@cognisphere-sh/cognisphere-harness 0.13.0 → 0.15.0
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/CHANGELOG.md +104 -0
- package/dist-web/assets/{file-editor-CqGYKmT3.js → file-editor-DJydzVQK.js} +1 -1
- package/dist-web/assets/{index-CDE-jsH8.js → index-C2uTXv8R.js} +2 -2
- package/dist-web/index.html +1 -1
- package/home-template/docs/base-harness/README.md +25 -7
- package/home-template/docs/base-harness/skills.md +1 -1
- package/package.json +5 -4
- package/skills/create-plugin/CHANGELOG.md +7 -0
- package/skills/create-plugin/SKILL.md +10 -6
- package/src/agents/base-agent/system_prompts/0-base_prompt.md +2 -0
- package/src/agents/base-agent/workspace/router.ts +47 -0
- package/src/agents/base-agent/workspace/workflows/README.md +71 -0
- package/src/api/admin.ts +1 -1
- package/src/cli/agent.ts +1 -2
- package/src/core/agent-manager.ts +17 -9
- package/src/core/queue.ts +19 -6
- package/src/core/router.ts +258 -0
- package/src/core/runner.ts +273 -66
- package/src/core/types.ts +2 -14
- package/src/plugins/admin/index.ts +2 -2
- package/src/plugins/agent-messaging/index.ts +4 -3
- package/src/plugins/gws/index.ts +8 -120
- package/src/plugins/gws/seed/system_prompts/plugin-gws.md +8 -4
- package/src/plugins/scheduler/index.ts +4 -4
- package/src/plugins/scheduler/seed/system_prompts/plugin-scheduler.md +1 -1
- package/src/plugins/telegram/index.ts +4 -98
- package/src/plugins/telegram/seed/system_prompts/plugin-telegram.md +4 -6
- package/src/plugins/gws/seed/scripts/gws/routes +0 -117
- package/src/plugins/gws/seed/skills/gws/route-email/CHANGELOG.md +0 -14
- package/src/plugins/gws/seed/skills/gws/route-email/SKILL.md +0 -84
- package/src/plugins/telegram/seed/scripts/telegram/routes +0 -95
- package/src/plugins/telegram/seed/skills/telegram/route-chats/CHANGELOG.md +0 -7
- package/src/plugins/telegram/seed/skills/telegram/route-chats/SKILL.md +0 -63
package/dist-web/index.html
CHANGED
|
@@ -16,7 +16,7 @@
|
|
|
16
16
|
} catch (_) {}
|
|
17
17
|
})();
|
|
18
18
|
</script>
|
|
19
|
-
<script type="module" crossorigin src="/assets/index-
|
|
19
|
+
<script type="module" crossorigin src="/assets/index-C2uTXv8R.js"></script>
|
|
20
20
|
<link rel="modulepreload" crossorigin href="/assets/react-BIlrhC09.js">
|
|
21
21
|
<link rel="modulepreload" crossorigin href="/assets/radix-DHW5y7jr.js">
|
|
22
22
|
<link rel="modulepreload" crossorigin href="/assets/markdown-sumxz7Eq.js">
|
|
@@ -61,7 +61,6 @@ plus an optional `X-App-User` header (see `app/README.md`).
|
|
|
61
61
|
|
|
62
62
|
- `agent.json` — `name`, `description?` (one-line role blurb shown to other
|
|
63
63
|
agents in the harness roster), `model: {provider, id, thinkingLevel?}`,
|
|
64
|
-
`threadIdStrategy` (`single` | `plugin` | `plugin_channel`),
|
|
65
64
|
`maxConcurrentSlots?`, `devAgent?` (marks the developer agent), optional
|
|
66
65
|
`secretsSchema`/`configSchema`/`config`.
|
|
67
66
|
- `system_prompts/*.md` — concatenated in lexical order into the system
|
|
@@ -83,6 +82,26 @@ plus an optional `X-App-User` header (see `app/README.md`).
|
|
|
83
82
|
upgrade or restart.
|
|
84
83
|
- `workspace/` — the agent's durable notes: `index.md`, `daily_notes/`, and
|
|
85
84
|
`threads/<ThreadId>/` (`notes.md`, `files/`, `tasks/<task-slug>/notes.md`).
|
|
85
|
+
- `workspace/router.ts` — **the only routing authority**: every event from
|
|
86
|
+
every plugin passes through it before it is queued. Its default export gets
|
|
87
|
+
the message (`notification`, `pluginId`, `channelId`, `text`, `metadata`,
|
|
88
|
+
`isSilent`) plus the workflow list, and returns a thread id, an array of
|
|
89
|
+
them, `workflow:<name>`, `undefined` (`metadata.threadId`, else `main`) or
|
|
90
|
+
`[]` (drop). Synchronous; edits apply within a few seconds (a broken edit
|
|
91
|
+
keeps the previous router). Agent-messaging, the scheduler and the console
|
|
92
|
+
set `metadata.threadId`; Telegram and Gmail don't, so they go to `main`
|
|
93
|
+
until a rule routes them (e.g. `return "telegram:" + msg.channelId`).
|
|
94
|
+
- `workspace/workflows/<name>/` — a workflow: `workflow.json`
|
|
95
|
+
(`{"description", "script"}`, both required) plus an executable script.
|
|
96
|
+
Its thread is `workflow:<name>`: input is queued like any thread's and the
|
|
97
|
+
script runs once per message from the agent folder (message JSON on stdin,
|
|
98
|
+
`PI_THREAD_ID=workflow:<name>`). Exit 0 = done; anything else, a 10-minute
|
|
99
|
+
timeout or an interruption = failed, **never retried**. The harness reports
|
|
100
|
+
only runs it killed (timeout, stop, crash) to `main` (`workflow_killed`). It invokes the agent with `scripts/agent-msg/send`; the
|
|
101
|
+
agent's reply to `workflow:<name>` runs it again. Best practice: wrap the
|
|
102
|
+
script in try/catch and send errors to the agent's `main` thread the same
|
|
103
|
+
way, so the agent can babysit it. See `workflows/README.md` in any
|
|
104
|
+
agent.
|
|
86
105
|
- `knowledge/` — cross-thread reference docs (`index.md`, `files/`) plus
|
|
87
106
|
`memory.md`, the agent's long-lived memory (grep-able sections separated by
|
|
88
107
|
`-----$-----$-----$-----`).
|
|
@@ -99,9 +118,8 @@ plus an optional `X-App-User` header (see `app/README.md`).
|
|
|
99
118
|
here: the harness runs the `pi` bundled with its own package, at the version
|
|
100
119
|
it was tested with.
|
|
101
120
|
|
|
102
|
-
A **thread** is one conversation.
|
|
103
|
-
|
|
104
|
-
plugin+channel).
|
|
121
|
+
A **thread** is one conversation. The agent's router (`workspace/router.ts`)
|
|
122
|
+
decides which thread each inbound message joins; by default `main`.
|
|
105
123
|
|
|
106
124
|
## Plugins
|
|
107
125
|
|
|
@@ -132,9 +150,9 @@ the open `/webhook/*` surface. Artifact HTML is always served with a CSP
|
|
|
132
150
|
sandbox, so it can't reach the app's cookies or storage.
|
|
133
151
|
|
|
134
152
|
The **telegram** plugin long-polls a bot; set
|
|
135
|
-
`secrets.json → <agent>.telegram.TELEGRAM_BOT_TOKEN`.
|
|
136
|
-
|
|
137
|
-
|
|
153
|
+
`secrets.json → <agent>.telegram.TELEGRAM_BOT_TOKEN`. It picks no thread:
|
|
154
|
+
messages go to `main` unless the agent's router routes them (e.g. one thread
|
|
155
|
+
per chat). To reset a conversation, delete its thread from the console.
|
|
138
156
|
|
|
139
157
|
## Secrets and models
|
|
140
158
|
|
|
@@ -63,7 +63,7 @@ A plugin may ship a skill in its `seed/` tree under
|
|
|
63
63
|
start and loaded like any other skill. Those directories are **plugin-owned** —
|
|
64
64
|
overwritten on each start, so never hand-edit them; change the plugin's seed
|
|
65
65
|
instead. Shipped this way: `artifacts` → `publish-artifact`, `gws` →
|
|
66
|
-
`read-email` + `search-email
|
|
66
|
+
`read-email` + `search-email`. A plugin's system
|
|
67
67
|
prompt fragment is kept short (identity, event shape, always-on rules) and
|
|
68
68
|
delegates every step-by-step procedure to one of these skills.
|
|
69
69
|
|
package/package.json
CHANGED
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
{
|
|
2
2
|
"name": "@cognisphere-sh/cognisphere-harness",
|
|
3
|
-
"version": "0.
|
|
3
|
+
"version": "0.15.0",
|
|
4
4
|
"license": "MIT",
|
|
5
5
|
"type": "module",
|
|
6
6
|
"repository": {
|
|
@@ -27,8 +27,8 @@
|
|
|
27
27
|
"access": "public"
|
|
28
28
|
},
|
|
29
29
|
"dependencies": {
|
|
30
|
-
"@earendil-works/pi-ai": "0.
|
|
31
|
-
"@earendil-works/pi-coding-agent": "0.
|
|
30
|
+
"@earendil-works/pi-ai": "1.0.0",
|
|
31
|
+
"@earendil-works/pi-coding-agent": "1.0.0",
|
|
32
32
|
"@hono/node-server": "^1.13.7",
|
|
33
33
|
"ajv": "^8.17.1",
|
|
34
34
|
"better-sqlite3": "^11.7.0",
|
|
@@ -51,6 +51,7 @@
|
|
|
51
51
|
"typecheck": "tsc --noEmit",
|
|
52
52
|
"lint": "eslint src",
|
|
53
53
|
"lint:fix": "eslint src --fix",
|
|
54
|
-
"check": "tsc --noEmit && eslint src"
|
|
54
|
+
"check": "tsc --noEmit && eslint src test",
|
|
55
|
+
"test": "node scripts/test.mjs"
|
|
55
56
|
}
|
|
56
57
|
}
|
|
@@ -1,5 +1,12 @@
|
|
|
1
1
|
# create-plugin — changelog
|
|
2
2
|
|
|
3
|
+
## 1.4.0
|
|
4
|
+
|
|
5
|
+
- Routing moved to the agent's router (`workspace/router.ts`): the `notify`
|
|
6
|
+
payload has no `threadIdOverride`, and plugins no longer ship routing rules.
|
|
7
|
+
§1 says what to publish instead (`channelId`/`metadata` for the router,
|
|
8
|
+
`metadata.threadId` only for thread-targeted events).
|
|
9
|
+
|
|
3
10
|
## 1.3.0
|
|
4
11
|
|
|
5
12
|
- Prompt-fragment budget: **over ~50 lines is the signal to split** — a
|
|
@@ -1,9 +1,9 @@
|
|
|
1
1
|
---
|
|
2
2
|
name: create-plugin
|
|
3
|
-
description: Author a new CogniSphere plugin in a harness's plugins/ directory — scaffold the index.ts + seed, enable it on an agent, and verify it loads and starts. Covers the duck-typed Plugin contract, the manifest (config/secrets schemas), the seed tree (prompt fragment, CLI scripts, plugin-owned skills), event/metadata conventions, enabling per agent, and the common load/start failures. Use when asked to "create a plugin", "author a new plugin", "add a custom plugin" or "write a plugin for the harness" — and whenever a capability an agent needs lives outside its own filesystem: polling or receiving from an external service, exposing an HTTP/webhook endpoint, running something on a schedule, or handing agents a CLI backed by harness-side state. (v1.
|
|
3
|
+
description: Author a new CogniSphere plugin in a harness's plugins/ directory — scaffold the index.ts + seed, enable it on an agent, and verify it loads and starts. Covers the duck-typed Plugin contract, the manifest (config/secrets schemas), the seed tree (prompt fragment, CLI scripts, plugin-owned skills), event/metadata conventions, enabling per agent, and the common load/start failures. Use when asked to "create a plugin", "author a new plugin", "add a custom plugin" or "write a plugin for the harness" — and whenever a capability an agent needs lives outside its own filesystem: polling or receiving from an external service, exposing an HTTP/webhook endpoint, running something on a schedule, or handing agents a CLI backed by harness-side state. (v1.4.0)
|
|
4
4
|
metadata:
|
|
5
5
|
author: cognisphere
|
|
6
|
-
version: "1.
|
|
6
|
+
version: "1.4.0"
|
|
7
7
|
argument-hint: <plugin-id>
|
|
8
8
|
---
|
|
9
9
|
|
|
@@ -48,8 +48,13 @@ harness types, so the plugin has no build step and no dependencies.
|
|
|
48
48
|
`notify(name, payload)`, `httpBaseUrl`, `log`.
|
|
49
49
|
|
|
50
50
|
`notify` payload needs at least `{ text, channelId }`; optional `metadata`,
|
|
51
|
-
`
|
|
52
|
-
|
|
51
|
+
`isSilent`, `doNotSteer`, `priority`. A plugin doesn't pick the thread: the
|
|
52
|
+
agent's router (`workspace/router.ts`) does, from the plugin id, channel,
|
|
53
|
+
text and metadata. Put whatever a routing rule would need (chat id, subject,
|
|
54
|
+
sender) in `channelId`/`metadata`. Set `metadata.threadId` only when the event
|
|
55
|
+
targets a specific thread (as agent-messaging and the scheduler do); the
|
|
56
|
+
router honours it by default. Without either, the event goes to `main`.
|
|
57
|
+
`notify` enqueues the event and wakes the agent (or steers a live run).
|
|
53
58
|
|
|
54
59
|
### Event & metadata conventions
|
|
55
60
|
|
|
@@ -113,8 +118,7 @@ loaded like any other). Version it and give it a `CHANGELOG.md` exactly as
|
|
|
113
118
|
`create-skill` describes; the scope dir `skills/<id>/` marks it plugin-owned,
|
|
114
119
|
so agents must never edit it — change the seed instead. The bundled `artifacts`
|
|
115
120
|
plugin is the worked example (`publish-artifact`); `gws` (`read-email`,
|
|
116
|
-
`
|
|
117
|
-
larger surface.
|
|
121
|
+
`search-email`) is the same split applied to a larger surface.
|
|
118
122
|
|
|
119
123
|
**A prompt fragment over ~50 lines is the signal to split.** It costs every
|
|
120
124
|
turn of every thread, whether or not the plugin is in play; a skill costs one
|
|
@@ -289,6 +289,8 @@ Always use the bare ThreadId as the thread directory name — don't prefix it wi
|
|
|
289
289
|
|
|
290
290
|
`workspace/index.md` is the entry point: pointers to every file and directory in the workspace. Keep it updated. Any subdirectory may carry its own `index.md` for a nested index.
|
|
291
291
|
|
|
292
|
+
**Routing.** `workspace/router.ts` is the only thing that decides which of your threads an incoming event joins — for every plugin. Without a rule, an event goes to the thread it names (agent-messaging, scheduler, console) or to `main`. To route events (e.g. one thread per Telegram chat or Gmail thread, or a workflow for a kind of email), edit the router; edits apply within a few seconds, and a broken edit keeps the previous router (check the harness log). A workflow is a script in `workspace/workflows/<name>/` that handles events routed to the thread `workflow:<name>` without you (see that folder's `README.md`). A message from a workflow arrives as an `agent_message` with `FromThread: workflow:<name>`; reply to that thread to hand the result back to the workflow (a reply runs it again). You babysit your workflows. A failed run is never retried, and the harness doesn't report ordinary failures, so when you write a workflow, wrap it in try/catch and send errors to yourself on `main` with `scripts/agent-msg/send` (see the workflows `README.md`). Only when the harness kills a script (timeout, stop, crash) does it post a `workflow_killed` notice to `main` itself (plugin `workflow`). When either arrives, read it and the script, fix the script, then re-send the work to `workflow:<name>` or ask the operator to requeue the event.
|
|
293
|
+
|
|
292
294
|
**Workspace is for what must persist — not scratch.** Write intermediate/throwaway files (temp conversions, scratch parsing output, working copies) under `/tmp`, or delete them once you're done; don't leave them in `workspace/`. Likewise, don't copy input files from `plugins/<id>/inbox/` into `workspace/` by default — read them in place. Copy one into `threads/<ThreadId>/files/` only when it genuinely needs to outlive the inbox.
|
|
293
295
|
|
|
294
296
|
## Notes
|
|
@@ -0,0 +1,47 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* Message router — the only thing that decides which thread a message goes to.
|
|
3
|
+
* Every event from every plugin (Telegram, Gmail, scheduler, agent-messaging,
|
|
4
|
+
* the console, …) passes through `route` before it is queued. Return:
|
|
5
|
+
*
|
|
6
|
+
* - `"<threadId>"` one thread
|
|
7
|
+
* - `["a", "b"]` several threads (one queued copy each)
|
|
8
|
+
* - `"workflow:<name>"` a workflow thread: the script in
|
|
9
|
+
* `workspace/workflows/<name>/` runs once per message
|
|
10
|
+
* - `[]` drop the message
|
|
11
|
+
* - `undefined` `msg.metadata.threadId` if set, else `"main"`
|
|
12
|
+
*
|
|
13
|
+
* Plugins that target a thread (agent-messaging, the scheduler, the console)
|
|
14
|
+
* set `metadata.threadId`; the others (Telegram, Gmail) leave it to you, so
|
|
15
|
+
* write rules for them here. Keep it synchronous and fast: it runs inside the
|
|
16
|
+
* harness for every message. Edits apply within a few seconds; an edit that
|
|
17
|
+
* fails to load keeps the previous router, and a router that throws or returns
|
|
18
|
+
* an unsafe or unknown id sends the message to the default above.
|
|
19
|
+
*/
|
|
20
|
+
|
|
21
|
+
export interface RouterMessage {
|
|
22
|
+
/** Event name, e.g. `message_received`, `email_received`, `schedule_fire`, `agent_message`. */
|
|
23
|
+
notification: string;
|
|
24
|
+
pluginId: string;
|
|
25
|
+
/** The source conversation: Telegram chat id, Gmail thread id, … */
|
|
26
|
+
channelId: string;
|
|
27
|
+
text: string;
|
|
28
|
+
metadata: Record<string, unknown>;
|
|
29
|
+
isSilent: boolean;
|
|
30
|
+
}
|
|
31
|
+
|
|
32
|
+
export interface WorkflowInfo {
|
|
33
|
+
name: string;
|
|
34
|
+
description: string;
|
|
35
|
+
}
|
|
36
|
+
|
|
37
|
+
export default function route(
|
|
38
|
+
msg: RouterMessage,
|
|
39
|
+
workflows: WorkflowInfo[],
|
|
40
|
+
): string | string[] | undefined {
|
|
41
|
+
void workflows;
|
|
42
|
+
// Rules go here, for example:
|
|
43
|
+
// if (msg.pluginId === "telegram") return `telegram:${msg.channelId}`; // one thread per chat
|
|
44
|
+
// if (msg.pluginId === "gws" && /invoice/i.test(String(msg.metadata.Subject))) return "workflow:invoices";
|
|
45
|
+
const threadId = msg.metadata.threadId;
|
|
46
|
+
return typeof threadId === "string" && threadId ? threadId : "main";
|
|
47
|
+
}
|
|
@@ -0,0 +1,71 @@
|
|
|
1
|
+
# Workflows
|
|
2
|
+
|
|
3
|
+
A workflow is a static script that runs instead of a Pi turn. The router
|
|
4
|
+
(`workspace/router.ts`) sends a message to it by returning the thread id
|
|
5
|
+
`workflow:<name>`. Each workflow is a folder:
|
|
6
|
+
|
|
7
|
+
```
|
|
8
|
+
workflows/<name>/
|
|
9
|
+
workflow.json {"description": "What it does and when to route to it.", "script": "run.sh"}
|
|
10
|
+
run.sh the script (executable: chmod +x)
|
|
11
|
+
```
|
|
12
|
+
|
|
13
|
+
The folder name is the workflow's name. `description` and `script` are
|
|
14
|
+
required; a folder missing either, or whose script isn't executable, is skipped.
|
|
15
|
+
|
|
16
|
+
Messages to a workflow are queued like any thread's, and the script runs once
|
|
17
|
+
per message, one at a time, from the agent folder:
|
|
18
|
+
|
|
19
|
+
- **stdin:** the message as JSON: `id`, `notification`, `pluginId`,
|
|
20
|
+
`channelId`, `threadId`, `text`, `metadata`, `isSilent`, `attempts`.
|
|
21
|
+
- **env:** the same as a Pi turn (`.venv`, secrets, `PI_AGENT_ID`,
|
|
22
|
+
`HARNESS_BASE_URL`), with `PI_THREAD_ID=workflow:<name>`.
|
|
23
|
+
- **result:** exit 0 marks the message done. Anything else (or running past
|
|
24
|
+
10 minutes, or being interrupted by a stop or crash) marks it **failed,
|
|
25
|
+
with no retry**. For an ordinary failure the harness only logs it and shows
|
|
26
|
+
it in the console's events; telling the agent is the script's job (below).
|
|
27
|
+
When the harness itself kills the script (timeout, stop, crash), the
|
|
28
|
+
script's handler can't run, so the harness posts a `workflow_killed`
|
|
29
|
+
notice to `main` instead. `attempts` > 0 on stdin means an operator
|
|
30
|
+
requeued it after an earlier run failed part-way.
|
|
31
|
+
- **To invoke the agent:** `scripts/agent-msg/send --to-agent "$PI_AGENT_ID" --thread-id <thread> --message "…"`.
|
|
32
|
+
The agent's reply to `workflow:<name>` arrives as a new message
|
|
33
|
+
(`notification: "agent_message"`), so don't wait for it in the script.
|
|
34
|
+
|
|
35
|
+
## Best practice: catch errors and tell the agent
|
|
36
|
+
|
|
37
|
+
The agent babysits its workflows, but only hears about a failure if the
|
|
38
|
+
script tells it. Wrap the work in try/catch and send the agent what it needs
|
|
39
|
+
to act, with agent-messaging, to `main` (or any thread you choose):
|
|
40
|
+
|
|
41
|
+
```python
|
|
42
|
+
#!/usr/bin/env python3
|
|
43
|
+
import json, os, subprocess, sys, traceback
|
|
44
|
+
|
|
45
|
+
msg = json.load(sys.stdin)
|
|
46
|
+
|
|
47
|
+
def tell_agent(text, thread="main"):
|
|
48
|
+
subprocess.run(["scripts/agent-msg/send", "--to-agent", os.environ["PI_AGENT_ID"],
|
|
49
|
+
"--thread-id", thread, "--subject", "workflow invoices", "--message", text])
|
|
50
|
+
|
|
51
|
+
try:
|
|
52
|
+
... # the workflow
|
|
53
|
+
except Exception:
|
|
54
|
+
tell_agent(f"invoices failed on event {msg['id']} ({msg['text'][:200]!r}):\n{traceback.format_exc()}")
|
|
55
|
+
sys.exit(1) # marks the event failed in the console
|
|
56
|
+
```
|
|
57
|
+
|
|
58
|
+
In a shell script, use `bash` with
|
|
59
|
+
`trap 'scripts/agent-msg/send --to-agent "$PI_AGENT_ID" --thread-id main --message "…failed at line $LINENO"; exit 1' ERR`.
|
|
60
|
+
A timeout, stop or crash kills the script before its handler runs; the
|
|
61
|
+
harness reports those to `main` itself.
|
|
62
|
+
|
|
63
|
+
Messages a workflow sends carry `FromThread: workflow:<name>`, so an agent
|
|
64
|
+
reply to them runs the workflow again. To redo the work after a fix, the agent
|
|
65
|
+
sends it to `workflow:<name>` (or the operator requeues the failed event in
|
|
66
|
+
the console).
|
|
67
|
+
|
|
68
|
+
Output goes to the harness log. A run ends when the script's stdout and stderr
|
|
69
|
+
close, so don't leave a child running in the background: it holds the worker
|
|
70
|
+
until it exits or the 10-minute timeout kills it. Edits to workflows and the
|
|
71
|
+
router apply within a few seconds.
|
package/src/api/admin.ts
CHANGED
package/src/cli/agent.ts
CHANGED
|
@@ -60,7 +60,7 @@ export function scaffoldAgent(
|
|
|
60
60
|
}
|
|
61
61
|
}
|
|
62
62
|
|
|
63
|
-
// Starter config — edit model
|
|
63
|
+
// Starter config — edit the model before first run. The agent stays
|
|
64
64
|
// "failed" until its model provider is configured (secrets + models.json).
|
|
65
65
|
writeJson(join(target, "agent.json"), {
|
|
66
66
|
name,
|
|
@@ -68,7 +68,6 @@ export function scaffoldAgent(
|
|
|
68
68
|
? "Developer agent: owns and modifies this deployment's platform code — agent prompts/scripts, forked plugins, the user-facing frontend app, deploy scripts — and keeps docs/harness + docs/app current. Send code, app, docs and software-install requests here."
|
|
69
69
|
: `TODO: one-line description of ${name}'s role (shown to other agents in the harness roster).`,
|
|
70
70
|
model: { provider: "anthropic", id: "claude-sonnet-4-6" },
|
|
71
|
-
threadIdStrategy: { type: "single" },
|
|
72
71
|
...(opts.dev ? { devAgent: true } : {}),
|
|
73
72
|
});
|
|
74
73
|
return target;
|
|
@@ -10,6 +10,7 @@ import type { PluginRegistry } from "./plugin-registry.js";
|
|
|
10
10
|
import { CORE_PLUGIN_IDS } from "./plugin-registry.js";
|
|
11
11
|
import { AgentDb } from "./queue.js";
|
|
12
12
|
import { AgentRunner } from "./runner.js";
|
|
13
|
+
import { loadRouting, seedRouting, watchRouting } from "./router.js";
|
|
13
14
|
import { findProviderInCatalog } from "./models-catalog.js";
|
|
14
15
|
import { modelUnavailableReason } from "./model-access.js";
|
|
15
16
|
import { ModelsStore } from "./models-store.js";
|
|
@@ -59,6 +60,8 @@ export interface AgentInstance {
|
|
|
59
60
|
* a fresh runner with the new config. Cleared synchronously when the
|
|
60
61
|
* swap fires. */
|
|
61
62
|
staleReason: string | null;
|
|
63
|
+
/** Stops the router/workflows reload watcher; set while running. */
|
|
64
|
+
stopRoutingWatch: (() => void) | null;
|
|
62
65
|
}
|
|
63
66
|
|
|
64
67
|
export type LifecycleErrorCode = "not_found" | "conflict";
|
|
@@ -207,7 +210,7 @@ export class AgentManager {
|
|
|
207
210
|
* Tear down one thread: refuse while a batch is in-flight (it would race
|
|
208
211
|
* the runner writing the session jsonl), delete the thread's queue rows +
|
|
209
212
|
* session binding, remove its `sessions/<threadId>/` dir. Shared by the
|
|
210
|
-
* HTTP DELETE route
|
|
213
|
+
* HTTP DELETE route.
|
|
211
214
|
*/
|
|
212
215
|
deleteThread(
|
|
213
216
|
id: string,
|
|
@@ -350,6 +353,7 @@ export class AgentManager {
|
|
|
350
353
|
changedAt: Date.now(),
|
|
351
354
|
plugins: new Map(),
|
|
352
355
|
staleReason: null,
|
|
356
|
+
stopRoutingWatch: null,
|
|
353
357
|
};
|
|
354
358
|
this.agents.set(id, inst);
|
|
355
359
|
try {
|
|
@@ -478,6 +482,12 @@ export class AgentManager {
|
|
|
478
482
|
// delays later agents.
|
|
479
483
|
await runBootstrap(dir, this.log);
|
|
480
484
|
|
|
485
|
+
// 2.6. Load the agent's router and workflows. Agents created before
|
|
486
|
+
// routing existed get the starter router (write-if-absent).
|
|
487
|
+
const routingLog = childLogger(`agent:${inst.id}`);
|
|
488
|
+
seedRouting(dir, routingLog);
|
|
489
|
+
const routing = await loadRouting(dir, routingLog);
|
|
490
|
+
|
|
481
491
|
// 3. Construct a fresh runner with the resolved env snapshot.
|
|
482
492
|
inst.runner = new AgentRunner({
|
|
483
493
|
rootDir: this.cfg.rootDir,
|
|
@@ -488,6 +498,7 @@ export class AgentManager {
|
|
|
488
498
|
serverBaseUrl: this.cfg.serverBaseUrl,
|
|
489
499
|
timezone: this.cfg.timezone,
|
|
490
500
|
envSecrets,
|
|
501
|
+
routing,
|
|
491
502
|
resolveProviderEnv: (providerId) =>
|
|
492
503
|
resolveProviderEnv(this.models, providerId, dir, this.log),
|
|
493
504
|
modelUnavailableReason: (providerId, modelId) =>
|
|
@@ -497,6 +508,9 @@ export class AgentManager {
|
|
|
497
508
|
// When a stale-reload is pending and the last active batch finishes,
|
|
498
509
|
// swap to a fresh runner. The listener captures `inst` (stable) and
|
|
499
510
|
// checks `inst.runner` at fire time so a swap-in-progress no-ops.
|
|
511
|
+
// The agent edits its own router; apply edits without a restart.
|
|
512
|
+
const runner = inst.runner;
|
|
513
|
+
inst.stopRoutingWatch = watchRouting(dir, routing, (r) => runner.setRouting(r), routingLog);
|
|
500
514
|
inst.runner.on("batch-completed", () => {
|
|
501
515
|
if (inst.staleReason && inst.runner?.activeCount === 0) {
|
|
502
516
|
void this.performStaleSwap(inst);
|
|
@@ -566,6 +580,8 @@ export class AgentManager {
|
|
|
566
580
|
entry.changedAt = Date.now();
|
|
567
581
|
}
|
|
568
582
|
|
|
583
|
+
inst.stopRoutingWatch?.();
|
|
584
|
+
inst.stopRoutingWatch = null;
|
|
569
585
|
if (inst.runner) {
|
|
570
586
|
try {
|
|
571
587
|
await inst.runner.stop();
|
|
@@ -676,14 +692,6 @@ export class AgentManager {
|
|
|
676
692
|
log.error({ err }, "notify failed");
|
|
677
693
|
}
|
|
678
694
|
},
|
|
679
|
-
resetThread: (channelId, threadIdOverride) => {
|
|
680
|
-
const runner = inst.runner;
|
|
681
|
-
if (!runner) throw new Error("agent not running");
|
|
682
|
-
const threadId =
|
|
683
|
-
threadIdOverride ?? runner.threadIdFor(pluginId, channelId);
|
|
684
|
-
this.deleteThread(inst.id, threadId);
|
|
685
|
-
log.info({ threadId }, "thread reset by plugin");
|
|
686
|
-
},
|
|
687
695
|
};
|
|
688
696
|
|
|
689
697
|
await pluginInstance.start(ctx);
|
package/src/core/queue.ts
CHANGED
|
@@ -226,12 +226,15 @@ export class AgentDb {
|
|
|
226
226
|
* this — `doNotSteer` only forbids steering into a live batch, not forming
|
|
227
227
|
* a fresh prompt.
|
|
228
228
|
*
|
|
229
|
+
* - `limit` — claim at most this many rows (workflow threads run one row
|
|
230
|
+
* per script run).
|
|
231
|
+
*
|
|
229
232
|
* A plain call (no options) claims everything still queued — resend + fresh
|
|
230
233
|
* rows — which the runner concatenates into one prompt.
|
|
231
234
|
*/
|
|
232
235
|
dequeueBatch(
|
|
233
236
|
threadId: string,
|
|
234
|
-
opts: { continueOnly?: boolean; excludeDoNotSteer?: boolean } = {},
|
|
237
|
+
opts: { continueOnly?: boolean; excludeDoNotSteer?: boolean; limit?: number } = {},
|
|
235
238
|
): BatchMessage[] {
|
|
236
239
|
const conds = ["thread_id = ?", "status = 'queued'"];
|
|
237
240
|
if (opts.continueOnly) conds.push("pi_entry_id IS NOT NULL");
|
|
@@ -240,7 +243,8 @@ export class AgentDb {
|
|
|
240
243
|
const txn = this.db.transaction((): BatchMessage[] => {
|
|
241
244
|
const rows = this.db
|
|
242
245
|
.prepare<unknown[], EventRow>(
|
|
243
|
-
`SELECT * FROM events WHERE ${where} ORDER BY priority DESC, id ASC
|
|
246
|
+
`SELECT * FROM events WHERE ${where} ORDER BY priority DESC, id ASC` +
|
|
247
|
+
(opts.limit ? ` LIMIT ${Math.floor(opts.limit)}` : ""),
|
|
244
248
|
)
|
|
245
249
|
.all(threadId);
|
|
246
250
|
if (rows.length === 0) return [];
|
|
@@ -488,13 +492,22 @@ export class AgentDb {
|
|
|
488
492
|
return { events };
|
|
489
493
|
}
|
|
490
494
|
|
|
491
|
-
/** Crash recovery: any row stuck in 'in_flight' is treated as a failed
|
|
492
|
-
|
|
495
|
+
/** Crash recovery: any row stuck in 'in_flight' is treated as a failed
|
|
496
|
+
* attempt. Rows on a thread starting with `noRetryPrefix` (workflow
|
|
497
|
+
* threads, which never retry) fail outright. */
|
|
498
|
+
sweepInFlight(maxAttempts: number, error = "process died mid-batch", noRetryPrefix?: string) {
|
|
493
499
|
const rows = this.db
|
|
494
|
-
.prepare<unknown[], { id: number }>(
|
|
495
|
-
`SELECT id FROM events WHERE status = 'in_flight'`,
|
|
500
|
+
.prepare<unknown[], { id: number; thread_id: string }>(
|
|
501
|
+
`SELECT id, thread_id FROM events WHERE status = 'in_flight'`,
|
|
496
502
|
)
|
|
497
503
|
.all();
|
|
504
|
+
if (noRetryPrefix) {
|
|
505
|
+
const noRetry = rows.filter((r) => r.thread_id.startsWith(noRetryPrefix)).map((r) => r.id);
|
|
506
|
+
const rest = rows.filter((r) => !r.thread_id.startsWith(noRetryPrefix)).map((r) => r.id);
|
|
507
|
+
const a = this.markBatchFailed(noRetry, `[crash] ${error}`, 1, null);
|
|
508
|
+
const b = this.markBatchFailed(rest, `[crash] ${error}`, maxAttempts, null);
|
|
509
|
+
return { retrying: b.retrying, dead: [...a.dead, ...b.dead] };
|
|
510
|
+
}
|
|
498
511
|
if (rows.length === 0) return { retrying: [], dead: [] };
|
|
499
512
|
// sessionId=null preserves any existing pi_session_id via COALESCE.
|
|
500
513
|
return this.markBatchFailed(rows.map((r) => r.id), `[crash] ${error}`, maxAttempts, null);
|