no other process sees it. A null value removes the variable.
- Host hooks (`ProcessSpawnHook`) are synchronous, pure rewrites of command, arguments, cwd and environment.
`ProcessSpawnEvent` handlers are ordinary (async) extension handlers that may replace `Context`; keep them free of
slow I/O, because they delay the spawn. A hook or handler that throws fails the spawn (the exception comes out of
`PrepareProcessAsync`), and nothing falls back to the inherited environment. Report it the way you report a
failed start; the built-in `bash` returns `process environment hook failed: <message>`.
- `ExecAsync` already does all of this with `Purpose = ExtensionExec` and `Origin` = your extension's name.
- Tool output is redacted by the agent loop: every `SecretValues` entry in a tool result, progress update or
error becomes `[REDACTED]` before the model, the transcript, hosts or `ToolResultEvent` handlers see it. Tools
don't redact themselves. Text you send to the model some other way (`SendMessage`, `SendUserMessage`) isn't
redacted: pass it through `api.Redactor.Redact` first.
## Where state should live
| State | Use |
|---|---|
| Something the model should see | `SendMessage` / `SendUserMessage` |
| Durable data outside the model context | `AppendEntry` and rebuild from `GetBranch()` on `SessionStartEvent` |
| Per-run data that follows the branch | tool result `Details` |
| Data spanning sessions | your own storage |
## Discovery and loading
Sources, in precedence order (earlier wins tool/command name collisions):
1. `AgenticCoreOptions.Extensions` (host code) and `AgentSessionOptions.ExtraExtensions`
2. `<cwd>/.volcanic/settings.json` → `{"extensions": ["relative/or/absolute/Ext.dll"]}`
3. `<cwd>/.volcanic/extensions/*.dll` and `<cwd>/.volcanic/extensions/<Name>/<Name>.dll`
4. `~/.volcanic/agent/settings.json` → `extensions`
5. `~/.volcanic/agent/extensions/*.dll`
An assembly is scanned for public, non-abstract `IExtension` types with a usable constructor (constructor
parameters are resolved from the host's `IServiceProvider`). Each assembly loads in its own
`AssemblyLoadContext`; `Volcanic.AgenticCore.*`, `Microsoft.Extensions.*` and `System.*` are shared with the
host so contract types unify. Build an extension with
`<ProjectReference Include="Volcanic.AgenticCore.Abstractions.csproj" Private="false" />` (or a package
reference) so the host's copy is used at runtime.
Activation or load failures are recorded (`AgentSession.ExtensionErrors`) and logged; other extensions still
load.
## Feature extensions in this repository
Besides the providers ([providers](providers.md), [claude-code-provider.md](claude-code-provider.md)), `extensions/`
holds feature extensions built on this API only: `Mcp` ([mcp.md](mcp.md)), `GoogleCalendar`
([google-calendar.md](google-calendar.md)), `Scheduler` ([telegram-bot.md](telegram-bot.md#scheduling)) and `A2A`.
`Volcanic.AgenticCore.Extensions.A2A` is the A2A client ([a2a.md](a2a.md#client)): `A2AClientExtension(settings,
target, registry)` registers the `a2a` tool (`agents`, `send`, `status`, `cancel` over the session's
`.volcanic/a2a/agents.json`), keeps the inbound hop count from the `a2a.inbound` event, and on `SessionStartEvent`
attaches the session for live delivery and flushes results kept for its target. It is meant to be added per session
by the host (`AgentSessionOptions.ExtraExtensions`, as the Telegram host does for each chat), not host-wide. Long
remote tasks go to the process-wide `A2AWatchRegistry`, which follows them by push, subscribe or poll, persists them
in `.volcanic/a2a/pending.json` and hands results to an `IA2AResultDelivery`: by default the target's live session
(`SendUserMessage` as a follow-up), or the host's own (`registry.UseDelivery(...)`, then `RestoreAsync` at startup and
`StopAsync` at shutdown). A host with an A2A server routes `A2AHostOptions.OnPush` to `registry.OnPushAsync` and sets
`A2AClientSettings.PushBaseUrl` to the server's public URL.