Skip to content

The ultimate harness for coding agents and beyond. All your agents. All your machines. One command center. Start with code, then follow your curiosity and build across disciplines: CAD, circuits, robots, games and music.

Merge batch 11-05 i1710b: #3156 (#3162)

0xken15197cc242de51
diff --git a/cli/src/lib/swarmSkills.ts b/cli/src/lib/swarmSkills.ts
index 8f6aa6c22..5eca1560f 100644
--- a/cli/src/lib/swarmSkills.ts
+++ b/cli/src/lib/swarmSkills.ts
@@ -842,9 +842,9 @@ function prepareBuddyFolder(root: string, key: string): void {
ignoreBuddyFiles(root)
}
/** Why a Buddy is not prepared in a Harness Team account: the server's own words (backend buddyContract.ts). */
-const BUDDY_NOT_HERE = 'A Buddy runs only on its owner\'s own machine, signed in as them. It is not started here.'
+export const BUDDY_NOT_HERE = 'A Buddy runs only on its owner\'s own machine, signed in as them. It is not started here.'
/** Why a Buddy saved before its start recorded its owner is not relaunched: its owner starts it again, and that start checks. */
-const BUDDY_START_AGAIN = 'This Buddy was started before Harness checked its machine at each launch. Start your Buddy again.'
+export const BUDDY_START_AGAIN = 'This Buddy was started before Harness checked its machine at each launch. Start your Buddy again.'
/** Which built-in text a session's skill file holds and of which mode, how many times the file was rewritten, and the text,
* mode and rewrite its agent was last confirmed told about (the launch prompt is rewrite 0). A file from before this record
diff --git a/cli/src/superTabs/buddyModel.spec.ts b/cli/src/superTabs/buddyModel.spec.ts
index 86a5122c6..94363794e 100644
--- a/cli/src/superTabs/buddyModel.spec.ts
+++ b/cli/src/superTabs/buddyModel.spec.ts
@@ -5,6 +5,7 @@ import { mkdtempSync } from 'node:fs'
import { tmpdir } from 'node:os'
import { join } from 'node:path'
import { describe, expect, it } from 'vitest'
+import { BUDDY_NOT_HERE, BUDDY_START_AGAIN } from '../lib/swarmSkills.js'
import { BuddyModelRequests, type BuddyModelRow } from './buddyModel.js'
const GRID = 'swarm-0123456789ab'
@@ -85,6 +86,25 @@ describe('carrying out a Buddy\'s model request', () => {
}
})
+ // Task #3256: a switch relaunches the Buddy, and the Buddy guard (lib/swarmSkills.ts prepareBuddyFolder) holds one started
+ // before its owner was recorded, or one not on its owner's machine. The launch reports SWARM_SKILL_FAILED with the guard's
+ // sentence as its detail (lib/launchOverrides.ts); its owner reads what to do, never the code or the detail.
+ it("says Start your Buddy again when the Buddy guard holds the relaunch, and never the guard's detail", async () => {
+ const held = (sentence: string) => ({ error: 'SWARM_SKILL_FAILED', detail: String(new Error(sentence)) })
+ const cases: Array<[Record<string, unknown>, string]> = [
+ [held(BUDDY_START_AGAIN), 'Start your Buddy again to switch its model'],
+ [held(BUDDY_NOT_HERE), 'Your Buddy runs only on your own computer, signed in as you'],
+ [{ error: 'SWARM_SKILL_FAILED', detail: 'Error: EACCES /Users/someone/.claude' }, 'its computer could not switch it (SWARM_SKILL_FAILED)'],
+ [{ error: 'RESPAWN_FAILED', detail: String(new Error(BUDDY_START_AGAIN)) }, 'its computer could not switch it (RESPAWN_FAILED)'],
+ ]
+ for (const [answer, error] of cases) {
+ const f = fixture({ answers: [answer] })
+ await f.runner.serve(main(), [thread('h-a')])
+ expect(f.frames).toHaveLength(1)
+ expect(f.posts).toEqual([{ asAgentId: 'buddy', type: 'buddy.modelResult', agentId: 'buddy', requestId: 'req1', ok: false, error }])
+ }
+ })
+
it('waits for the turn to end before sending a Buddy home by itself, then moves it', async () => {
const f = fixture({ answers: [{ error: 'AGENT_BUSY' }, { retargeted: true }] })
const home = main(request({ gridModel: null, gridName: undefined, reason: 'limit' }))
diff --git a/cli/src/superTabs/buddyModel.ts b/cli/src/superTabs/buddyModel.ts
index 5cb4f1a26..233628deb 100644
--- a/cli/src/superTabs/buddyModel.ts
+++ b/cli/src/superTabs/buddyModel.ts
@@ -13,6 +13,7 @@ import { existsSync, renameSync, writeFileSync } from 'node:fs'
import { join } from 'node:path'
import type { Daemon } from '../agentClient/daemon.js'
import { readPrivateStateFile, secureStateDirectory } from '../lib/secureState.js'
+import { BUDDY_NOT_HERE, BUDDY_START_AGAIN } from '../lib/swarmSkills.js'
import { identifier, object } from './wire.js'
/** A conversation this share of the model's window or more starts fresh: the rest is the room its next turns need. */
@@ -51,9 +52,14 @@ function requestOf(value: unknown): ModelRequest | null {
...(typeof r.reason === 'string' ? { reason: r.reason } : {}) }
}
-/** Why a move was refused, in words its owner can act on: never the daemon's detail, which can name paths on this machine. */
-function refusal(code: string, request: ModelRequest): string {
+/** Why a move was refused, in words its owner can act on: never the daemon's detail, which can name paths on this machine.
+ * The detail is only read for the Buddy guard's two holds (lib/swarmSkills.ts prepareBuddyFolder), which every relaunch runs
+ * and the launch reports as SWARM_SKILL_FAILED with the guard's own sentence (lib/launchOverrides.ts). Task #3256: Scholes's
+ * switch to his saved API, on a Buddy started before its owner was recorded, said only the code. */
+function refusal(code: string, request: ModelRequest, detail?: string): string {
if (code === 'AGENT_BUSY') return 'Your Buddy is answering. Try again when it has finished'
+ if (code === 'SWARM_SKILL_FAILED' && detail?.includes(BUDDY_START_AGAIN)) return 'Start your Buddy again to switch its model'
+ if (code === 'SWARM_SKILL_FAILED' && detail?.includes(BUDDY_NOT_HERE)) return 'Your Buddy runs only on your own computer, signed in as you'
if ((code === 'GRID_UNAVAILABLE' || code === 'API_UNAVAILABLE') && request.gridModel) {
const name = request.name ?? request.gridModel
return request.servedBy?.length ? `${request.servedBy.join(', ')} isn't serving ${name} now` : `Nobody serves ${name} now`
@@ -101,23 +107,24 @@ export class BuddyModelRequests {
private async carryOut(row: BuddyModelRow, sessions: readonly BuddyModelRow[], request: ModelRequest): Promise<Outcome | 'wait'> {
const daemon = await this.deps.open()
try {
- const move = async (harnessId: string): Promise<{ error?: string; fresh: boolean }> => {
+ const move = async (harnessId: string): Promise<{ error?: string; detail?: string; fresh: boolean }> => {
const fresh = await this.tooLong(harnessId, request)
const frame = request.gridModel === null ? { agentId: harnessId, clearGrid: true }
: { agentId: harnessId, gridModel: request.gridModel, gridName: request.gridName, ...(fresh ? { fresh: true } : {}) }
- const reply = await daemon.request('agent_retarget', frame, MOVE_TIMEOUT_MS).catch(() => ({ error: 'TIMEOUT' }))
- return { ...(typeof reply.error === 'string' ? { error: reply.error } : {}), fresh }
+ const reply = await daemon.request('agent_retarget', frame, MOVE_TIMEOUT_MS).catch((): Record<string, unknown> => ({ error: 'TIMEOUT' }))
+ return { ...(typeof reply.error === 'string' ? { error: reply.error } : {}),
+ ...(typeof reply.detail === 'string' ? { detail: reply.detail } : {}), fresh }
}
const moved = await move(row.agent.harnessId)
if (moved.error) {
// Sent home by the Swarm (a limit, a model gone) while it answers: moved once the turn ends, within the request's lease.
if (moved.error === 'AGENT_BUSY' && request.reason) return 'wait'
- return { ok: false, error: refusal(moved.error, request) }
+ return { ok: false, error: refusal(moved.error, request, moved.detail) }
}
const stayed: string[] = []
for (const session of sessions) {
const result = await move(session.agent.harnessId)
- if (result.error) stayed.push(result.error === 'AGENT_BUSY' ? 'it was answering' : refusal(result.error, request))
+ if (result.error) stayed.push(result.error === 'AGENT_BUSY' ? 'it was answering' : refusal(result.error, request, result.detail))
}
return { ok: true, ...(moved.fresh ? { fresh: true as const } : {}),
...(stayed.length === 1 ? { error: `A thread session stayed where it was: ${stayed[0]}` }
diff --git a/docs/plans/2026-11-05-buddy-switch-start-again-plan.md b/docs/plans/2026-11-05-buddy-switch-start-again-plan.md
new file mode 100644
index 000000000..c8cee3e9b
--- /dev/null
+++ b/docs/plans/2026-11-05-buddy-switch-start-again-plan.md
@@ -0,0 +1,43 @@
+# A Buddy model switch that can't run says why and offers Start again — Plan
+
+**Goal:** When a Buddy's model switch is refused by the Buddy guard, its owner reads why in plain words ("Start your Buddy
+again to switch its model") instead of `its computer could not switch it (SWARM_SKILL_FAILED)`, and Desktop offers a
+one-click Start again.
+**Architecture:** The switch relaunches the Buddy through the daemon's `agent_retarget`; the relaunch's launch overrides
+(cli/src/lib/launchOverrides.ts) run `prepareBuddyFolder` (cli/src/lib/swarmSkills.ts), whose guard throws one of two fixed
+sentences, and return `SWARM_SKILL_FAILED` with that sentence as the detail. `refusal()` in cli/src/superTabs/buddyModel.ts
+reads only the code. It now also recognises the guard's two sentences in the detail (the constants exported from
+swarmSkills.ts, unchanged) and maps each to its own sentence; the detail itself is still never forwarded. The launch error
+code stays `SWARM_SKILL_FAILED`, so the launch telemetry families and every other relaunch caller keep their contract.
+**Tech Stack:** TypeScript CLI, vitest; Flutter Desktop, flutter_test.
+**Spec:** Task #3256 (Scholes, 11-05; Guide2's evidence from task #3069). This plan is implemented by its Engineer, test first,
+with no execution handoff.
+
+## Global Constraints
+- The Buddy guard stays as strict: no condition in `prepareBuddyFolder` changes, only which words reach the owner.
+- No path or daemon detail reaches the owner: refusals are fixed sentences.
+- Swarm code stays out of cli/src/core. Two security reviews (Buddy guard).
+- Part 2 (Desktop) is built only after Scholes approves the mockup posted in the task.
+
+## Review Focus
+- The guard's conditions are unchanged (only `export` added to its two sentences).
+- A detail that merely mentions a sentence cannot make any other failure look like success: it only changes refusal words.
+
+## Tasks
+
+### 1. Plain refusal words for the guard's two holds
+**Files:** cli/src/superTabs/buddyModel.ts, cli/src/superTabs/buddyModel.spec.ts, cli/src/lib/swarmSkills.ts (export only)
+**Interfaces:** `export const BUDDY_START_AGAIN`, `export const BUDDY_NOT_HERE` (swarmSkills.ts).
+- [ ] Failing test: an `agent_retarget` answer `{ error: 'SWARM_SKILL_FAILED', detail: String(new Error(BUDDY_START_AGAIN)) }`
+ reports `Start your Buddy again to switch its model`; with `BUDDY_NOT_HERE` it reports
+ `Your Buddy runs only on your own computer, signed in as you`; any other `SWARM_SKILL_FAILED` keeps the code sentence.
+- [ ] Watch it fail, commit.
+- [ ] `refusal()` reads the detail of a `SWARM_SKILL_FAILED` answer for the two sentences. Test passes, commit.
+
+### 2. Desktop offers Start again before a switch (after Scholes's yes)
+**Files:** desktop/lib/super_tabs/buddies.dart, desktop/test/
+- [ ] Failing widget test: a Swarm Buddy bar whose Buddy's owner isn't recorded shows "Start your Buddy again to switch its
+ model." with a Start again button that calls `addBuddy` for its machine and engine, and no model picker.
+- [ ] Watch it fail, commit; build the row; test passes, commit.
+- The signal Desktop reads for "owner not recorded" is decided with the PM in the task (it needs the Buddy's machine to
+ report it, outside this task's files today).