@akanjs/cli 3.2.0-rc.20 → 3.2.0-rc.22
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/.build-stamp
CHANGED
|
@@ -1,2 +1,2 @@
|
|
|
1
|
-
|
|
2
|
-
|
|
1
|
+
0018601bd7b80c0a8f0b71960a59cf4fc033f87b77335b612f0d4f351f543d35
|
|
2
|
+
b32cfaf2-7c00-4a10-a90a-784494553fc4
|
|
@@ -235,8 +235,8 @@ globalPrefix?: false;
|
|
|
235
235
|
/** HTTP verb of a `mutation`, `POST` unless named; every other endpoint type ignores it. */
|
|
236
236
|
method?: HttpMutationMethod;
|
|
237
237
|
fileUpload?: boolean;
|
|
238
|
-
/** `false` keeps it out of the MCP catalogue (default `true`); curation, not authorization — HTTP still serves it. */
|
|
239
|
-
mcp?:
|
|
238
|
+
/** `false` keeps it out of the MCP catalogue (default `true`), and `"agent"` keeps it off `/mcp` while the in-process agent (`AgentToolsRole`, `agentService.loop`) still lists it; curation, not authorization — HTTP still serves it. */
|
|
239
|
+
mcp?: McpShelf;
|
|
240
240
|
/** Overrides what the key implies for a `mutation`'s MCP `destructiveHint`: `true` for one no verb gives away, such as `archiveOrder`, `false` for a `drop…` that loses nothing. */
|
|
241
241
|
destructive?: boolean;
|
|
242
242
|
/** The errors this endpoint throws on purpose, by status. */
|
|
@@ -46,6 +46,12 @@ export class FileSlice extends slice(srv.file, {
|
|
|
46
46
|
- **A view keeps reaching it.** To a host that renders MCP Apps views, an endpoint refused for `mcp: false` alone
|
|
47
47
|
is published with `visibility: ["app"]`: an app's view may call it, the model is never shown it (below). Every
|
|
48
48
|
other refusal still holds there.
|
|
49
|
+
- **`mcp: "agent"` keeps it off `/mcp` and on the in-process agent.** It goes wherever `mcp: false` goes, a slice
|
|
50
|
+
map key included. `/mcp` neither lists nor runs it and `listingCost` leaves it out, while `AgentToolsRole` and
|
|
51
|
+
`agentService.loop` list it under the same guards and exclusions. Use it when the listing's size, not the step,
|
|
52
|
+
keeps an entry off. A lib's slice and the app's keep the narrowest: `false`, then `"agent"`, then listed. Code
|
|
53
|
+
that reads `signalOption.mcp` to tell whether an entry is listed tests `(mcp ?? true) === true`, never
|
|
54
|
+
`mcp !== false`.
|
|
49
55
|
- **`mcp: true` is the default and is a no-op**; write it only to say out loud that an entry is deliberately on
|
|
50
56
|
the shelf. It does not override a shape refusal (an `Any` return stays refused) and it cannot publish something
|
|
51
57
|
the guard rules keep out.
|
|
@@ -504,14 +510,14 @@ costs something to assemble.
|
|
|
504
510
|
|
|
505
511
|
A server-side agent loop or a queue job reaches the same catalogue with no token and no loopback request:
|
|
506
512
|
`plug(AgentToolsRole)` in a service, then `listAs(account, { allow })` for the tools a model is offered and
|
|
507
|
-
`callAs(account, { id, name, args })` to run one.
|
|
508
|
-
`readOnly`, `Person`, the account guards, masking and the refusal text — and
|
|
509
|
-
`isAgentCall` and `.with(AgentCall)` treat it as an agent's. `allow` only
|
|
510
|
-
registry, so a batch server, which mounts no `/mcp`, has it too; the per-caller
|
|
511
|
-
endpoint's own `rateLimit` does. A tool's failure never throws: it comes back as
|
|
512
|
-
An adaptor may plug the role too. Neither method works before the server has built
|
|
513
|
-
`onInit` throws `InjectError` `notInitialized`; call from a request, a job or a
|
|
514
|
-
are left out: `/mcp` `tools/list` shows them, `listAs` does not, and `callAs` answers `Unknown tool`.
|
|
513
|
+
`callAs(account, { id, name, args })` to run one. It adds every `mcp: "agent"` entry, and everything above still
|
|
514
|
+
holds — `mcp: false`, the exclusions, `readOnly`, `Person`, the account guards, masking and the refusal text — and
|
|
515
|
+
the call's `origin` is `"mcp"`, so `isAgentCall` and `.with(AgentCall)` treat it as an agent's. `allow` only
|
|
516
|
+
narrows. It is built from the DI registry, so a batch server, which mounts no `/mcp`, has it too; the per-caller
|
|
517
|
+
`/mcp` budget does not apply, an endpoint's own `rateLimit` does. A tool's failure never throws: it comes back as
|
|
518
|
+
the result's `error` and `code`. An adaptor may plug the role too. Neither method works before the server has built
|
|
519
|
+
its endpoints, so a call from an `onInit` throws `InjectError` `notInitialized`; call from a request, a job or a
|
|
520
|
+
schedule. A page's `page().app()` tools are left out: `/mcp` `tools/list` shows them, `listAs` does not, and `callAs` answers `Unknown tool`.
|
|
515
521
|
|
|
516
522
|
- **The account is trusted as given.** It is put on the call after the middleware chain, so pass one the server
|
|
517
523
|
resolved, never client input, and resolve a stored one again before reusing it: a revoked session is not noticed.
|
|
@@ -532,6 +538,14 @@ at every event.
|
|
|
532
538
|
Neither is ever client input. The model is offered `allow` ∩ what
|
|
533
539
|
the account sees on `/mcp`, and a call to a name that round did not list is answered as an unknown tool, neither
|
|
534
540
|
asked about nor run. A request with no `allow` array is an `error`.
|
|
541
|
+
- **Toolkits.** `toolkits: { calendar: { description, tools } }` groups names of `allow` so a round need not carry
|
|
542
|
+
every tool. A tool in no toolkit is listed every round, beside one more, `openToolkit({ name })`, whose description
|
|
543
|
+
lists the toolkits. An opened toolkit's tools are listed from the next round on; `openToolkit` never reaches
|
|
544
|
+
`approve` and counts as no tool call. A resume reopens what stored calls and a summary's `toolkits` opened.
|
|
545
|
+
- **The open screen's tools.** `clientTools` takes `new AgentClientTools().list()` from a page the member has open
|
|
546
|
+
(`akanjs/ui`), sent over your own transport. It is client input, and a name an endpoint holds is dropped. The run
|
|
547
|
+
never runs one: `approve` is asked, then the run ends `awaiting` with `client: true`. Have the page `run(call)`,
|
|
548
|
+
append its result — or your own error when no page answers — and resume.
|
|
535
549
|
- **Writes after untrusted input.** `untrusted(result)` marks a result third parties wrote, and every later call
|
|
536
550
|
reaches `approve(call, { tainted, tool })` with `tainted` set. A batch is judged before it runs, since the model
|
|
537
551
|
wrote it before reading its results. Without `approve`, a `needsConfirm` tool — on this catalogue, a mutation
|
package/package.json
CHANGED
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
{
|
|
2
2
|
"name": "@akanjs/cli",
|
|
3
|
-
"version": "3.2.0-rc.
|
|
3
|
+
"version": "3.2.0-rc.22",
|
|
4
4
|
"sourceType": "module",
|
|
5
5
|
"type": "module",
|
|
6
6
|
"publishConfig": {
|
|
@@ -39,7 +39,7 @@
|
|
|
39
39
|
"@earendil-works/pi-tui": "1.3.0",
|
|
40
40
|
"@inquirer/prompts": "^8.7.3",
|
|
41
41
|
"@tailwindcss/node": "^4.3.3",
|
|
42
|
-
"akanjs": "3.2.0-rc.
|
|
42
|
+
"akanjs": "3.2.0-rc.22",
|
|
43
43
|
"chalk": "^6.0.1",
|
|
44
44
|
"commander": "^15.0.0",
|
|
45
45
|
"dayjs": "^1.11.23",
|
package/skills/akan-mcp/SKILL.md
CHANGED
|
@@ -13,7 +13,7 @@ an endpoint with a real guard is published, one with none is refused, and so is
|
|
|
13
13
|
|
|
14
14
|
- `mcp: false` on an endpoint, or a `slice()` `mcp` map mirroring `guards`, keeps a perfectly guarded step (a UI
|
|
15
15
|
state-machine step a model would only call by mistake) off the shelf. It is an opt-out; everything else still
|
|
16
|
-
publishes.
|
|
16
|
+
publishes. `mcp: "agent"` keeps an entry off `/mcp` but listed to `AgentToolsRole` and `agentService.loop`.
|
|
17
17
|
- `guards: [Every, Person]` reserves an act for a person; `.with(AgentCall)` narrows what an agent's call may do.
|
|
18
18
|
- A guard's `static scope` decides whether a listing can hide the entry: `"account"` guards are evaluated per
|
|
19
19
|
caller, `"resource"` guards stay visible and stop the call.
|
|
@@ -35,10 +35,12 @@ an endpoint with a real guard is published, one with none is refused, and so is
|
|
|
35
35
|
## Calling it from the server
|
|
36
36
|
|
|
37
37
|
- `plug(AgentToolsRole)` gives a service `listAs(account, { allow })` / `callAs(account, call)`: the same catalogue
|
|
38
|
-
and refusals in-process, batch servers included. Pass only an account the server resolved, never client input.
|
|
38
|
+
and refusals in-process plus the `mcp: "agent"` entries, batch servers included. Pass only an account the server resolved, never client input.
|
|
39
39
|
- `this.agentService.loop({ account | principal, allow, messages, approve, untrusted, onMessage, meta, budget })`
|
|
40
40
|
runs a model over those tools until it answers; `option.setAgentAccountResolver` resolves a stored `principal`
|
|
41
|
-
before every round. Each round is metered as its own turn with `turn.loop` and `turn.meta`.
|
|
41
|
+
before every round. Each round is metered as its own turn with `turn.loop` and `turn.meta`. `toolkits` holds named
|
|
42
|
+
groups of `allow` back until the model calls `openToolkit`; `clientTools` (from `AgentClientTools` on an open page)
|
|
43
|
+
parks a call as `awaiting { client: true }` for the page to run.
|
|
42
44
|
|
|
43
45
|
## Verify
|
|
44
46
|
|