@akanjs/cli 3.2.0-rc.22 → 3.2.0-rc.24
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
|
+
dcbcd94203ef4c5c9ad76f291cfe7eb19ce55485ab0d278f65cb2ef180b0563f
|
|
2
|
+
57fb187c-e934-47af-b251-de7eb0dfce25
|
|
@@ -69,7 +69,7 @@ pages declaring one name never clash, and a registry kept outside the surface re
|
|
|
69
69
|
panel the app already titles, and `defaultDraft` opens the composer with text in it without sending it. A panel driven by a
|
|
70
70
|
controlled `open` with **no** `onOpenChange` draws no close button at all rather than an inert one. Then the
|
|
71
71
|
slots: `AgentLauncher`, `AgentSteps`, `AgentBubble`, `AgentComposer`, `AgentApproval`,
|
|
72
|
-
`AgentQuestion`, `AgentToolCard`, `AgentQueued`, `AgentMenu`, `AgentMarkdown` and `AgentCode` each replace one part in `_overrides.tsx`, and
|
|
72
|
+
`AgentQuestion`, `AgentToolCard`, `AgentToolRow`, `AgentQueued`, `AgentMenu`, `AgentMarkdown` and `AgentCode` each replace one part in `_overrides.tsx`, and
|
|
73
73
|
`akanjs/ui` exports every default beside them (`DefaultBubble`, `DefaultComposer`, …) so a skin composes the
|
|
74
74
|
one it is replacing. **`AgentSteps` is one agent turn** — the messages between the user message that opened it
|
|
75
75
|
and the next one, with `isRunning` saying whether more are still coming — which is the grain a chat needs to
|
|
@@ -126,9 +126,10 @@ pages declaring one name never clash, and a registry kept outside the surface re
|
|
|
126
126
|
carries `{ provider, status }`, the hostname and the HTTP status; the provider's own sentence goes to the server
|
|
127
127
|
log only, because it can quote an account id or part of a key and `Err` data reaches the browser.
|
|
128
128
|
`LlmOverflow.refusal(host, status, reason)` builds it, and reads the one refusal a chat can undo — a prompt past
|
|
129
|
-
the window, recognized from each provider's own sentence — as `agent.error.contextOverflow` with the `limit`
|
|
130
|
-
named, which the relay flags so the chat compacts and asks again. An adaptor an app wrote throws through the same function, so its refusals
|
|
131
|
-
translate and recover the way the shipped ones do without a key in the framework's dictionary it cannot add.
|
|
129
|
+
the window, recognized from each provider's own sentence — as `agent.error.contextOverflow` with the `limit` and
|
|
130
|
+
`requested` it named, which the relay flags so the chat compacts and asks again. An adaptor an app wrote throws through the same function, so its refusals
|
|
131
|
+
translate and recover the way the shipped ones do without a key in the framework's dictionary it cannot add. An
|
|
132
|
+
adaptor's own size cap throws that error with `{ limit, requested }` in its own unit, bytes included.
|
|
132
133
|
- **`accepts` is answered per model, through `option.setLlm({ accepts })`.** An adaptor answers for an API and one
|
|
133
134
|
API serves models that differ, so the override rides beside the `model` it is a fact about. It is not a table
|
|
134
135
|
the framework keeps: a table is a claim about models that ship after it, and getting this wrong is the worst
|
|
@@ -603,8 +604,8 @@ pages declaring one name never clash, and a registry kept outside the surface re
|
|
|
603
604
|
`window − answer ceiling (8,192 when unreported) − compact.buffer (13,000, the margin Claude Code leaves)` —
|
|
604
605
|
measured by the provider's own count of the last turn (`done.usage`, kept on that assistant message) plus the
|
|
605
606
|
four-character estimate for what arrived since, so a Korean transcript the character rule reads as small is
|
|
606
|
-
still caught. The window comes from `option.setLlm({ contextWindow })` or from
|
|
607
|
-
with neither, only the ceiling runs. **A refusal for length is answered once**: a relay flags it `overflow` on
|
|
607
|
+
still caught. The window comes from `option.setLlm({ contextWindow })` or from a refusal that names it, scaled by
|
|
608
|
+
its `limit / requested` to this estimate when it gives both; with neither, only the ceiling runs. **A refusal for length is answered once**: a relay flags it `overflow` on
|
|
608
609
|
the wire, and the session compacts and sends the same turn again, and a second refusal in the same send fails
|
|
609
610
|
with `agent.error.contextOverflow`. Either way the history above the last `keep` messages
|
|
610
611
|
becomes one message standing in for it, flagged `summary` on the wire — `<Agent.Chat compact={{ at, keep, buffer }} />`
|
|
@@ -791,7 +792,9 @@ pages declaring one name never clash, and a registry kept outside the surface re
|
|
|
791
792
|
keeps a free-text row either way, because the model wrote the options and only the user knows whether the answer
|
|
792
793
|
is among them. The loop parks on the question exactly as it parks on an approval, a dismissal is the tool's
|
|
793
794
|
error result rather than a silent empty answer, and the settled exchange renders as question-and-answer instead
|
|
794
|
-
of a tool row.
|
|
795
|
+
of a tool row. A restored transcript that ends on an unanswered `askUser` — a server run that parked on it while
|
|
796
|
+
no chat was open, or a reload mid-question — opens on that question, and settling it runs the turn on through the
|
|
797
|
+
runner; a turn started first closes it as unanswered instead. **Never re-implement it per screen** — a `st.tool("askAboutX")` that opens a modal is the same
|
|
795
798
|
thing with a worse transcript — and a hook tool named `askUser` shadows it like any other built-in. The line
|
|
796
799
|
between it and a `.card()` tool is what the answer *is*: a sentence or a pick off a list is `askUser` and belongs
|
|
797
800
|
to the conversation, while a shape the model must not invent — fields, a form, a signature — is a card the screen
|
|
@@ -546,6 +546,10 @@ at every event.
|
|
|
546
546
|
(`akanjs/ui`), sent over your own transport. It is client input, and a name an endpoint holds is dropped. The run
|
|
547
547
|
never runs one: `approve` is asked, then the run ends `awaiting` with `client: true`. Have the page `run(call)`,
|
|
548
548
|
append its result — or your own error when no page answers — and resume.
|
|
549
|
+
- **Asking the member.** `ask: true` lists `askUser({ question, choices?, multiple? })`, the in-page chat's own tool.
|
|
550
|
+
A call ends the run `awaiting` with `question: true`; put the question to the member, append
|
|
551
|
+
`{ id, name: "askUser", result }` — or an `error` when they dismiss it — and resume. It never reaches `approve`,
|
|
552
|
+
counts as no tool call, and `untrusted` never sees the answer. A `clientTools` entry by that name is dropped.
|
|
549
553
|
- **Writes after untrusted input.** `untrusted(result)` marks a result third parties wrote, and every later call
|
|
550
554
|
reaches `approve(call, { tainted, tool })` with `tainted` set. A batch is judged before it runs, since the model
|
|
551
555
|
wrote it before reading its results. Without `approve`, a `needsConfirm` tool — on this catalogue, a mutation
|
|
@@ -571,6 +575,10 @@ at every event.
|
|
|
571
575
|
short neutral prompt stands in, as each adaptor's own fallback is about a screen. `compact { at, keep }` summarizes
|
|
572
576
|
the head, and a tool result is capped at 20,000 characters as the browser caps it. Its calls carry no ip, so an
|
|
573
577
|
ip-scoped `rateLimit` counts them under one unknown address.
|
|
578
|
+
- **A refusal for length.** With `compact` set, a round whose call fails as `agent.error.contextOverflow` compacts
|
|
579
|
+
regardless of `at` and asks once more, as one round; a second refusal, or nothing to cut, ends the run with the
|
|
580
|
+
error. An adaptor with its own size cap throws that error with `{ limit, requested }` in its own unit, bytes
|
|
581
|
+
included, and the run lowers `at` to where that request would have fit.
|
|
574
582
|
|
|
575
583
|
## Rolling MCP Out To An App
|
|
576
584
|
|
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.24",
|
|
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.24",
|
|
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
|
@@ -40,7 +40,8 @@ an endpoint with a real guard is published, one with none is refused, and so is
|
|
|
40
40
|
runs a model over those tools until it answers; `option.setAgentAccountResolver` resolves a stored `principal`
|
|
41
41
|
before every round. Each round is metered as its own turn with `turn.loop` and `turn.meta`. `toolkits` holds named
|
|
42
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
|
|
43
|
+
parks a call as `awaiting { client: true }` for the page to run; `ask: true` lists `askUser`, which parks as
|
|
44
|
+
`awaiting { question: true }` until you append the member's answer.
|
|
44
45
|
|
|
45
46
|
## Verify
|
|
46
47
|
|