typeclaw 0.42.0 → 0.43.0
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/auth.schema.json +16 -0
- package/package.json +3 -3
- package/secrets.schema.json +16 -0
- package/src/agent/adaptive-thinking-compat.ts +79 -0
- package/src/agent/index.ts +79 -101
- package/src/agent/live-subagents.ts +4 -4
- package/src/agent/plugin-tools.ts +85 -132
- package/src/agent/provider-error.ts +37 -0
- package/src/agent/reload-tool.ts +1 -1
- package/src/agent/session-origin.ts +6 -3
- package/src/agent/subagents.ts +2 -2
- package/src/agent/system-prompt.ts +24 -8
- package/src/agent/tool-result-budget.ts +1 -1
- package/src/agent/tools/channel-react.ts +5 -4
- package/src/agent/tools/spawn-subagent.ts +80 -24
- package/src/agent/tools/stream-snapshot.ts +8 -3
- package/src/bundled-plugins/agent-browser/index.ts +18 -8
- package/src/bundled-plugins/memory/README.md +4 -1
- package/src/bundled-plugins/memory/belief-sentence.ts +50 -0
- package/src/bundled-plugins/memory/dreaming.ts +13 -6
- package/src/bundled-plugins/memory/load-memory.ts +22 -2
- package/src/bundled-plugins/memory/memory-logger.ts +1 -1
- package/src/bundled-plugins/planner/planner.ts +1 -1
- package/src/bundled-plugins/researcher/researcher.ts +2 -2
- package/src/bundled-plugins/reviewer/reviewer.ts +1 -1
- package/src/capabilities/control-plane.ts +24 -0
- package/src/capabilities/index.ts +2 -0
- package/src/capabilities/runtime.ts +30 -0
- package/src/channels/adapters/discord-bot-thread-room.ts +73 -0
- package/src/channels/adapters/discord-bot.ts +9 -1
- package/src/channels/adapters/slack-bot-classify.ts +1 -0
- package/src/channels/adapters/slack-classify.ts +1 -0
- package/src/channels/engagement.ts +33 -0
- package/src/channels/manager.ts +41 -47
- package/src/channels/router.ts +150 -21
- package/src/channels/types.ts +10 -0
- package/src/cli/channel.ts +49 -12
- package/src/cli/command-meta.ts +2 -0
- package/src/cli/hostd.ts +4 -3
- package/src/cli/index.ts +1 -0
- package/src/cli/logs.ts +2 -2
- package/src/cli/mcp.ts +233 -0
- package/src/cli/restart.ts +5 -3
- package/src/cli/shell.ts +2 -2
- package/src/cli/start.ts +2 -2
- package/src/cli/status.ts +2 -2
- package/src/cli/stop.ts +2 -2
- package/src/compose/restart.ts +4 -3
- package/src/compose/start.ts +2 -2
- package/src/compose/stop.ts +2 -2
- package/src/config/config.ts +10 -1
- package/src/config/providers.ts +61 -9
- package/src/container/controller.ts +104 -0
- package/src/container/index.ts +9 -0
- package/src/hostd/daemon.ts +28 -9
- package/src/hostd/protocol.ts +12 -9
- package/src/init/index.ts +1 -1
- package/src/init/instagram-auth.ts +189 -34
- package/src/inspect/logs-item.ts +2 -2
- package/src/mcp/client.ts +16 -6
- package/src/mcp/index.ts +12 -0
- package/src/mcp/manager.ts +9 -4
- package/src/mcp/oauth.ts +212 -0
- package/src/run/index.ts +36 -11
- package/src/run/{codex-fetch-observer.ts → llm-fetch-observer.ts} +182 -82
- package/src/sandbox/availability.ts +2 -2
- package/src/sandbox/build.ts +1 -1
- package/src/sandbox/hidden-paths.ts +59 -0
- package/src/sandbox/index.ts +1 -1
- package/src/sandbox/policy.ts +2 -2
- package/src/secrets/discord-store.ts +4 -13
- package/src/secrets/index.ts +1 -1
- package/src/secrets/instagram-store.ts +4 -13
- package/src/secrets/kakao-store.ts +4 -13
- package/src/secrets/line-store.ts +4 -13
- package/src/secrets/schema.ts +13 -0
- package/src/secrets/secrets-provider.ts +86 -0
- package/src/secrets/slack-store.ts +4 -13
- package/src/secrets/storage.ts +108 -1
- package/src/secrets/webex-store.ts +4 -13
- package/src/server/index.ts +2 -2
- package/src/skills/typeclaw-channel-github/SKILL.md +2 -2
- package/src/skills/typeclaw-channel-instagram/SKILL.md +1 -1
- package/src/skills/typeclaw-config/SKILL.md +5 -3
- package/src/skills/typeclaw-plugins/SKILL.md +1 -1
- package/src/skills/typeclaw-troubleshooting/SKILL.md +1 -1
- package/src/stream/broker.ts +4 -3
- package/src/stream/types.ts +1 -0
- package/typeclaw.schema.json +14 -0
|
@@ -1,20 +1,19 @@
|
|
|
1
1
|
import { AsyncLocalStorage } from 'node:async_hooks'
|
|
2
2
|
import { join } from 'node:path'
|
|
3
3
|
|
|
4
|
-
import type { AgentTool } from '@mariozechner/pi-agent-core'
|
|
5
4
|
import {
|
|
6
|
-
|
|
5
|
+
createBashToolDefinition as piCreateBashToolDefinition,
|
|
6
|
+
createEditToolDefinition as piCreateEditToolDefinition,
|
|
7
|
+
createFindToolDefinition as piCreateFindToolDefinition,
|
|
8
|
+
createGrepToolDefinition as piCreateGrepToolDefinition,
|
|
9
|
+
createLsToolDefinition as piCreateLsToolDefinition,
|
|
10
|
+
createReadToolDefinition as piCreateReadToolDefinition,
|
|
11
|
+
createWriteToolDefinition as piCreateWriteToolDefinition,
|
|
7
12
|
defineTool as piDefineTool,
|
|
8
|
-
editTool as piEditTool,
|
|
9
|
-
findTool as piFindTool,
|
|
10
|
-
grepTool as piGrepTool,
|
|
11
|
-
lsTool as piLsTool,
|
|
12
|
-
readTool as piReadTool,
|
|
13
|
-
writeTool as piWriteTool,
|
|
14
13
|
} from '@mariozechner/pi-coding-agent'
|
|
15
14
|
import type { BashSpawnContext, ToolDefinition } from '@mariozechner/pi-coding-agent'
|
|
16
|
-
import type { Static, TSchema } from '
|
|
17
|
-
import { Type } from '
|
|
15
|
+
import type { Static, TSchema } from 'typebox'
|
|
16
|
+
import { Type } from 'typebox'
|
|
18
17
|
import { z } from 'zod'
|
|
19
18
|
|
|
20
19
|
import {
|
|
@@ -41,6 +40,7 @@ import {
|
|
|
41
40
|
commandNeedsRealProc,
|
|
42
41
|
DEFAULT_SANDBOX_ENV,
|
|
43
42
|
ensureBwrapAvailable,
|
|
43
|
+
ensureHiddenMaskTargets,
|
|
44
44
|
ensureSessionTmpDir,
|
|
45
45
|
getProcBindSafetyVerdict,
|
|
46
46
|
isPackageInstallCommand,
|
|
@@ -102,8 +102,6 @@ function bashSpawnHookWithOverlay(context: BashSpawnContext): BashSpawnContext {
|
|
|
102
102
|
return { ...context, env: { ...context.env, ...overlay } }
|
|
103
103
|
}
|
|
104
104
|
|
|
105
|
-
const piBashTool = piCreateBashTool(process.cwd(), { spawnHook: bashSpawnHookWithOverlay })
|
|
106
|
-
|
|
107
105
|
const ACKNOWLEDGE_GUARDS_SCHEMA = Type.Optional(
|
|
108
106
|
Type.Object(
|
|
109
107
|
{
|
|
@@ -115,31 +113,50 @@ const ACKNOWLEDGE_GUARDS_SCHEMA = Type.Optional(
|
|
|
115
113
|
),
|
|
116
114
|
)
|
|
117
115
|
|
|
118
|
-
// pi-coding-agent 0.
|
|
119
|
-
// - `createAgentSession({ tools:
|
|
120
|
-
//
|
|
121
|
-
//
|
|
122
|
-
//
|
|
116
|
+
// pi-coding-agent 0.73 contract (load-bearing for hook coverage):
|
|
117
|
+
// - `createAgentSession({ tools: string[] })` is a name allowlist: only the
|
|
118
|
+
// listed names stay active, and that allowlist gates BOTH builtins and
|
|
119
|
+
// custom tools (see `allowedToolNames` in pi's `_refreshToolRegistry`).
|
|
120
|
+
// - `noTools: "builtin"` drops pi's read/bash/edit/write from the INITIAL
|
|
121
|
+
// active set; pi still registers its base builtins, so a same-named entry
|
|
122
|
+
// in `customTools` is what overrides the implementation (registry
|
|
123
|
+
// last-write-wins). Passing an explicit `tools:` allowlist plus shipping all
|
|
124
|
+
// seven wrapped builtins is what makes the wrapped versions the only
|
|
125
|
+
// callable ones.
|
|
123
126
|
//
|
|
124
|
-
// Consequence:
|
|
125
|
-
//
|
|
126
|
-
//
|
|
127
|
-
//
|
|
128
|
-
//
|
|
129
|
-
|
|
130
|
-
// name-filter path); the wrapped customTools just replace the implementation
|
|
131
|
-
// underneath so subagent and channel sessions share the same hook coverage.
|
|
132
|
-
type PiAgentToolName = 'read' | 'bash' | 'edit' | 'write' | 'grep' | 'find' | 'ls'
|
|
127
|
+
// Consequence: every builtin enters as a `ToolDefinition` (pi now exposes
|
|
128
|
+
// `create*ToolDefinition` factories), TypeClaw wraps each one with its hook +
|
|
129
|
+
// guard + sandbox pipeline, and the call site routes them through `customTools`
|
|
130
|
+
// while narrowing via `tools:` names. There is no longer an `AgentTool` vs
|
|
131
|
+
// `ToolDefinition` split.
|
|
132
|
+
type PiBuiltinToolName = 'read' | 'bash' | 'edit' | 'write' | 'grep' | 'find' | 'ls'
|
|
133
133
|
type TypeclawToolName = 'web_search' | 'web_fetch'
|
|
134
134
|
|
|
135
|
-
const
|
|
136
|
-
|
|
137
|
-
|
|
138
|
-
|
|
139
|
-
|
|
140
|
-
|
|
141
|
-
|
|
142
|
-
|
|
135
|
+
const PI_BUILTIN_TOOL_NAMES: readonly PiBuiltinToolName[] = ['read', 'bash', 'edit', 'write', 'grep', 'find', 'ls']
|
|
136
|
+
|
|
137
|
+
// pi builtins resolve relative paths (and, for trusted/owner bash, the spawn
|
|
138
|
+
// cwd) against the cwd baked in at factory time, so the definitions are built
|
|
139
|
+
// per session from the session's agentDir rather than the module-load
|
|
140
|
+
// process.cwd() — otherwise a session whose agentDir differs from the import-
|
|
141
|
+
// time cwd would read/write in the wrong tree. bash keeps the spawnHook that
|
|
142
|
+
// threads the internal env overlay to the (non-sandboxed) spawn.
|
|
143
|
+
function createPiBuiltinToolDefinition(name: PiBuiltinToolName, cwd: string): ToolDefinition<any, any, any> {
|
|
144
|
+
switch (name) {
|
|
145
|
+
case 'read':
|
|
146
|
+
return piCreateReadToolDefinition(cwd)
|
|
147
|
+
case 'bash':
|
|
148
|
+
return piCreateBashToolDefinition(cwd, { spawnHook: bashSpawnHookWithOverlay })
|
|
149
|
+
case 'edit':
|
|
150
|
+
return piCreateEditToolDefinition(cwd)
|
|
151
|
+
case 'write':
|
|
152
|
+
return piCreateWriteToolDefinition(cwd)
|
|
153
|
+
case 'grep':
|
|
154
|
+
return piCreateGrepToolDefinition(cwd)
|
|
155
|
+
case 'find':
|
|
156
|
+
return piCreateFindToolDefinition(cwd)
|
|
157
|
+
case 'ls':
|
|
158
|
+
return piCreateLsToolDefinition(cwd)
|
|
159
|
+
}
|
|
143
160
|
}
|
|
144
161
|
|
|
145
162
|
const TYPECLAW_TOOL_DEFINITION_MAP: Record<TypeclawToolName, ToolDefinition<any, any, any>> = {
|
|
@@ -147,33 +164,25 @@ const TYPECLAW_TOOL_DEFINITION_MAP: Record<TypeclawToolName, ToolDefinition<any,
|
|
|
147
164
|
web_fetch: webFetchTool,
|
|
148
165
|
}
|
|
149
166
|
|
|
150
|
-
function
|
|
151
|
-
return
|
|
167
|
+
function isPiBuiltinToolName(name: string): name is PiBuiltinToolName {
|
|
168
|
+
return (PI_BUILTIN_TOOL_NAMES as readonly string[]).includes(name)
|
|
152
169
|
}
|
|
153
170
|
|
|
154
|
-
function
|
|
155
|
-
return name
|
|
171
|
+
export function isPiCodingBuiltinName(name: string): boolean {
|
|
172
|
+
return isPiBuiltinToolName(name)
|
|
156
173
|
}
|
|
157
174
|
|
|
158
|
-
|
|
159
|
-
|
|
160
|
-
toolDefinitions: ToolDefinition<any, any, any>[]
|
|
175
|
+
function isTypeclawToolName(name: string): name is TypeclawToolName {
|
|
176
|
+
return name in TYPECLAW_TOOL_DEFINITION_MAP
|
|
161
177
|
}
|
|
162
178
|
|
|
163
|
-
export function resolveBuiltinToolRefs(refs: BuiltinToolRef[]):
|
|
164
|
-
|
|
165
|
-
const toolDefinitions: ToolDefinition<any, any, any>[] = []
|
|
166
|
-
for (const ref of refs) {
|
|
179
|
+
export function resolveBuiltinToolRefs(refs: BuiltinToolRef[], cwd: string): ToolDefinition<any, any, any>[] {
|
|
180
|
+
return refs.map((ref) => {
|
|
167
181
|
const name = ref.__builtinTool
|
|
168
|
-
if (
|
|
169
|
-
|
|
170
|
-
|
|
171
|
-
|
|
172
|
-
} else {
|
|
173
|
-
throw new Error(`unknown built-in tool ref: ${name}`)
|
|
174
|
-
}
|
|
175
|
-
}
|
|
176
|
-
return { agentTools, toolDefinitions }
|
|
182
|
+
if (isPiBuiltinToolName(name)) return createPiBuiltinToolDefinition(name, cwd)
|
|
183
|
+
if (isTypeclawToolName(name)) return TYPECLAW_TOOL_DEFINITION_MAP[name]
|
|
184
|
+
throw new Error(`unknown built-in tool ref: ${name}`)
|
|
185
|
+
})
|
|
177
186
|
}
|
|
178
187
|
|
|
179
188
|
export type WrapToolOptions = {
|
|
@@ -363,81 +372,20 @@ export function wrapSystemTool<TParams extends TSchema, TDetails = unknown, TSta
|
|
|
363
372
|
})
|
|
364
373
|
}
|
|
365
374
|
|
|
366
|
-
|
|
367
|
-
|
|
368
|
-
|
|
369
|
-
|
|
370
|
-
|
|
371
|
-
|
|
372
|
-
|
|
373
|
-
|
|
374
|
-
const mutableArgs = params as Record<string, unknown>
|
|
375
|
-
const liveOrigin = opts.getOrigin?.()
|
|
376
|
-
const blockResult = await opts.hooks.runToolBefore({
|
|
377
|
-
tool: tool.name,
|
|
378
|
-
sessionId: opts.sessionId,
|
|
379
|
-
callId: toolCallId,
|
|
380
|
-
args: mutableArgs,
|
|
381
|
-
...(liveOrigin !== undefined ? { origin: liveOrigin } : {}),
|
|
382
|
-
})
|
|
383
|
-
if (blockResult !== undefined) {
|
|
384
|
-
throw new Error(`blocked: ${blockResult.reason}`)
|
|
385
|
-
}
|
|
386
|
-
const loopGate = gateLoopGuard(opts.sessionId, tool.name, mutableArgs)
|
|
387
|
-
if (loopGate.blockNow) {
|
|
388
|
-
fireLoopAbort(opts.getAbort, 'loop_guard:block')
|
|
389
|
-
throw new Error(loopGate.message)
|
|
390
|
-
}
|
|
391
|
-
const guardResult = await runFinalWriteGuards({
|
|
392
|
-
tool: tool.name,
|
|
393
|
-
args: mutableArgs,
|
|
394
|
-
agentDir: opts.agentDir,
|
|
395
|
-
})
|
|
396
|
-
if (guardResult !== undefined) {
|
|
397
|
-
throw new Error(`blocked: ${guardResult.reason}`)
|
|
398
|
-
}
|
|
399
|
-
const readGuardResult = runFinalReadGuards({ tool: tool.name, args: mutableArgs })
|
|
400
|
-
if (readGuardResult !== undefined) {
|
|
401
|
-
throw new Error(`blocked: ${readGuardResult.reason}`)
|
|
402
|
-
}
|
|
403
|
-
stripGuardAcknowledgements(mutableArgs)
|
|
404
|
-
|
|
405
|
-
const result = await tool.execute(toolCallId, mutableArgs as Static<TParams>, signal, onUpdate)
|
|
406
|
-
const resolved = loopGate.resolve({ content: result.content as ContentPart[], details: result.details })
|
|
407
|
-
if ('deferredBlock' in resolved) {
|
|
408
|
-
fireLoopAbort(opts.getAbort, 'loop_guard:deferred_block')
|
|
409
|
-
throw new Error(resolved.deferredBlock)
|
|
410
|
-
}
|
|
411
|
-
const hookResult = resolved.result
|
|
412
|
-
await opts.hooks.runToolAfter({
|
|
413
|
-
tool: tool.name,
|
|
414
|
-
sessionId: opts.sessionId,
|
|
415
|
-
callId: toolCallId,
|
|
416
|
-
result: hookResult,
|
|
417
|
-
})
|
|
418
|
-
return {
|
|
419
|
-
content: hookResult.content as ContentPart[],
|
|
420
|
-
details: hookResult.details as TDetails,
|
|
421
|
-
}
|
|
422
|
-
},
|
|
423
|
-
}
|
|
424
|
-
}
|
|
425
|
-
|
|
426
|
-
// Wraps a pi-coding-agent AgentTool into a ToolDefinition so it can ride in
|
|
427
|
-
// `customTools` and override pi's same-named builtin (see top-of-file contract
|
|
428
|
-
// block). The hook + guard pipeline matches `wrapSystemAgentTool`; only the
|
|
429
|
-
// input/output shape differs.
|
|
430
|
-
export function wrapAgentToolAsCustomToolDefinition<TParams extends TSchema, TDetails = unknown>(
|
|
431
|
-
tool: AgentTool<TParams, TDetails>,
|
|
375
|
+
// Wraps a pi builtin `ToolDefinition` (read/bash/edit/write/grep/find/ls) with
|
|
376
|
+
// TypeClaw's full pipeline — hooks, loop guard, write/read guards, per-subagent
|
|
377
|
+
// bash policy, bwrap sandbox, /tmp redirect, and the internal bash env overlay —
|
|
378
|
+
// then ships it through `customTools`. `noTools: "builtin"` at the call site
|
|
379
|
+
// disables pi's own unwrapped copies, so this wrapped definition is the ONLY
|
|
380
|
+
// registered implementation of each builtin name.
|
|
381
|
+
export function wrapBuiltinToolDefinition<TParams extends TSchema, TDetails = unknown, TState = any>(
|
|
382
|
+
tool: ToolDefinition<TParams, TDetails, TState>,
|
|
432
383
|
opts: WrapSystemToolOptions,
|
|
433
|
-
): ToolDefinition<TParams, TDetails> {
|
|
384
|
+
): ToolDefinition<TParams, TDetails, TState> {
|
|
434
385
|
return piDefineTool({
|
|
435
|
-
|
|
436
|
-
label: tool.label,
|
|
437
|
-
description: tool.description,
|
|
386
|
+
...tool,
|
|
438
387
|
parameters: withGuardAcknowledgements(tool.name, tool.parameters),
|
|
439
|
-
|
|
440
|
-
async execute(toolCallId, params, signal, onUpdate) {
|
|
388
|
+
async execute(toolCallId, params, signal, onUpdate, ctx) {
|
|
441
389
|
const mutableArgs = params as Record<string, unknown>
|
|
442
390
|
const liveOrigin = opts.getOrigin?.()
|
|
443
391
|
// Defense-in-depth: strip any pre-existing internal env-overlay key
|
|
@@ -499,7 +447,7 @@ export function wrapAgentToolAsCustomToolDefinition<TParams extends TSchema, TDe
|
|
|
499
447
|
let rawResult: ToolResult
|
|
500
448
|
try {
|
|
501
449
|
rawResult = await bashEnvStore.run(bashEnvOverlay, () =>
|
|
502
|
-
tool.execute(toolCallId, mutableArgs as Static<TParams>, signal, onUpdate),
|
|
450
|
+
tool.execute(toolCallId, mutableArgs as Static<TParams>, signal, onUpdate, ctx),
|
|
503
451
|
)
|
|
504
452
|
} catch (error) {
|
|
505
453
|
// A throwing tool (pi's bash rejects on non-zero exit) must still run
|
|
@@ -549,12 +497,12 @@ async function runToolAfterSafely(
|
|
|
549
497
|
}
|
|
550
498
|
}
|
|
551
499
|
|
|
552
|
-
export function
|
|
553
|
-
return
|
|
500
|
+
export function defaultBuiltinPiToolDefinitions(cwd: string): ToolDefinition<any, any, any>[] {
|
|
501
|
+
return PI_BUILTIN_TOOL_NAMES.map((name) => createPiBuiltinToolDefinition(name, cwd))
|
|
554
502
|
}
|
|
555
503
|
|
|
556
504
|
export function buildBuiltinPiToolOverrides(opts: WrapSystemToolOptions): ToolDefinition<any, any>[] {
|
|
557
|
-
return
|
|
505
|
+
return defaultBuiltinPiToolDefinitions(opts.agentDir).map((tool) => wrapBuiltinToolDefinition(tool, opts))
|
|
558
506
|
}
|
|
559
507
|
|
|
560
508
|
// Rewrites mutableArgs.command in place so the bash builtin runs inside bwrap
|
|
@@ -579,6 +527,11 @@ async function applyBashSandbox(
|
|
|
579
527
|
const sandboxEnvOverlay = buildRoleScopedConfigEnv(agentDir, dirs, envOverlay)
|
|
580
528
|
|
|
581
529
|
await ensureBwrapAvailable()
|
|
530
|
+
// bwrap's --ro-bind-data/--tmpfs mask ops abort when the target does not exist
|
|
531
|
+
// on the (read-only, virtiofs/OrbStack) agent-folder bind. Materialize the mask
|
|
532
|
+
// targets on the real host FS first; the full {dirs, files} still feeds
|
|
533
|
+
// subtractMasked below so a masked path is never re-exposed by a later RW bind.
|
|
534
|
+
const maskTargets = await ensureHiddenMaskTargets({ dirs, files })
|
|
582
535
|
// Per-session /tmp: bind this session's scratch dir over the default
|
|
583
536
|
// --tmpfs /tmp so writes survive across the role's sandboxed bash calls AND
|
|
584
537
|
// match what the write/edit wrapper redirected a /tmp path to. The bind is
|
|
@@ -666,8 +619,8 @@ async function applyBashSandbox(
|
|
|
666
619
|
{ type: 'bind', source: sessionTmp, dest: '/tmp' },
|
|
667
620
|
],
|
|
668
621
|
...(packageInstall !== undefined
|
|
669
|
-
? { writableRoot: { dir: packageInstall.root }, masks:
|
|
670
|
-
: { masks:
|
|
622
|
+
? { writableRoot: { dir: packageInstall.root }, masks: maskTargets, protected: packageInstall.protected }
|
|
623
|
+
: { masks: maskTargets, writable, protected: protectedZones }),
|
|
671
624
|
symlinks,
|
|
672
625
|
network: 'inherit',
|
|
673
626
|
cwd: agentDir,
|
|
@@ -26,6 +26,13 @@ export type DetectedProviderError = {
|
|
|
26
26
|
|
|
27
27
|
const GENERIC_SAFE_NOTICE = 'The upstream LLM provider failed. Operators can check `typeclaw logs` for details.'
|
|
28
28
|
|
|
29
|
+
// The fetch observer (`llm-fetch-observer`) aborts a stalled provider stream
|
|
30
|
+
// with a message ending in this marker (TTFB / idle / overall deadline). Such a
|
|
31
|
+
// stall is a hard throw, not a `stopReason: 'error'` soft error, so it never
|
|
32
|
+
// reaches `detectProviderError`; the marker lets the hard-throw path recognize
|
|
33
|
+
// and surface it. The literal is the system's own token — English by design.
|
|
34
|
+
const OBSERVER_TIMEOUT = /\(typeclaw observer timeout\)/i
|
|
35
|
+
|
|
29
36
|
// Each entry pairs a narrow matcher against the raw provider text with the
|
|
30
37
|
// canonical, leak-free sentence shown in channels. Matchers are intentionally
|
|
31
38
|
// specific: a miss falls through to GENERIC_SAFE_NOTICE rather than echoing raw
|
|
@@ -60,6 +67,10 @@ const SAFE_CLASSES: ReadonlyArray<{ match: RegExp; safe: string }> = [
|
|
|
60
67
|
match: /\b(billing|quota|insufficient.*(credit|fund|balance)|payment|account is not active)\b/i,
|
|
61
68
|
safe: 'The upstream LLM provider rejected the request for a billing/quota reason. Operators can check `typeclaw logs` for details.',
|
|
62
69
|
},
|
|
70
|
+
{
|
|
71
|
+
match: OBSERVER_TIMEOUT,
|
|
72
|
+
safe: 'The upstream LLM provider stopped responding and the request timed out. Try again shortly.',
|
|
73
|
+
},
|
|
63
74
|
]
|
|
64
75
|
|
|
65
76
|
function toSafeMessage(raw: string): string {
|
|
@@ -90,6 +101,17 @@ export function isThrottleOrOverload(raw: string): boolean {
|
|
|
90
101
|
return THROTTLE_OR_OVERLOAD.test(raw)
|
|
91
102
|
}
|
|
92
103
|
|
|
104
|
+
// Failover predicate for turn-drivers' `shouldFailover`: rotate to the next model
|
|
105
|
+
// ref on a transient capacity signal (throttle/overload) OR an observer stall
|
|
106
|
+
// timeout. A stall — the provider accepted the stream but stopped sending bytes —
|
|
107
|
+
// is as transient as an overload and another ref may well succeed, so it must
|
|
108
|
+
// fail OVER before the caller surfaces a failure notice. Account-wide faults
|
|
109
|
+
// (billing/quota/auth) still return false via `isThrottleOrOverload` — a
|
|
110
|
+
// different ref shares the same account problem.
|
|
111
|
+
export function isFailoverWorthy(raw: string): boolean {
|
|
112
|
+
return isThrottleOrOverload(raw) || OBSERVER_TIMEOUT.test(raw)
|
|
113
|
+
}
|
|
114
|
+
|
|
93
115
|
export function detectProviderError(message: unknown): DetectedProviderError | null {
|
|
94
116
|
if (typeof message !== 'object' || message === null) return null
|
|
95
117
|
const m = message as { role?: unknown; stopReason?: unknown; errorMessage?: unknown }
|
|
@@ -102,6 +124,21 @@ export function detectProviderError(message: unknown): DetectedProviderError | n
|
|
|
102
124
|
return { message: text, safeMessage: toSafeMessage(text) }
|
|
103
125
|
}
|
|
104
126
|
|
|
127
|
+
// Classifies a HARD-thrown error (an exception out of `session.prompt()`, after
|
|
128
|
+
// model fallback is exhausted) into a user-facing notice — the throw counterpart
|
|
129
|
+
// to `detectProviderError`'s soft-error path. Returns `null` for errors that are
|
|
130
|
+
// NOT an operator-actionable provider failure (internal bugs, network blips),
|
|
131
|
+
// so the caller stays silent-with-log rather than spamming channels. The three
|
|
132
|
+
// recognized classes are: failover-worthy throttle/overload, account-wide faults
|
|
133
|
+
// (billing/quota/auth) that must surface, and observer stall timeouts.
|
|
134
|
+
export function detectHardProviderError(err: unknown): DetectedProviderError | null {
|
|
135
|
+
const message = err instanceof Error ? err.message : String(err)
|
|
136
|
+
const isProviderFailure =
|
|
137
|
+
isThrottleOrOverload(message) || NON_FAILOVER_FAULT.test(message) || OBSERVER_TIMEOUT.test(message)
|
|
138
|
+
if (!isProviderFailure) return null
|
|
139
|
+
return { message, safeMessage: toSafeMessage(message) }
|
|
140
|
+
}
|
|
141
|
+
|
|
105
142
|
export type ProviderErrorListener = (error: DetectedProviderError) => void
|
|
106
143
|
export type Unsubscribe = () => void
|
|
107
144
|
|
package/src/agent/reload-tool.ts
CHANGED
|
@@ -429,8 +429,11 @@ function renderChannelOrigin(
|
|
|
429
429
|
'**React like a teammate would.** `channel_react({ emoji })` adds only a',
|
|
430
430
|
'reaction: `+1` approve, `rocket` shipping/exciting, `tada` celebrate,',
|
|
431
431
|
'`heart` appreciate, `laugh` funny, `eyes` looking. Use it when it adds',
|
|
432
|
-
'real signal, not every turn.
|
|
433
|
-
'
|
|
432
|
+
'real signal, not every turn. The reaction is applied ONLY if you also',
|
|
433
|
+
'reply to this message this turn; if you stay silent or `skip_response`,',
|
|
434
|
+
'it is dropped — so never react to a message you are only observing and',
|
|
435
|
+
'not answering. It does NOT satisfy the reply obligation; substantive',
|
|
436
|
+
'answers still go through `channel_reply`.',
|
|
434
437
|
)
|
|
435
438
|
}
|
|
436
439
|
|
|
@@ -492,7 +495,7 @@ function renderChannelOrigin(
|
|
|
492
495
|
]),
|
|
493
496
|
'',
|
|
494
497
|
'**Backgrounded work does not end the obligation.** If you spawn a',
|
|
495
|
-
'subagent
|
|
498
|
+
'subagent in the background to answer the current inbound,',
|
|
496
499
|
'you promised a reply you have not delivered. Do not skip: the system will',
|
|
497
500
|
'not surface the result. When the subagent-completion `<system-reminder>` arrives,',
|
|
498
501
|
'call `subagent_output` and send the result via `channel_reply`.',
|
package/src/agent/subagents.ts
CHANGED
|
@@ -10,7 +10,7 @@ import { applyTurnThinkingLevel } from './attention-escalation'
|
|
|
10
10
|
import { type AgentSession, createSession, type PluginSessionWiring } from './index'
|
|
11
11
|
import { resolveFallbackChain } from './model-fallback'
|
|
12
12
|
import { applyModelRuntimeOverrides } from './model-overrides'
|
|
13
|
-
import {
|
|
13
|
+
import { isFailoverWorthy, subscribeProviderErrors } from './provider-error'
|
|
14
14
|
import type { SubagentBashPolicy } from './reviewer-bash-policy'
|
|
15
15
|
import type { SessionOrigin } from './session-origin'
|
|
16
16
|
import {
|
|
@@ -349,7 +349,7 @@ export async function invokeSubagent(name: string, options: InvokeSubagentOption
|
|
|
349
349
|
text: turnText,
|
|
350
350
|
skipProviderErrorSubscription: true,
|
|
351
351
|
detectSoftErrorFromLeaf: true,
|
|
352
|
-
shouldFailover: (err) =>
|
|
352
|
+
shouldFailover: (err) => isFailoverWorthy(err.message),
|
|
353
353
|
setModelForRef: async (ref) => {
|
|
354
354
|
await session.setModel(applyModelRuntimeOverrides(resolveModel(ref), ref))
|
|
355
355
|
activeModelRef = ref
|
|
@@ -11,10 +11,15 @@ const PACKAGE_JSON_INSTALL_RULE =
|
|
|
11
11
|
// by the no-registry path (back-compat callers, the debug dumper); production
|
|
12
12
|
// full-mode sessions pass the real registry-rendered roster via
|
|
13
13
|
// `composeSystemPrompt`'s `subagentRoster` field.
|
|
14
|
-
export function buildDefaultSystemPrompt(subagentRoster: string): string {
|
|
15
|
-
|
|
14
|
+
export function buildDefaultSystemPrompt(subagentRoster: string, branding = true): string {
|
|
15
|
+
const opening = branding
|
|
16
|
+
? `You are a general-purpose AI agent running inside TypeClaw.
|
|
16
17
|
|
|
17
|
-
TypeClaw is domain-agnostic: \`IDENTITY.md\` defines your role, \`SOUL.md\` your voice, and \`AGENTS.md\` your operating manual. This prompt describes only the runtime
|
|
18
|
+
TypeClaw is domain-agnostic: \`IDENTITY.md\` defines your role, \`SOUL.md\` your voice, and \`AGENTS.md\` your operating manual. This prompt describes only the runtime.`
|
|
19
|
+
: `You are a general-purpose AI agent.
|
|
20
|
+
|
|
21
|
+
Your runtime is domain-agnostic: \`IDENTITY.md\` defines your role, \`SOUL.md\` your voice, and \`AGENTS.md\` your operating manual. This prompt describes only the runtime.`
|
|
22
|
+
return `${opening}
|
|
18
23
|
|
|
19
24
|
## Your agent folder
|
|
20
25
|
|
|
@@ -93,7 +98,7 @@ Use tmux only for work that belongs in your session. Delegate self-contained lon
|
|
|
93
98
|
|
|
94
99
|
## Version control
|
|
95
100
|
|
|
96
|
-
Your agent folder is a git repository, but **it is your own private backup repo — not a software project you develop.** TypeClaw snapshots identity files, \`sessions/\`, and \`memory/\` there over time. It normally has no remote, nothing is pushed, and it is **not a checkout of any project**. Commits here save your state, not a codebase contribution.
|
|
101
|
+
Your agent folder is a git repository, but **it is your own private backup repo — not a software project you develop.** ${branding ? 'TypeClaw snapshots' : 'The runtime snapshots'} identity files, \`sessions/\`, and \`memory/\` there over time. It normally has no remote, nothing is pushed, and it is **not a checkout of any project**. Commits here save your state, not a codebase contribution.
|
|
97
102
|
|
|
98
103
|
For project work (bug, feature, PR), clone the project repo into \`/tmp/<repo>\`, work there, and open the PR from that clone with \`gh\`. Never \`git init\`, add a remote, or push your agent folder as the project. If there is no remote or you cannot find the repo, ask the user where it lives. Your agent folder is where you live; the clone is where you work.
|
|
99
104
|
|
|
@@ -116,13 +121,15 @@ Commits to your agent folder (your own state):
|
|
|
116
121
|
|
|
117
122
|
Delegate focused work with \`spawn_subagent\`, \`subagent_output\`, and \`subagent_cancel\`. Each subagent has its own context/tools; re-read the tool description before delegating. Briefly: ${subagentRoster}.
|
|
118
123
|
|
|
124
|
+
Spawns run in the BACKGROUND by default: the tool returns a task_id immediately and a \`<system-reminder>\` arrives when the subagent completes — do NOT poll \`subagent_output\`. Pass \`run_in_foreground: true\` for a quick call whose result you need inline this turn. Deep-profile subagents (deep investigation, review, planning) always run in the background from here — requesting foreground for one is overridden, and the tool result tells you so.
|
|
125
|
+
|
|
119
126
|
Pick one of three modes:
|
|
120
127
|
|
|
121
|
-
**Mode A — Research fan-out.** Broad search: spawn 2-5 \`explorer\`/\`scout\` workers in parallel
|
|
128
|
+
**Mode A — Research fan-out.** Broad search: spawn 2-5 \`explorer\`/\`scout\` workers in parallel (background by default), end your response, then collect each completion once via \`subagent_output\`. Use \`scout\` for narrow lookups; \`researcher\` for decomposed, multi-source, cross-validated synthesis. When the user *explicitly* says "research"/"investigate" (or equivalent), you MUST spawn \`researcher\` — answering from training memory or a single inline \`web_search\` does not satisfy the request, even if you think you know the answer. (Fanning out \`scout\`/\`explorer\` underneath is fine, but it does not replace \`researcher\`.)
|
|
122
129
|
|
|
123
|
-
**Mode B — Delegate-and-converse.** For >~30s side-effectful/noisy work (installs, builds, \`docker\`, scrapes, long tests, multi-host loops, fetch-and-synthesize chains), spawn one background
|
|
130
|
+
**Mode B — Delegate-and-converse.** For >~30s side-effectful/noisy work (installs, builds, \`docker\`, scrapes, long tests, multi-host loops, fetch-and-synthesize chains), spawn one subagent (background) and stay responsive: \`operator\` for side effects, \`scout\` quick lookup, \`researcher\` deep investigation, \`planner\` risk-aware sequencing. For a quick lookup whose answer you need inline, pass \`run_in_foreground: true\`. For hard execution work that needs stronger reasoning than the default tier — gnarly bugs, non-obvious refactors, failures that resisted a quick fix — pass \`profile: "deep"\` to run that one spawn on a stronger model; leave it off for routine work. When the completion \`<system-reminder>\` lands, Surface the result via \`channel_reply\`/\`channel_send\` in channel sessions because reminders are not user messages.
|
|
124
131
|
|
|
125
|
-
**Mode C — Troubleshooting.** After ~3 non-converging attempts at the same failure, stop and hand the loop to \`operator\`
|
|
132
|
+
**Mode C — Troubleshooting.** After ~3 non-converging attempts at the same failure, stop and hand the loop to \`operator\` (background; add \`profile: "deep"\` when the failure is genuinely hard), symptom, attempts, and success condition. Read \`typeclaw-troubleshooting\` before spawning.
|
|
126
133
|
|
|
127
134
|
**Status queries.** If the user asks "status?" on a running subagent, call \`subagent_output({ task_id })\` and report its \`status_summary\` — don't guess.
|
|
128
135
|
|
|
@@ -262,7 +269,9 @@ export function renderTurnRoleAnchor(role: string): string | undefined {
|
|
|
262
269
|
// The full DEFAULT_SYSTEM_PROMPT remains the right choice for TUI + channel
|
|
263
270
|
// sessions because there IS a human reading the output, the agent IS expected
|
|
264
271
|
// to maintain its agent folder over time, and conversational register matters.
|
|
265
|
-
export
|
|
272
|
+
export function buildSlimSystemPrompt(branding = true): string {
|
|
273
|
+
const opening = branding ? 'You are an AI agent running inside TypeClaw.' : 'You are an AI agent.'
|
|
274
|
+
return `${opening}
|
|
266
275
|
|
|
267
276
|
Never echo secrets from \`secrets.json\` or \`.env\`, or any credential you see in the environment. Never include them in tool calls, logs, or commit messages.
|
|
268
277
|
|
|
@@ -277,3 +286,10 @@ Your free-write zone is \`workspace/\`. Do not create files at the root of the a
|
|
|
277
286
|
The agent folder is a private backup repo with no remote, not a project checkout. To work on a software project (fix a bug, open a PR), clone its repo elsewhere (e.g. \`/tmp/<repo>\`) and work there — never push the agent folder as if it were the project.
|
|
278
287
|
|
|
279
288
|
See the session-origin block below for what kind of session this is and what's expected of you.`
|
|
289
|
+
}
|
|
290
|
+
|
|
291
|
+
// Back-compat constant: the slim prompt with branding on. Retained because
|
|
292
|
+
// tests assert `prompt.startsWith(SLIM_SYSTEM_PROMPT)` on the default path;
|
|
293
|
+
// production slim composition substitutes the branding flag via
|
|
294
|
+
// `buildSlimSystemPrompt`.
|
|
295
|
+
export const SLIM_SYSTEM_PROMPT = buildSlimSystemPrompt()
|
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
import type { AgentTool } from '@mariozechner/pi-agent-core'
|
|
2
2
|
import type { ToolDefinition } from '@mariozechner/pi-coding-agent'
|
|
3
|
-
import type { TSchema } from '
|
|
3
|
+
import type { TSchema } from 'typebox'
|
|
4
4
|
|
|
5
5
|
// Subagents that read large files (memory-logger and dreaming each read parent
|
|
6
6
|
// session transcripts that can run hundreds of KB) are vulnerable to a class
|
|
@@ -41,8 +41,9 @@ export function createChannelReactTool({
|
|
|
41
41
|
'Pick the reaction a thoughtful teammate would leave: :+1: to agree or approve, :rocket: for something ' +
|
|
42
42
|
'shipping or exciting, :tada: to celebrate, :heart: to show appreciation, :eyes: to signal "I am looking at this", ' +
|
|
43
43
|
':laugh: for something funny. Use it when a reaction adds genuine social signal — not on every message. ' +
|
|
44
|
-
'
|
|
45
|
-
'
|
|
44
|
+
'The reaction is only applied if you ALSO reply to this message in the same turn (via `channel_reply`/' +
|
|
45
|
+
'`channel_send`); if you stay silent or `skip_response`, the reaction is dropped. So do not use a reaction ' +
|
|
46
|
+
'as a standalone response to a message you are only observing — react to the messages you actually answer. ' +
|
|
46
47
|
'Pass the bare emoji name, no colons.',
|
|
47
48
|
parameters: Type.Object({
|
|
48
49
|
emoji: Type.String({
|
|
@@ -66,7 +67,7 @@ export function createChannelReactTool({
|
|
|
66
67
|
const reactionRef = getReactionRef()
|
|
67
68
|
if (reactionRef === undefined) return deny('this conversation has no message to react to')
|
|
68
69
|
|
|
69
|
-
const result = await router.
|
|
70
|
+
const result = await router.queueReactionAfterReply({
|
|
70
71
|
adapter: origin.adapter,
|
|
71
72
|
workspace: origin.workspace,
|
|
72
73
|
chat: origin.chat,
|
|
@@ -82,7 +83,7 @@ export function createChannelReactTool({
|
|
|
82
83
|
content: [
|
|
83
84
|
{
|
|
84
85
|
type: 'text' as const,
|
|
85
|
-
text: `${TOOL_RESULT_PREFIX}
|
|
86
|
+
text: `${TOOL_RESULT_PREFIX}will react with :${params.emoji}: on ${origin.adapter}:${origin.workspace}/${origin.chat} if you reply this turn`,
|
|
86
87
|
},
|
|
87
88
|
],
|
|
88
89
|
details,
|