owenloop 0.5.3 → 0.5.5
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 +25 -0
- package/README.md +6 -2
- package/dist/packages/work/src/agent/brief.d.ts +6 -0
- package/dist/packages/work/src/agent/brief.js +102 -2
- package/dist/packages/work/src/agent/loop.d.ts +10 -1
- package/dist/packages/work/src/agent/loop.js +28 -0
- package/dist/packages/work/src/agent/workdir.d.ts +22 -0
- package/dist/packages/work/src/agent/workdir.js +80 -0
- package/dist/packages/work/src/exec/loop.d.ts +13 -1
- package/dist/packages/work/src/exec/loop.js +184 -9
- package/dist/packages/work/src/harness/child-env.js +8 -0
- package/dist/packages/work/src/harness/claude.js +25 -6
- package/dist/packages/work/src/harness/codex.js +7 -0
- package/dist/packages/work/src/harness/contract.d.ts +22 -3
- package/dist/packages/work/src/harness/contract.js +22 -3
- package/dist/packages/work/src/hold/mcp.d.ts +11 -2
- package/dist/packages/work/src/hold/mcp.js +69 -2
- package/dist/packages/work/src/hub/client.d.ts +8 -1
- package/dist/packages/work/src/hub/client.js +1 -0
- package/dist/packages/work/src/hub/types.d.ts +29 -0
- package/dist/packages/work/src/roles/agent-run.js +9 -1
- package/dist/packages/work/src/roles/exec.js +9 -0
- package/dist/packages/work/src/roles/shift.js +9 -1
- package/dist/packages/work/src/settings/settings.d.ts +32 -5
- package/dist/packages/work/src/settings/settings.js +37 -12
- package/dist/packages/work/src/shift/runtime.d.ts +7 -0
- package/dist/packages/work/src/shift/runtime.js +29 -4
- package/dist/packages/work/src/shift/spawn.d.ts +12 -2
- package/dist/packages/work/src/shift/spawn.js +16 -3
- package/dist/src/cli.js +257 -7
- package/dist/src/config-dir.d.ts +62 -0
- package/dist/src/config-dir.js +87 -0
- package/dist/src/crypto/trust-roots.d.ts +8 -4
- package/dist/src/crypto/trust-roots.js +18 -12
- package/dist/src/defs.js +57 -16
- package/dist/src/engine.d.ts +46 -1
- package/dist/src/engine.js +97 -3
- package/dist/src/hub.d.ts +5 -1
- package/dist/src/hub.js +7 -8
- package/dist/src/index.d.ts +1 -1
- package/dist/src/index.js +1 -1
- package/dist/src/model.d.ts +29 -5
- package/dist/src/model.js +63 -14
- package/dist/src/paths.d.ts +47 -4
- package/dist/src/paths.js +36 -5
- package/dist/src/types.d.ts +8 -2
- package/dist/src/work-settings.d.ts +8 -5
- package/dist/src/work-settings.js +15 -12
- package/docs/authoring.md +40 -13
- package/docs/bundles.md +9 -0
- package/docs/cli.md +155 -24
- package/docs/crypto.md +8 -8
- package/docs/design.md +20 -2
- package/docs/embedding.md +5 -4
- package/docs/wire-contracts.md +3 -2
- package/package.json +1 -1
- package/plugins/claude-code/plugin/.claude-plugin/plugin.json +1 -1
- package/plugins/claude-code/plugin/.mcp.json +1 -1
- package/plugins/codex/plugins/owenloop/.codex-plugin/plugin.json +1 -1
- package/plugins/codex/plugins/owenloop/.mcp.json +1 -1
|
@@ -108,6 +108,14 @@ export const ADMITTED_OWENLOOP_KEYS = new Set([
|
|
|
108
108
|
'OWENLOOP_NO_KEYCHAIN',
|
|
109
109
|
// `holdSession` — src/roles/hold.ts. The fallback when `--session` is absent.
|
|
110
110
|
'OWENLOOP_SESSION',
|
|
111
|
+
// `owenloopConfigDir` — src/config-dir.ts. Scopes owenloop's OWN config
|
|
112
|
+
// directory (`settings.json`, `credentials.json`, `allowed_signers`) without
|
|
113
|
+
// touching `XDG_CONFIG_HOME`, which the rest of the operator's toolchain also
|
|
114
|
+
// reads. A path, not credential material — but the credential FILE backend
|
|
115
|
+
// resolves under it (`readStoredCredential` → `configDir` → src/hub.ts), so a
|
|
116
|
+
// mount that does not see this variable falls back to `XDG_CONFIG_HOME` and
|
|
117
|
+
// reads a DIFFERENT credential store than the shift that spawned it.
|
|
118
|
+
'OWENLOOP_CONFIG_DIR',
|
|
111
119
|
]);
|
|
112
120
|
/** The namespace this filter governs. Nothing outside it is ever touched. */
|
|
113
121
|
const NAMESPACE = 'OWENLOOP_';
|
|
@@ -186,11 +186,24 @@ const PERMISSION_MODES = [
|
|
|
186
186
|
* `canUseTool` callback — a human decision point — and denies when no callback
|
|
187
187
|
* is wired. That is the "human is the gate" position.
|
|
188
188
|
* - `auto-safe` → `auto`. `auto` is the SDK's model-classifier mode: it decides
|
|
189
|
-
* per call
|
|
190
|
-
*
|
|
191
|
-
*
|
|
192
|
-
*
|
|
193
|
-
*
|
|
189
|
+
* per call, with no prompt. NOT `acceptEdits`, which auto-approves filesystem
|
|
190
|
+
* operations by category and applies no judgment to anything else, and NOT
|
|
191
|
+
* `dontAsk`, which DENIES an unapproved tool rather than consulting anyone —
|
|
192
|
+
* silently narrower than what the step asked for.
|
|
193
|
+
*
|
|
194
|
+
* MEASURED against SDK 0.3.220, `canUseTool` wired throughout: the classifier
|
|
195
|
+
* consulted the host on NONE of five probe commands, and emitted no
|
|
196
|
+
* `permission_denied` system message for any of them. Four ran, including
|
|
197
|
+
* `rm -rf` on an ABSOLUTE path outside the session cwd (verified by the
|
|
198
|
+
* target directory being gone afterwards), a `curl` to the public internet,
|
|
199
|
+
* and a read of a file under `$HOME` outside cwd.
|
|
200
|
+
*
|
|
201
|
+
* So `auto` does not, in practice, escalate to whoever hosts the session.
|
|
202
|
+
* Wiring `canUseTool` does not give `auto-safe` a human exception path — the
|
|
203
|
+
* callback is simply never reached. Anything that needs a person in the loop
|
|
204
|
+
* has to run under `ask`, whose `default` mode does route to the callback
|
|
205
|
+
* (measured on the same SDK build, including calls originating inside a
|
|
206
|
+
* subagent, which arrive carrying that subagent's `agentID`).
|
|
194
207
|
* - `full-access` → `bypassPermissions`. The only SDK mode that runs tools with
|
|
195
208
|
* no prompt at all, and the one the companion flag below exists for.
|
|
196
209
|
*
|
|
@@ -245,7 +258,13 @@ const OWENLOOP_ONLY_NETWORK_TOOLS = new Set([
|
|
|
245
258
|
* unrestricted-network readers. */
|
|
246
259
|
const READ_ONLY_OWENLOOP_ONLY_NETWORK_TOOLS = [...READ_ONLY_TOOLS]
|
|
247
260
|
.filter((tool) => OWENLOOP_ONLY_NETWORK_TOOLS.has(tool));
|
|
248
|
-
|
|
261
|
+
// The tools a born-bound step agent gets no matter how restricted it is.
|
|
262
|
+
// `ask` is here for the same reason `submit` is: without it a restricted agent
|
|
263
|
+
// has no legal way to END. It could previously only submit (fabricating when it
|
|
264
|
+
// did not know) or fall silent (re-arming the step into a retry storm). `ask` is
|
|
265
|
+
// about the agent's OWN owed artifact, so unlike `reject` it grants no authority
|
|
266
|
+
// over anyone else's work and there is no reason to withhold it under isolation.
|
|
267
|
+
const BORN_BOUND_OWENLOOP_TOOL_NAMES = ['get_order', 'submit', 'ask'];
|
|
249
268
|
const BORN_BOUND_OWENLOOP_TOOLS = BORN_BOUND_OWENLOOP_TOOL_NAMES.map((name) => `mcp__owenloop__${name}`);
|
|
250
269
|
const RESTRICTED_OWENLOOP_DENIED_TOOLS = ['mcp__owenloop__reject'];
|
|
251
270
|
const OWENLOOP_CONTROL_TOOLS = new Set([
|
|
@@ -264,6 +264,13 @@ function isPlainMap(v) {
|
|
|
264
264
|
* keeps this list a complete statement of what the child needs.
|
|
265
265
|
*/
|
|
266
266
|
const MOUNT_ENV_KEEP = ['HOME', 'XDG_CONFIG_HOME', 'XDG_STATE_HOME', 'XDG_CACHE_HOME'];
|
|
267
|
+
/**
|
|
268
|
+
* `OWENLOOP_CONFIG_DIR` rides the `OWENLOOP_*` allowlist below rather than this
|
|
269
|
+
* list, but it belongs to the same job: without it a mount whose parent scoped
|
|
270
|
+
* owenloop's config with `OWENLOOP_CONFIG_DIR` would fall back to
|
|
271
|
+
* `XDG_CONFIG_HOME` and read a DIFFERENT settings file than the shift that
|
|
272
|
+
* spawned it.
|
|
273
|
+
*/
|
|
267
274
|
/**
|
|
268
275
|
* The environment for the owenloop mount, forwarded from THIS process.
|
|
269
276
|
*
|
|
@@ -138,9 +138,28 @@ export interface StepPermissions {
|
|
|
138
138
|
* - `full-access` — nothing is gated. The harness never consults a human, and
|
|
139
139
|
* a failed action is simply reported back to the model.
|
|
140
140
|
*
|
|
141
|
-
* `auto-safe` is NOT a milder `full-access`: it
|
|
142
|
-
* headless step that must not stall wants
|
|
143
|
-
* separately so that choice is stated rather
|
|
141
|
+
* `auto-safe` is NOT a milder `full-access`: the position it names still has a
|
|
142
|
+
* human exception path. A headless step that must not stall wants
|
|
143
|
+
* `full-access`, and the two exist separately so that choice is stated rather
|
|
144
|
+
* than inferred.
|
|
145
|
+
*
|
|
146
|
+
* WHAT `auto-safe` COSTS YOU TODAY, MEASURED. The three positions above
|
|
147
|
+
* describe the vocabulary, not a guarantee about how far any given classifier
|
|
148
|
+
* actually goes before it consults someone. On one shipped adapter the
|
|
149
|
+
* classifier was probed with five actions — among them a recursive delete on an
|
|
150
|
+
* absolute path outside the session working directory, an outbound network
|
|
151
|
+
* request, and a read of a file outside that directory — and it consulted the
|
|
152
|
+
* host on none of them and reported no auto-denial for any of them; four ran.
|
|
153
|
+
* On the other, the position that maps to `auto-safe` keys its judgment on
|
|
154
|
+
* filesystem access being restricted, so a step that also declares
|
|
155
|
+
* `filesystem: unrestricted` moves it close to `full-access` as well.
|
|
156
|
+
*
|
|
157
|
+
* The practical consequence for a definition author: choosing `auto-safe`
|
|
158
|
+
* because you want a person consulted on the dangerous cases may get you
|
|
159
|
+
* neither the prompt nor a record that one was skipped. Choose `ask` when a
|
|
160
|
+
* human decision genuinely has to happen, and `full-access` when it genuinely
|
|
161
|
+
* must not. Treat `auto-safe` as the classifier's own judgment with no
|
|
162
|
+
* guaranteed appeal, not as a supervised `full-access`.
|
|
144
163
|
*
|
|
145
164
|
* `ask` and `auto-safe` are meaningful only where something can answer a
|
|
146
165
|
* prompt. In a Shift-dispatched run nothing can, so a prompt is a stall: the
|
|
@@ -24,9 +24,28 @@
|
|
|
24
24
|
* - `full-access` — nothing is gated. The harness never consults a human, and
|
|
25
25
|
* a failed action is simply reported back to the model.
|
|
26
26
|
*
|
|
27
|
-
* `auto-safe` is NOT a milder `full-access`: it
|
|
28
|
-
* headless step that must not stall wants
|
|
29
|
-
* separately so that choice is stated rather
|
|
27
|
+
* `auto-safe` is NOT a milder `full-access`: the position it names still has a
|
|
28
|
+
* human exception path. A headless step that must not stall wants
|
|
29
|
+
* `full-access`, and the two exist separately so that choice is stated rather
|
|
30
|
+
* than inferred.
|
|
31
|
+
*
|
|
32
|
+
* WHAT `auto-safe` COSTS YOU TODAY, MEASURED. The three positions above
|
|
33
|
+
* describe the vocabulary, not a guarantee about how far any given classifier
|
|
34
|
+
* actually goes before it consults someone. On one shipped adapter the
|
|
35
|
+
* classifier was probed with five actions — among them a recursive delete on an
|
|
36
|
+
* absolute path outside the session working directory, an outbound network
|
|
37
|
+
* request, and a read of a file outside that directory — and it consulted the
|
|
38
|
+
* host on none of them and reported no auto-denial for any of them; four ran.
|
|
39
|
+
* On the other, the position that maps to `auto-safe` keys its judgment on
|
|
40
|
+
* filesystem access being restricted, so a step that also declares
|
|
41
|
+
* `filesystem: unrestricted` moves it close to `full-access` as well.
|
|
42
|
+
*
|
|
43
|
+
* The practical consequence for a definition author: choosing `auto-safe`
|
|
44
|
+
* because you want a person consulted on the dangerous cases may get you
|
|
45
|
+
* neither the prompt nor a record that one was skipped. Choose `ask` when a
|
|
46
|
+
* human decision genuinely has to happen, and `full-access` when it genuinely
|
|
47
|
+
* must not. Treat `auto-safe` as the classifier's own judgment with no
|
|
48
|
+
* guaranteed appeal, not as a supervised `full-access`.
|
|
30
49
|
*
|
|
31
50
|
* `ask` and `auto-safe` are meaningful only where something can answer a
|
|
32
51
|
* prompt. In a Shift-dispatched run nothing can, so a prompt is a stall: the
|
|
@@ -4,7 +4,7 @@
|
|
|
4
4
|
* A born-bound work-holder: the D2 stamped Step Agent's frontmatter declares
|
|
5
5
|
* `mcpServers.owenloop = owenloop work hold --order <wf>/<run> --origin <url> --mcp`,
|
|
6
6
|
* so when the Step Agent session boots it launches THIS as a stdio MCP server. The
|
|
7
|
-
* server exposes
|
|
7
|
+
* server exposes four bare tools the model uses to do its order:
|
|
8
8
|
* - `get_order` → the order packet (prompt, inputs, owed outputs) for the run
|
|
9
9
|
* this holder is bound to. No ids are arguments — they came in on argv, never
|
|
10
10
|
* through the model.
|
|
@@ -14,6 +14,15 @@
|
|
|
14
14
|
* - `reject` → invalidate a consumed artifact through the claiming step's
|
|
15
15
|
* server-derived authority; the client never supplies `by`. A CLOSED reject
|
|
16
16
|
* also stops the lease loop without releasing the already-closed claim.
|
|
17
|
+
* - `ask` → stop and escalate to a human about an OWED output path. The
|
|
18
|
+
* third exit. `submit` and `reject` both assume the worker can finish; until
|
|
19
|
+
* `ask` existed a worker that genuinely could not had only two moves, and
|
|
20
|
+
* both damaged the run: submit something it did not believe (greens a
|
|
21
|
+
* fabrication that every downstream step then builds on), or end its turn
|
|
22
|
+
* silently (the task re-arms and a fresh worker relearns the same blocker
|
|
23
|
+
* until the stall cap). `ask` holds the artifact with no counter movement
|
|
24
|
+
* and surfaces the question to an operator, who answers with
|
|
25
|
+
* `owenloop retry <workflow> <path> --text "<answer>"`.
|
|
17
26
|
*
|
|
18
27
|
* Underneath the tools, the SAME lease loop the CLI `hold` runs keeps the order's
|
|
19
28
|
* lease warm: `createHoldMcp` builds the loop with `onOrder` wired to capture the
|
|
@@ -33,7 +42,7 @@ import { type SubmissionKeyManager } from '../submit-proof.ts';
|
|
|
33
42
|
import type { SshProcessAdapter } from '../../../../src/crypto/ssh.ts';
|
|
34
43
|
import type { ConsumedVerifier } from '../consumed-verifier.ts';
|
|
35
44
|
import { type HoldLoop } from './loop.ts';
|
|
36
|
-
export declare const HOLD_MCP_TOOL_NAMES: readonly ["get_order", "submit", "reject"];
|
|
45
|
+
export declare const HOLD_MCP_TOOL_NAMES: readonly ["get_order", "submit", "reject", "ask"];
|
|
37
46
|
export type HoldMcpToolName = (typeof HOLD_MCP_TOOL_NAMES)[number];
|
|
38
47
|
export interface HoldMcpDeps {
|
|
39
48
|
hub: HubClient;
|
|
@@ -4,7 +4,7 @@
|
|
|
4
4
|
* A born-bound work-holder: the D2 stamped Step Agent's frontmatter declares
|
|
5
5
|
* `mcpServers.owenloop = owenloop work hold --order <wf>/<run> --origin <url> --mcp`,
|
|
6
6
|
* so when the Step Agent session boots it launches THIS as a stdio MCP server. The
|
|
7
|
-
* server exposes
|
|
7
|
+
* server exposes four bare tools the model uses to do its order:
|
|
8
8
|
* - `get_order` → the order packet (prompt, inputs, owed outputs) for the run
|
|
9
9
|
* this holder is bound to. No ids are arguments — they came in on argv, never
|
|
10
10
|
* through the model.
|
|
@@ -14,6 +14,15 @@
|
|
|
14
14
|
* - `reject` → invalidate a consumed artifact through the claiming step's
|
|
15
15
|
* server-derived authority; the client never supplies `by`. A CLOSED reject
|
|
16
16
|
* also stops the lease loop without releasing the already-closed claim.
|
|
17
|
+
* - `ask` → stop and escalate to a human about an OWED output path. The
|
|
18
|
+
* third exit. `submit` and `reject` both assume the worker can finish; until
|
|
19
|
+
* `ask` existed a worker that genuinely could not had only two moves, and
|
|
20
|
+
* both damaged the run: submit something it did not believe (greens a
|
|
21
|
+
* fabrication that every downstream step then builds on), or end its turn
|
|
22
|
+
* silently (the task re-arms and a fresh worker relearns the same blocker
|
|
23
|
+
* until the stall cap). `ask` holds the artifact with no counter movement
|
|
24
|
+
* and surfaces the question to an operator, who answers with
|
|
25
|
+
* `owenloop retry <workflow> <path> --text "<answer>"`.
|
|
17
26
|
*
|
|
18
27
|
* Underneath the tools, the SAME lease loop the CLI `hold` runs keeps the order's
|
|
19
28
|
* lease warm: `createHoldMcp` builds the loop with `onOrder` wired to capture the
|
|
@@ -30,7 +39,7 @@ import { textResult } from "../mcp/server.js";
|
|
|
30
39
|
import { buildSubmitProof } from "../submit-proof.js";
|
|
31
40
|
import { normalizeSubmitValue } from "../submit-value.js";
|
|
32
41
|
import { createHoldLoop } from "./loop.js";
|
|
33
|
-
export const HOLD_MCP_TOOL_NAMES = ['get_order', 'submit', 'reject'];
|
|
42
|
+
export const HOLD_MCP_TOOL_NAMES = ['get_order', 'submit', 'reject', 'ask'];
|
|
34
43
|
function errMsg(e) {
|
|
35
44
|
return e instanceof Error ? e.message : String(e);
|
|
36
45
|
}
|
|
@@ -261,10 +270,68 @@ export function createHoldMcp(deps) {
|
|
|
261
270
|
}
|
|
262
271
|
},
|
|
263
272
|
};
|
|
273
|
+
const askTool = {
|
|
274
|
+
name: 'ask',
|
|
275
|
+
description: 'Stop and ask a human about an output path you OWE, when you cannot produce it honestly — a required input is missing, wrong, or contradictory, or the order asks for a decision only a person can make. Use this INSTEAD of submitting a guess and INSTEAD of ending your turn without submitting: a guess greens and poisons every step downstream, and ending silently just re-runs this same step until it burns its retry budget. Asking does not count as a failure and costs you no attempts. The step is held until a human answers; your run ends here.',
|
|
276
|
+
inputSchema: {
|
|
277
|
+
type: 'object',
|
|
278
|
+
required: ['path', 'question'],
|
|
279
|
+
properties: {
|
|
280
|
+
path: {
|
|
281
|
+
type: 'string',
|
|
282
|
+
description: 'The owed output path you are blocked on (one of the `owes` paths from get_order).',
|
|
283
|
+
},
|
|
284
|
+
question: {
|
|
285
|
+
type: 'string',
|
|
286
|
+
description: 'The specific question for the human. State what decision or fact you need, not just that you are stuck — this is the whole message they receive.',
|
|
287
|
+
},
|
|
288
|
+
context: {
|
|
289
|
+
type: 'string',
|
|
290
|
+
description: 'Optional: what you already tried, what you read, and why it was not enough. Saves the human from reconstructing it from logs.',
|
|
291
|
+
},
|
|
292
|
+
},
|
|
293
|
+
additionalProperties: false,
|
|
294
|
+
},
|
|
295
|
+
handler: async (args) => {
|
|
296
|
+
const gone = terminalGuard();
|
|
297
|
+
if (gone !== undefined)
|
|
298
|
+
return gone;
|
|
299
|
+
const path = args['path'];
|
|
300
|
+
const question = args['question'];
|
|
301
|
+
const context = args['context'];
|
|
302
|
+
if (typeof path !== 'string' || path.trim() === '') {
|
|
303
|
+
return textResult({ error: 'ask requires a non-empty string "path"' }, true);
|
|
304
|
+
}
|
|
305
|
+
if (typeof question !== 'string' || question.trim() === '') {
|
|
306
|
+
return textResult({ error: 'ask requires a non-empty string "question"' }, true);
|
|
307
|
+
}
|
|
308
|
+
if (context !== undefined && typeof context !== 'string') {
|
|
309
|
+
return textResult({ error: 'ask "context" must be a string when present' }, true);
|
|
310
|
+
}
|
|
311
|
+
try {
|
|
312
|
+
const res = await hub.ask({
|
|
313
|
+
workflow,
|
|
314
|
+
run,
|
|
315
|
+
path,
|
|
316
|
+
question,
|
|
317
|
+
...(typeof context === 'string' && context.trim() !== '' ? { context } : {}),
|
|
318
|
+
});
|
|
319
|
+
// Asking ENDS the run. Same shape as submit/reject's closed branch:
|
|
320
|
+
// stop without releasing, because the hub already closed the claim.
|
|
321
|
+
if (res.closed === true)
|
|
322
|
+
loop.stop('asked', { release: false });
|
|
323
|
+
return textResult({ ok: res.ok, closed: res.closed ?? false, text: res.text });
|
|
324
|
+
}
|
|
325
|
+
catch (e) {
|
|
326
|
+
return textResult({ error: errMsg(e) }, true);
|
|
327
|
+
}
|
|
328
|
+
},
|
|
329
|
+
};
|
|
264
330
|
const registrations = {
|
|
265
331
|
get_order: getOrderTool,
|
|
266
332
|
submit: submitTool,
|
|
267
333
|
reject: rejectTool,
|
|
334
|
+
ask: askTool,
|
|
268
335
|
};
|
|
269
336
|
const selected = deps.tools ?? HOLD_MCP_TOOL_NAMES;
|
|
270
337
|
return { tools: selected.map((name) => registrations[name]), loop };
|
|
@@ -1,4 +1,4 @@
|
|
|
1
|
-
import type { GetOrderRequest, GetOrderResponse, HeartbeatRequest, HeartbeatResponse, PresencePingRequest, PresencePingResponse, ReleaseRequest, ReleaseResponse, RejectRequest, RejectResponse, ReportResolutionRequest, ReportResolutionResponse, SubmitRequest, SubmitResponse, WakeResponse, WhatsNextRequest, WhatsNextResponse, WhoamiResponse } from './types.ts';
|
|
1
|
+
import type { AskRequest, AskResponse, GetOrderRequest, GetOrderResponse, HeartbeatRequest, HeartbeatResponse, PresencePingRequest, PresencePingResponse, ReleaseRequest, ReleaseResponse, RejectRequest, RejectResponse, ReportResolutionRequest, ReportResolutionResponse, SubmitRequest, SubmitResponse, WakeResponse, WhatsNextRequest, WhatsNextResponse, WhoamiResponse } from './types.ts';
|
|
2
2
|
export interface HubClientOptions {
|
|
3
3
|
/** Hub origin, e.g. `https://hub.owenloop.dev` (no trailing slash needed). */
|
|
4
4
|
origin: string;
|
|
@@ -14,6 +14,13 @@ export interface HubClient {
|
|
|
14
14
|
release(req: ReleaseRequest): Promise<ReleaseResponse>;
|
|
15
15
|
submit(req: SubmitRequest): Promise<SubmitResponse>;
|
|
16
16
|
reject(req: RejectRequest): Promise<RejectResponse>;
|
|
17
|
+
/**
|
|
18
|
+
* ESCALATION: the worker stops and asks a human about an artifact it OWES.
|
|
19
|
+
* Distinct from `reject`, which is a verdict on somebody else's delivered
|
|
20
|
+
* work. Holds the artifact (no counter moves) until a human answers with
|
|
21
|
+
* `owenloop retry <workflow> <path> --text "<answer>"`.
|
|
22
|
+
*/
|
|
23
|
+
ask(req: AskRequest): Promise<AskResponse>;
|
|
17
24
|
/**
|
|
18
25
|
* Plan §6: record what this shift resolved the order's compound capability to,
|
|
19
26
|
* BEFORE the harness launches. Idempotent on the hub by order id, so a
|
|
@@ -82,6 +82,7 @@ export function createHubClient(opts) {
|
|
|
82
82
|
release: (req) => post('release', req),
|
|
83
83
|
submit: (req) => post('submit', req),
|
|
84
84
|
reject: (req) => post('reject', req),
|
|
85
|
+
ask: (req) => post('ask', req),
|
|
85
86
|
reportResolution: (req) => post('report_resolution', req),
|
|
86
87
|
whoami: () => get('whoami'),
|
|
87
88
|
// Cursor is an opaque non-negative integer; omit it entirely to bootstrap
|
|
@@ -367,6 +367,35 @@ export interface RejectResponse extends HubResponse {
|
|
|
367
367
|
ok: boolean;
|
|
368
368
|
closed?: boolean;
|
|
369
369
|
}
|
|
370
|
+
/**
|
|
371
|
+
* ESCALATION: the claiming worker cannot honestly build what it owes and stops
|
|
372
|
+
* to ask a person. `path` is one of the order's OWED paths — unlike `reject`,
|
|
373
|
+
* whose `path` is somebody else's delivered work, this is the worker's own debt.
|
|
374
|
+
*
|
|
375
|
+
* `by` is absent for the same reason it is absent on `reject`: the hub derives
|
|
376
|
+
* the actor from the claiming run's step, so a worker cannot ask on another
|
|
377
|
+
* step's behalf. `question` is what the human is being asked; `context` is
|
|
378
|
+
* optional supporting detail (what the worker already tried, which files it
|
|
379
|
+
* read) that rides the same reason-thread entry so the answering human does not
|
|
380
|
+
* have to reconstruct it from logs.
|
|
381
|
+
*/
|
|
382
|
+
export interface AskRequest {
|
|
383
|
+
workflow: string;
|
|
384
|
+
run: string;
|
|
385
|
+
path: string;
|
|
386
|
+
question: string;
|
|
387
|
+
context?: string;
|
|
388
|
+
}
|
|
389
|
+
/**
|
|
390
|
+
* The ask verb's flattened envelope. `closed` reports whether the escalation
|
|
391
|
+
* closed the worker's run — it does, always: a worker that has asked has
|
|
392
|
+
* nothing further to do on this order, and holding the lease open would only
|
|
393
|
+
* let it expire into a reap.
|
|
394
|
+
*/
|
|
395
|
+
export interface AskResponse extends HubResponse {
|
|
396
|
+
ok: boolean;
|
|
397
|
+
closed?: boolean;
|
|
398
|
+
}
|
|
370
399
|
export interface WhoamiResponse extends HubResponse {
|
|
371
400
|
orgId: string;
|
|
372
401
|
orgName: string;
|
|
@@ -96,7 +96,7 @@ import { adapterFor, defaultHarnessId, registeredHarnessIds } from "../harness/r
|
|
|
96
96
|
import { parseHarnessCarrier } from "../bundle/fetch.js";
|
|
97
97
|
import { normalizeStepPermissions, validateHarnessOptions } from "../harness/permissions.js";
|
|
98
98
|
import { appendSession, latestFor, sessionsPath } from "../harness/session-store.js";
|
|
99
|
-
import { ensureWorkDir, resolveWorkRepo, resolveWorkRoot } from "../agent/workdir.js";
|
|
99
|
+
import { ensureWorkDir, resolveAllowedWorkdirRoots, resolveWorkRepo, resolveWorkRoot, } from "../agent/workdir.js";
|
|
100
100
|
import { resolveShiftId, resolveTarget } from "./hold.js";
|
|
101
101
|
import { installSignalHandlers } from "./signals.js";
|
|
102
102
|
const DEFAULT_INTERVAL_MS = 60_000;
|
|
@@ -195,6 +195,7 @@ export function exitCodeFor(outcome) {
|
|
|
195
195
|
case 'completed':
|
|
196
196
|
return 0;
|
|
197
197
|
case 'misroute':
|
|
198
|
+
case 'workdir-denied':
|
|
198
199
|
case 'no-template':
|
|
199
200
|
case 'no-harness':
|
|
200
201
|
case 'incompatible-harness-policy':
|
|
@@ -495,6 +496,13 @@ export async function run(args, deps = {}) {
|
|
|
495
496
|
account,
|
|
496
497
|
...(shiftId !== undefined ? { shiftId } : {}),
|
|
497
498
|
cwd: workCwd,
|
|
499
|
+
// Machine policy, resolved from the SAME two sources the shift resolves it
|
|
500
|
+
// from: `OWENLOOP_ALLOWED_WORKDIR_ROOTS` (which `owenloop shift start`
|
|
501
|
+
// exports when the operator passed `--work-root`) then the settings file.
|
|
502
|
+
// Resolved against `process.cwd()`, not `workCwd`: a relative root in the
|
|
503
|
+
// env means "relative to where this process was launched", and `workCwd`
|
|
504
|
+
// may be a directory owenloop just created under the cache root.
|
|
505
|
+
allowedWorkdirRoots: resolveAllowedWorkdirRoots(env, settings.allowedWorkdirRoots, process.cwd()),
|
|
498
506
|
loadStep,
|
|
499
507
|
resolveAdapter,
|
|
500
508
|
...(capabilityModels !== undefined ? { capabilityModels } : {}),
|
|
@@ -34,6 +34,7 @@
|
|
|
34
34
|
import { hostname } from 'node:os';
|
|
35
35
|
import { createHubClient } from "../hub/client.js";
|
|
36
36
|
import { resolveBearer } from "../credentials/resolve.js";
|
|
37
|
+
import { resolveAllowedWorkdirRoots } from "../agent/workdir.js";
|
|
37
38
|
import { loadSettings } from "../settings/settings.js";
|
|
38
39
|
import { createExecLoop } from "../exec/loop.js";
|
|
39
40
|
import { createDefaultStoreInstructionResolver } from "../exec/instructions.js";
|
|
@@ -121,6 +122,7 @@ export function exitCodeFor(outcome) {
|
|
|
121
122
|
case 'judge-rejected':
|
|
122
123
|
return 0;
|
|
123
124
|
case 'misroute':
|
|
125
|
+
case 'workdir-denied':
|
|
124
126
|
case 'unresolved-instructions':
|
|
125
127
|
case 'killed':
|
|
126
128
|
case 'lease-lost':
|
|
@@ -173,6 +175,12 @@ export async function run(args, deps = {}) {
|
|
|
173
175
|
return 2;
|
|
174
176
|
}
|
|
175
177
|
const cwd = deps.cwd ?? process.cwd();
|
|
178
|
+
// Machine policy, resolved from the SAME two sources the shift resolves it
|
|
179
|
+
// from: `OWENLOOP_ALLOWED_WORKDIR_ROOTS` (which `owenloop shift start` exports
|
|
180
|
+
// when the operator passed `--work-root`) then the settings file. An
|
|
181
|
+
// independently launched `owenloop work exec` therefore honours the settings
|
|
182
|
+
// file on its own, with no shift involved.
|
|
183
|
+
const allowedWorkdirRoots = resolveAllowedWorkdirRoots(env, settings.allowedWorkdirRoots, cwd);
|
|
176
184
|
let instructions = deps.instructions;
|
|
177
185
|
if (instructions === undefined) {
|
|
178
186
|
try {
|
|
@@ -208,6 +216,7 @@ export async function run(args, deps = {}) {
|
|
|
208
216
|
holder,
|
|
209
217
|
instructions,
|
|
210
218
|
cwd,
|
|
219
|
+
allowedWorkdirRoots,
|
|
211
220
|
sleep: (ms) => new Promise((resolve) => setTimeout(resolve, ms)),
|
|
212
221
|
now: () => Date.now(),
|
|
213
222
|
out,
|
|
@@ -15,6 +15,7 @@ function usage() {
|
|
|
15
15
|
' [--cap <n>] [--max-agents <n>] [--poll-interval <ms>] [--once]\n' +
|
|
16
16
|
' [--cache-dir <p>] [--state-dir <p>]\n' +
|
|
17
17
|
' [--log-dir <p>] [--log-max-age <ms>]\n' +
|
|
18
|
+
' [--work-root <dir>]... restrict this shift to these directories\n' +
|
|
18
19
|
' owenloop shift next [--wait <seconds>] [--state-dir <p>]\n' +
|
|
19
20
|
' owenloop shift status [--state-dir <p>]\n' +
|
|
20
21
|
' owenloop shift end [--state-dir <p>]\n');
|
|
@@ -30,7 +31,7 @@ function parseStartArgs(args) {
|
|
|
30
31
|
const parsed = { crews: [] };
|
|
31
32
|
const valueFlags = new Set([
|
|
32
33
|
'--origin', '--as', '--name', '--cap', '--max-agents', '--poll-interval',
|
|
33
|
-
'--cache-dir', '--state-dir', '--log-dir', '--log-max-age',
|
|
34
|
+
'--cache-dir', '--state-dir', '--log-dir', '--log-max-age', '--work-root',
|
|
34
35
|
]);
|
|
35
36
|
for (let i = 0; i < args.length; i++) {
|
|
36
37
|
const raw = args[i];
|
|
@@ -87,6 +88,13 @@ function parseStartArgs(args) {
|
|
|
87
88
|
return { crews: parsed.crews, error: n };
|
|
88
89
|
parsed.pollIntervalMs = n;
|
|
89
90
|
}
|
|
91
|
+
else if (name === '--work-root') {
|
|
92
|
+
// REPEATABLE — one directory per occurrence. See `ParsedArgs.workRoots`
|
|
93
|
+
// in `src/shift/runtime.ts` for why this accumulates instead of taking
|
|
94
|
+
// a separated list, and `src/agent/workdir.ts` for how this differs
|
|
95
|
+
// from the singular `workRoot` setting.
|
|
96
|
+
(parsed.workRoots ??= []).push(value);
|
|
97
|
+
}
|
|
90
98
|
continue;
|
|
91
99
|
}
|
|
92
100
|
if (raw.startsWith('-'))
|
|
@@ -99,6 +99,25 @@ export interface Settings {
|
|
|
99
99
|
* of the MACHINE running the shift, not of the workflow.
|
|
100
100
|
*/
|
|
101
101
|
workRepo?: string;
|
|
102
|
+
/**
|
|
103
|
+
* The directories a shift on THIS machine may accept as an order's working
|
|
104
|
+
* directory. Absolute paths; a directory permits itself and everything under
|
|
105
|
+
* it.
|
|
106
|
+
*
|
|
107
|
+
* NOT `workRoot` (above) UNDER A DIFFERENT NAME. `workRoot` is a directory
|
|
108
|
+
* owenloop writes per-run subdirectories INTO. This is a list of directories
|
|
109
|
+
* a HUB-SUPPLIED `OrderPacket.workdir` is allowed to name. See
|
|
110
|
+
* `src/agent/workdir.ts` for the side-by-side table and the full rationale.
|
|
111
|
+
*
|
|
112
|
+
* UNSET OR EMPTY MEANS NO RESTRICTION, which is what every shift running
|
|
113
|
+
* today does. Default-closed would refuse work on upgrade, silently, on
|
|
114
|
+
* machines whose operator never asked for a boundary.
|
|
115
|
+
*
|
|
116
|
+
* Overridden by `OWENLOOP_ALLOWED_WORKDIR_ROOTS` (PATH-style, `:`-separated)
|
|
117
|
+
* and by `owenloop shift start --work-root <dir>` (repeatable). The override
|
|
118
|
+
* REPLACES this list; it does not extend it.
|
|
119
|
+
*/
|
|
120
|
+
allowedWorkdirRoots?: string[];
|
|
102
121
|
/**
|
|
103
122
|
* How THIS machine serves each capability: a flat map from capability to
|
|
104
123
|
* `{model, effort}`, keyed by the composed capability (`wise:deep`) or by the
|
|
@@ -113,7 +132,7 @@ export interface Settings {
|
|
|
113
132
|
capabilityModels?: Record<string, CapabilityModelRow>;
|
|
114
133
|
}
|
|
115
134
|
/** The known settings keys, in the order `owenloop work settings` prints them. */
|
|
116
|
-
export declare const KNOWN_SETTINGS_KEYS: readonly ["hubOrigin", "cacheDir", "stateDir", "dispatchCap", "shiftLogDir", "shiftLogMaxAgeMs", "commandRouting", "defPolicy", "artifactPolicy", "maxConcurrentAgents", "workRoot", "workRepo", "capabilityModels"];
|
|
135
|
+
export declare const KNOWN_SETTINGS_KEYS: readonly ["hubOrigin", "cacheDir", "stateDir", "dispatchCap", "shiftLogDir", "shiftLogMaxAgeMs", "commandRouting", "defPolicy", "artifactPolicy", "maxConcurrentAgents", "workRoot", "workRepo", "allowedWorkdirRoots", "capabilityModels"];
|
|
117
136
|
/**
|
|
118
137
|
* Keys owenloop still RECOGNIZES but no longer acts on. Each one produces a
|
|
119
138
|
* warning naming what replaced it, and is excluded from `unrecognized` so the
|
|
@@ -154,10 +173,18 @@ export interface SettingsInspection extends ValidatedSettings {
|
|
|
154
173
|
exists: boolean;
|
|
155
174
|
}
|
|
156
175
|
/**
|
|
157
|
-
* Resolve the settings file path from the caller's env
|
|
158
|
-
* `$
|
|
159
|
-
* non-
|
|
160
|
-
*
|
|
176
|
+
* Resolve the settings file path from the caller's env, through the one shared
|
|
177
|
+
* ladder in `src/config-dir.ts`: `$OWENLOOP_CONFIG_DIR/settings.json` when
|
|
178
|
+
* `OWENLOOP_CONFIG_DIR` is set and non-blank (it must be absolute), else
|
|
179
|
+
* `$XDG_CONFIG_HOME/owenloop/settings.json`, else
|
|
180
|
+
* `$HOME/.config/owenloop/settings.json`. Throws when no rung yields a value
|
|
181
|
+
* rather than guessing a home directory.
|
|
182
|
+
*
|
|
183
|
+
* `OWENLOOP_CONFIG_DIR` is the variable an operator running SEVERAL shifts on
|
|
184
|
+
* one machine should reach for. Setting `XDG_CONFIG_HOME` per shift also works,
|
|
185
|
+
* but a shift spreads its whole environment into every worker and every command
|
|
186
|
+
* step's child, so relocating `XDG_CONFIG_HOME` relocates the config of `gh`,
|
|
187
|
+
* `git`, `gcloud`, and everything else those steps shell out to.
|
|
161
188
|
*/
|
|
162
189
|
export declare function settingsPath(env: Record<string, string | undefined>): string;
|
|
163
190
|
/**
|
|
@@ -13,7 +13,8 @@
|
|
|
13
13
|
* own store (the CredentialReader seam) — there is deliberately no token knob.
|
|
14
14
|
*/
|
|
15
15
|
import { existsSync, readFileSync } from 'node:fs';
|
|
16
|
-
import { join } from 'node:path';
|
|
16
|
+
import { isAbsolute, join } from 'node:path';
|
|
17
|
+
import { owenloopConfigDir } from "../../../../src/config-dir.js";
|
|
17
18
|
import { validateCapabilityModels } from "../agent/capability-model.js";
|
|
18
19
|
/** The known settings keys, in the order `owenloop work settings` prints them. */
|
|
19
20
|
export const KNOWN_SETTINGS_KEYS = [
|
|
@@ -29,6 +30,7 @@ export const KNOWN_SETTINGS_KEYS = [
|
|
|
29
30
|
'maxConcurrentAgents',
|
|
30
31
|
'workRoot',
|
|
31
32
|
'workRepo',
|
|
33
|
+
'allowedWorkdirRoots',
|
|
32
34
|
'capabilityModels',
|
|
33
35
|
];
|
|
34
36
|
/**
|
|
@@ -42,19 +44,21 @@ export const KNOWN_SETTINGS_KEYS = [
|
|
|
42
44
|
*/
|
|
43
45
|
export const RETIRED_IGNORED_KEYS = ['escalateAt', 'escalationExtensionKey'];
|
|
44
46
|
/**
|
|
45
|
-
* Resolve the settings file path from the caller's env
|
|
46
|
-
* `$
|
|
47
|
-
* non-
|
|
48
|
-
*
|
|
47
|
+
* Resolve the settings file path from the caller's env, through the one shared
|
|
48
|
+
* ladder in `src/config-dir.ts`: `$OWENLOOP_CONFIG_DIR/settings.json` when
|
|
49
|
+
* `OWENLOOP_CONFIG_DIR` is set and non-blank (it must be absolute), else
|
|
50
|
+
* `$XDG_CONFIG_HOME/owenloop/settings.json`, else
|
|
51
|
+
* `$HOME/.config/owenloop/settings.json`. Throws when no rung yields a value
|
|
52
|
+
* rather than guessing a home directory.
|
|
53
|
+
*
|
|
54
|
+
* `OWENLOOP_CONFIG_DIR` is the variable an operator running SEVERAL shifts on
|
|
55
|
+
* one machine should reach for. Setting `XDG_CONFIG_HOME` per shift also works,
|
|
56
|
+
* but a shift spreads its whole environment into every worker and every command
|
|
57
|
+
* step's child, so relocating `XDG_CONFIG_HOME` relocates the config of `gh`,
|
|
58
|
+
* `git`, `gcloud`, and everything else those steps shell out to.
|
|
49
59
|
*/
|
|
50
60
|
export function settingsPath(env) {
|
|
51
|
-
|
|
52
|
-
if (xdg !== undefined && xdg.trim() !== '')
|
|
53
|
-
return join(xdg, 'owenloop', 'settings.json');
|
|
54
|
-
const home = env.HOME;
|
|
55
|
-
if (home !== undefined && home.trim() !== '')
|
|
56
|
-
return join(home, '.config', 'owenloop', 'settings.json');
|
|
57
|
-
throw new Error('cannot locate a settings directory: set HOME or XDG_CONFIG_HOME');
|
|
61
|
+
return join(owenloopConfigDir(env), 'settings.json');
|
|
58
62
|
}
|
|
59
63
|
/**
|
|
60
64
|
* Validate a parsed settings object. Checks the TYPE of every known key
|
|
@@ -111,6 +115,27 @@ export function validateSettings(raw, path) {
|
|
|
111
115
|
`Depth now rides on the capability's modifier suffix, not on a tier rung.`);
|
|
112
116
|
}
|
|
113
117
|
}
|
|
118
|
+
if ('allowedWorkdirRoots' in obj) {
|
|
119
|
+
// ABSOLUTE PATHS ONLY, and that is the one rule worth a hard error rather
|
|
120
|
+
// than a warning. A relative entry would resolve against whatever directory
|
|
121
|
+
// the shift process happened to start in — the very "wherever the shift was
|
|
122
|
+
// launched" assumption this key exists to remove — so a boundary written
|
|
123
|
+
// relatively would move with the launch directory and permit a different
|
|
124
|
+
// set of paths on every start.
|
|
125
|
+
const v = obj['allowedWorkdirRoots'];
|
|
126
|
+
if (!Array.isArray(v) || v.some((entry) => typeof entry !== 'string')) {
|
|
127
|
+
throw bad('allowedWorkdirRoots', 'an array of absolute directory paths', v);
|
|
128
|
+
}
|
|
129
|
+
for (const entry of v) {
|
|
130
|
+
if (entry.trim() === '') {
|
|
131
|
+
throw bad('allowedWorkdirRoots', 'an array of NON-EMPTY absolute directory paths', v);
|
|
132
|
+
}
|
|
133
|
+
if (!isAbsolute(entry)) {
|
|
134
|
+
throw new Error(`invalid settings file at ${path}: 'allowedWorkdirRoots' entry ${JSON.stringify(entry)} must be an ` +
|
|
135
|
+
`absolute path — a relative root would resolve against whatever directory the shift was launched in`);
|
|
136
|
+
}
|
|
137
|
+
}
|
|
138
|
+
}
|
|
114
139
|
if ('capabilityModels' in obj) {
|
|
115
140
|
const v = obj['capabilityModels'];
|
|
116
141
|
if (typeof v !== 'object' || v === null || Array.isArray(v)) {
|
|
@@ -35,6 +35,13 @@ export interface ParsedArgs {
|
|
|
35
35
|
logDir?: string;
|
|
36
36
|
/** `--log-max-age` — worker-log retention in milliseconds. `0` reaps eagerly. */
|
|
37
37
|
logMaxAgeMs?: number;
|
|
38
|
+
/**
|
|
39
|
+
* `--work-root <dir>` — REPEATABLE. Each occurrence adds one directory the
|
|
40
|
+
* shift may accept as an order's working directory; passing none leaves the
|
|
41
|
+
* shift unrestricted. Distinct from `settings.workRoot` (singular), which is
|
|
42
|
+
* where owenloop CREATES per-run directories — see `src/agent/workdir.ts`.
|
|
43
|
+
*/
|
|
44
|
+
workRoots?: string[];
|
|
38
45
|
error?: string;
|
|
39
46
|
}
|
|
40
47
|
/** Parse shift's internal `--flag value` and `--flag=value` forms. */
|