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.
Files changed (60) hide show
  1. package/CHANGELOG.md +25 -0
  2. package/README.md +6 -2
  3. package/dist/packages/work/src/agent/brief.d.ts +6 -0
  4. package/dist/packages/work/src/agent/brief.js +102 -2
  5. package/dist/packages/work/src/agent/loop.d.ts +10 -1
  6. package/dist/packages/work/src/agent/loop.js +28 -0
  7. package/dist/packages/work/src/agent/workdir.d.ts +22 -0
  8. package/dist/packages/work/src/agent/workdir.js +80 -0
  9. package/dist/packages/work/src/exec/loop.d.ts +13 -1
  10. package/dist/packages/work/src/exec/loop.js +184 -9
  11. package/dist/packages/work/src/harness/child-env.js +8 -0
  12. package/dist/packages/work/src/harness/claude.js +25 -6
  13. package/dist/packages/work/src/harness/codex.js +7 -0
  14. package/dist/packages/work/src/harness/contract.d.ts +22 -3
  15. package/dist/packages/work/src/harness/contract.js +22 -3
  16. package/dist/packages/work/src/hold/mcp.d.ts +11 -2
  17. package/dist/packages/work/src/hold/mcp.js +69 -2
  18. package/dist/packages/work/src/hub/client.d.ts +8 -1
  19. package/dist/packages/work/src/hub/client.js +1 -0
  20. package/dist/packages/work/src/hub/types.d.ts +29 -0
  21. package/dist/packages/work/src/roles/agent-run.js +9 -1
  22. package/dist/packages/work/src/roles/exec.js +9 -0
  23. package/dist/packages/work/src/roles/shift.js +9 -1
  24. package/dist/packages/work/src/settings/settings.d.ts +32 -5
  25. package/dist/packages/work/src/settings/settings.js +37 -12
  26. package/dist/packages/work/src/shift/runtime.d.ts +7 -0
  27. package/dist/packages/work/src/shift/runtime.js +29 -4
  28. package/dist/packages/work/src/shift/spawn.d.ts +12 -2
  29. package/dist/packages/work/src/shift/spawn.js +16 -3
  30. package/dist/src/cli.js +257 -7
  31. package/dist/src/config-dir.d.ts +62 -0
  32. package/dist/src/config-dir.js +87 -0
  33. package/dist/src/crypto/trust-roots.d.ts +8 -4
  34. package/dist/src/crypto/trust-roots.js +18 -12
  35. package/dist/src/defs.js +57 -16
  36. package/dist/src/engine.d.ts +46 -1
  37. package/dist/src/engine.js +97 -3
  38. package/dist/src/hub.d.ts +5 -1
  39. package/dist/src/hub.js +7 -8
  40. package/dist/src/index.d.ts +1 -1
  41. package/dist/src/index.js +1 -1
  42. package/dist/src/model.d.ts +29 -5
  43. package/dist/src/model.js +63 -14
  44. package/dist/src/paths.d.ts +47 -4
  45. package/dist/src/paths.js +36 -5
  46. package/dist/src/types.d.ts +8 -2
  47. package/dist/src/work-settings.d.ts +8 -5
  48. package/dist/src/work-settings.js +15 -12
  49. package/docs/authoring.md +40 -13
  50. package/docs/bundles.md +9 -0
  51. package/docs/cli.md +155 -24
  52. package/docs/crypto.md +8 -8
  53. package/docs/design.md +20 -2
  54. package/docs/embedding.md +5 -4
  55. package/docs/wire-contracts.md +3 -2
  56. package/package.json +1 -1
  57. package/plugins/claude-code/plugin/.claude-plugin/plugin.json +1 -1
  58. package/plugins/claude-code/plugin/.mcp.json +1 -1
  59. package/plugins/codex/plugins/owenloop/.codex-plugin/plugin.json +1 -1
  60. package/plugins/codex/plugins/owenloop/.mcp.json +1 -1
package/CHANGELOG.md CHANGED
@@ -4,6 +4,31 @@ All notable changes to this project are documented here. The format is based on
4
4
  [Keep a Changelog](https://keepachangelog.com/en/1.1.0/), and this project
5
5
  adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
6
6
 
7
+ ## [0.5.5](https://github.com/typicalday/owenloop/compare/v0.5.4...v0.5.5) (2026-08-15)
8
+
9
+
10
+ ### Features
11
+
12
+ * give every step agent a channel to ask a human ([#168](https://github.com/typicalday/owenloop/issues/168)) ([fa9e5f0](https://github.com/typicalday/owenloop/commit/fa9e5f08d3a9f55c6e95cc504966a6389ec699c4))
13
+ * let a shift operator declare where work may happen ([#166](https://github.com/typicalday/owenloop/issues/166)) ([71e9412](https://github.com/typicalday/owenloop/commit/71e9412c534ba826490dc62464989b47a6f551fe))
14
+
15
+ ## [0.5.4](https://github.com/typicalday/owenloop/compare/v0.5.3...v0.5.4) (2026-08-15)
16
+
17
+
18
+ ### Features
19
+
20
+ * add `owenloop cancel` to stop a running hub instance ([#158](https://github.com/typicalday/owenloop/issues/158)) ([76ccc75](https://github.com/typicalday/owenloop/commit/76ccc75cf4b1c9312484e691663f9e5c9e1ebe9a))
21
+ * add `owenloop instance show` to read a hub run's live state ([#161](https://github.com/typicalday/owenloop/issues/161)) ([9ab7cec](https://github.com/typicalday/owenloop/commit/9ab7cec6f72dc79c11c2c7079174f9581ff97304))
22
+ * let workdirFrom name a declared input, not only a consume ([#165](https://github.com/typicalday/owenloop/issues/165)) ([fff886b](https://github.com/typicalday/owenloop/commit/fff886b86c097a1b85de745faa322cf92d0e5615))
23
+ * state the submit contract in every rendered brief ([#160](https://github.com/typicalday/owenloop/issues/160)) ([f62430d](https://github.com/typicalday/owenloop/commit/f62430d1febc77597cafb25e8b2928f1ab7e7fb3))
24
+
25
+
26
+ ### Bug Fixes
27
+
28
+ * deliver a payload reject before the owed submits, not after ([#164](https://github.com/typicalday/owenloop/issues/164)) ([b38b709](https://github.com/typicalday/owenloop/commit/b38b709aededd0ebef2684f8c64abb8fe68226de))
29
+ * relay a failed command step's own output to the exec log ([#162](https://github.com/typicalday/owenloop/issues/162)) ([d40595d](https://github.com/typicalday/owenloop/commit/d40595dbfe40bfbf6f5d729296189e5b88424eec))
30
+ * scope owenloop's config dir without hijacking XDG_CONFIG_HOME ([#163](https://github.com/typicalday/owenloop/issues/163)) ([72838b4](https://github.com/typicalday/owenloop/commit/72838b45e3fb6f52038001148bca62f879a90efc))
31
+
7
32
  ## [0.5.3](https://github.com/typicalday/owenloop/compare/v0.5.2...v0.5.3) (2026-08-14)
8
33
 
9
34
 
package/README.md CHANGED
@@ -487,8 +487,12 @@ hub's own def hashes. `owenloop agent new
487
487
  Setting up a machine from scratch? `owenloop setup` may sign you in as a human
488
488
  and, when needed, mint or rekey and store a Scoped Identity. Setup writes only
489
489
  `hubOrigin` into the execution settings file, preserving the other keys. The
490
- file is `$XDG_CONFIG_HOME/owenloop/settings.json` when `XDG_CONFIG_HOME` is
491
- set to a non-blank value; otherwise `$HOME/.config/owenloop/settings.json`. For
490
+ file is `$OWENLOOP_CONFIG_DIR/settings.json` when `OWENLOOP_CONFIG_DIR` is set
491
+ to a non-blank absolute path; otherwise `$XDG_CONFIG_HOME/owenloop/settings.json`
492
+ when `XDG_CONFIG_HOME` is non-blank; otherwise
493
+ `$HOME/.config/owenloop/settings.json`. Prefer `OWENLOOP_CONFIG_DIR` when
494
+ running several shifts on one machine — `XDG_CONFIG_HOME` also relocates `gh`,
495
+ `git`, and `gcloud` for every command step the shift dispatches. For
492
496
  a non-default account,
493
497
  setup only prints the `OWENLOOP_ACCOUNT=<name>` instruction. Setup probes and,
494
498
  when needed, converges the bundled `owenloop` plugins for Claude Code and Codex.
@@ -69,6 +69,12 @@ export interface BriefSpec {
69
69
  * repeated rejections, not a first attempt.
70
70
  */
71
71
  escalated?: boolean;
72
+ /**
73
+ * The owed output paths for this order, in packet order — `packet.owes`
74
+ * mapped to its `path`. Feeds the submit contract; absent or empty renders
75
+ * the contract without naming any path.
76
+ */
77
+ owes?: readonly string[];
72
78
  }
73
79
  /**
74
80
  * Substitute the four dispatch tokens into a step template.
@@ -28,8 +28,108 @@ export function renderBrief(templateContent, spec) {
28
28
  .join(spec.account)
29
29
  .split(SHIFT_TOKEN)
30
30
  .join(spec.shiftId ?? '');
31
- const routing = renderRoutingLine(spec);
32
- return routing === '' ? substituted : `${routing}\n\n${substituted}`;
31
+ const blocks = [
32
+ renderRoutingLine(spec),
33
+ renderSubmitContract(spec),
34
+ renderEscalationContract(spec),
35
+ substituted,
36
+ ];
37
+ return blocks.filter((b) => b !== '').join('\n\n');
38
+ }
39
+ /**
40
+ * How an order is COMPLETED, stated to every agent on every run.
41
+ *
42
+ * WHY THIS EXISTS. Nothing else ever told an agent that finishing means calling
43
+ * a tool. The engine mounts the work-holder MCP server and then relies entirely
44
+ * on the def author's prose; owenloop's own house verb for it is "green <path>
45
+ * with <value>", which names no tool and is not a phrase any model has seen
46
+ * before. A strong enough model infers the tool call from the mounted surface
47
+ * and submits; a weaker one reads the same sentence as "print this JSON",
48
+ * writes the value into a code fence, and ends its turn. The hub then records
49
+ * nothing, the confirm grace lapses, and the order is re-offered — so the step
50
+ * loops forever while every attempt does the work correctly. That is a defect
51
+ * in what the engine states, not in the def or the model: the engine is the
52
+ * only party that knows the submit contract, so the engine has to state it.
53
+ *
54
+ * WHY IT NAMES THE OWED PATHS. `submit` is per-path and takes the path as an
55
+ * argument, so "call submit" alone is not actionable — an agent that owes two
56
+ * outputs and submits one is as stuck as one that submits none. The list comes
57
+ * from `packet.owes`, which is the hub's own answer to "what does this order
58
+ * still owe", so it stays correct on a re-offer where some paths are already
59
+ * paid.
60
+ *
61
+ * WHY THE COUNTER-EXAMPLE IS SPELLED OUT. The observed failure is not a refusal
62
+ * to submit — it is a belief that printing IS submitting. Naming the tool
63
+ * without ruling out the near-miss leaves that belief intact.
64
+ *
65
+ * Vendor-neutral by construction: it names owenloop's own tool and mount, and
66
+ * no harness. `test/vendor-gate.test.ts` fails any shipped file outside
67
+ * `src/harness/` that names a vendor.
68
+ */
69
+ function renderSubmitContract(spec) {
70
+ const owed = (spec.owes ?? []).filter((p) => p !== '');
71
+ // Nothing owed, nothing to say — the same stance the routing line takes on an
72
+ // absent modifier. A contract with no path to name would have to write
73
+ // `<path>` into its own example, which teaches the shape of the call while
74
+ // leaving the one argument that matters unanswered.
75
+ if (owed.length === 0)
76
+ return '';
77
+ return [
78
+ `How this order completes: submit your result with the \`submit\` tool on the mounted \`owenloop\` MCP server. You owe ${owed.map((p) => `\`${p}\``).join(', ')} — call it once for each.`,
79
+ `Example: \`submit({"path": "${owed[0]}", "value": <the value this brief asks for>})\`.`,
80
+ 'Printing that value as text, or inside a code fence, does NOT submit it — the turn ends, the hub records nothing, and this order is re-offered from scratch. Wherever this brief says "green <path> with <value>", it means exactly this tool call.',
81
+ ].join('\n');
82
+ }
83
+ /**
84
+ * What an agent does when it CANNOT complete the order, stated to every agent on
85
+ * every run, right after the submit contract.
86
+ *
87
+ * WHY THIS EXISTS. The submit contract closes one hole (an agent that finished
88
+ * but did not know how to say so). This closes the opposite one: an agent that
89
+ * did NOT finish and has no legal way to say so. Until `ask` shipped there were
90
+ * exactly two exits from a step, and a blocked agent had to pick the less
91
+ * damaging:
92
+ * - Submit anyway. The artifact goes green on a value the agent does not
93
+ * believe, and every downstream step builds on it. Observed on a real run:
94
+ * a planner that could read nothing produced a confident, fictional plan and
95
+ * a builder started executing it.
96
+ * - End the turn without submitting. The lease lapses, the task re-arms, and a
97
+ * fresh worker starts from the identical missing information. It repeats
98
+ * until the stall cap, spending a full model run per attempt to relearn the
99
+ * same blocker, and no human is ever told.
100
+ * Neither reaches a person. Both were watched happening. The tool is the fix and
101
+ * this paragraph is what makes the tool reachable — a mounted tool nobody is
102
+ * told about is dead code.
103
+ *
104
+ * WHY IT NAMES THE COST. Models are heavily biased toward producing an answer,
105
+ * and an escalation path that reads as failure will not be taken. So the text
106
+ * states the two facts that make asking the correct move rather than the
107
+ * cowardly one: it costs no attempts, and guessing is worse than asking.
108
+ *
109
+ * WHY IT STATES THE BAR. "Ask whenever unsure" would turn every ambiguity into a
110
+ * human interrupt and the channel would be ignored within a day. The bar is
111
+ * concrete: ask when the missing thing cannot be recovered by working — reading
112
+ * the repo, running a read-only command, re-reading the inputs — because those
113
+ * are exactly the cases where another attempt cannot help.
114
+ *
115
+ * Rendered only when the order owes something, for the same reason the submit
116
+ * contract is: `ask` takes an owed path as its argument, and a contract that has
117
+ * to write `<path>` into its own example teaches the call shape while leaving
118
+ * the one argument that matters unanswered.
119
+ *
120
+ * Vendor-neutral by construction — names owenloop's own tool and mount only.
121
+ */
122
+ function renderEscalationContract(spec) {
123
+ const owed = (spec.owes ?? []).filter((p) => p !== '');
124
+ if (owed.length === 0)
125
+ return '';
126
+ return [
127
+ `If you CANNOT produce what this order asks for, do not guess and do not end your turn silently — call the \`ask\` tool on the same \`owenloop\` MCP server and stop.`,
128
+ `Example: \`ask({"path": "${owed[0]}", "question": "<the specific decision or fact you need from a person>", "context": "<what you already tried>"})\`.`,
129
+ 'Ask when the thing you are missing cannot be recovered by working: a required input is absent, wrong, or contradicts another input; the order asks for a judgment only the operator can make; or you cannot reach something the order depends on. Do NOT ask for anything you could resolve by reading the repository, re-reading your inputs, or running a read-only command — do that first.',
130
+ 'Asking costs you nothing: it is not a failed attempt, it consumes none of this step\'s retry budget, and it is the correct outcome when the honest answer is "I do not know". Submitting a value you do not believe is worse than asking — it goes green and every later step builds on it. Ending your turn without submitting is worse than asking — this step simply re-runs with the same information missing, and no person is ever told.',
131
+ 'After you ask, your run is over. A human answers on the same artifact and a fresh attempt starts with their answer attached.',
132
+ ].join('\n');
33
133
  }
34
134
  /**
35
135
  * The run's depth, stated to the agent in one line, PREPENDED rather than
@@ -6,7 +6,7 @@ import type { HarnessAdapter } from '../harness/contract.ts';
6
6
  import { type SessionRecord } from '../harness/session-store.ts';
7
7
  import { type CapabilityModelMap, type CapabilityResolution } from './capability-model.ts';
8
8
  /** Discriminated result of one worker life; the role maps these to exit codes. */
9
- export type AgentRunOutcome = 'submitted' | 'completed' | 'misroute' | 'no-template' | 'no-harness' | 'incompatible-harness-policy' | 'unresolvable-capability' | 'unverified-consumed' | 'session-store-failed' | 'no-submit' | 'killed' | 'lease-lost' | 'ownership-error' | 'hub-unreachable' | 'stopped';
9
+ export type AgentRunOutcome = 'submitted' | 'completed' | 'misroute' | 'workdir-denied' | 'no-template' | 'no-harness' | 'incompatible-harness-policy' | 'unresolvable-capability' | 'unverified-consumed' | 'session-store-failed' | 'no-submit' | 'killed' | 'lease-lost' | 'ownership-error' | 'hub-unreachable' | 'stopped';
10
10
  /**
11
11
  * Locate the normalized step spec for an order. Injected because the lookup is
12
12
  * I/O (the verified workflow store) and because resolving the local definition
@@ -52,6 +52,15 @@ export interface AgentRunLoopOptions {
52
52
  shiftId?: string;
53
53
  /** cwd for the step agent when the order packet carries no `workdir`. */
54
54
  cwd: string;
55
+ /**
56
+ * The directories this MACHINE's operator declared as places work may happen,
57
+ * already resolved to absolute paths by `resolveAllowedWorkdirRoots`.
58
+ *
59
+ * NOT `workRoot`. `workRoot` is the single directory owenloop CREATES worktrees
60
+ * under; this is the set of directories an ORDER is permitted to name as its
61
+ * cwd. UNSET or EMPTY means NO RESTRICTION.
62
+ */
63
+ allowedWorkdirRoots?: string[];
55
64
  loadStep: StepLoader;
56
65
  resolveAdapter: AdapterResolver;
57
66
  /**
@@ -57,6 +57,7 @@
57
57
  * back would waste a whole re-offer cycle to learn something already known.
58
58
  */
59
59
  import { existsSync } from 'node:fs';
60
+ import { isWorkdirAllowed } from "./workdir.js";
60
61
  import { createLeaseLoop } from "../lease/loop.js";
61
62
  import { isResumeUnavailable } from "../harness/contract.js";
62
63
  import { preflightStepPermissions } from "../harness/permissions.js";
@@ -363,6 +364,30 @@ export function createAgentRunLoop(opts) {
363
364
  }
364
365
  stepName = packet.step;
365
366
  recordCwd = packet.workdir ?? opts.cwd;
367
+ // MACHINE POLICY. The operator who started this shift named the directories
368
+ // work may happen in (`owenloop shift start --work-root`, or
369
+ // `allowedWorkdirRoots` in settings). This worker is one of only two
370
+ // processes that ever sees `OrderPacket.workdir` — the shift's `whats_next`
371
+ // sweep receives a `WorkOrder`, which carries no such field — so the check
372
+ // belongs here and cannot be hoisted into dispatch.
373
+ //
374
+ // RELEASED, not failed, and BEFORE the step spec is loaded or any provider
375
+ // session is opened. Declining work this machine was not configured to host
376
+ // is not a failure of the work: the targeted release returns the order to
377
+ // the hub's pickup window, where a differently-configured machine can take
378
+ // it. `misroute` takes the same shape for the same reason.
379
+ //
380
+ // ONLY an order-NAMED workdir is checked. When the packet names none this
381
+ // worker uses `<workRoot>/<workflow>/<run>/` — a directory owenloop ITSELF
382
+ // created under the operator's own cache root — and denying that would deny
383
+ // every agent order on a machine that declared any root at all.
384
+ if (packet.workdir !== undefined &&
385
+ !isWorkdirAllowed(packet.workdir, opts.allowedWorkdirRoots ?? [])) {
386
+ const roots = (opts.allowedWorkdirRoots ?? []).join(', ');
387
+ opts.err(`owenloop work agent-run: step '${packet.step}' (${order}) names workdir '${packet.workdir}', ` +
388
+ `which is outside every work root this machine declared (${roots}) — releasing for the pickup window`);
389
+ return releaseWith('workdir-denied', 'workdir-denied');
390
+ }
366
391
  // The step spec. No bundle / no spec ⇒ we cannot brief anybody; release so
367
392
  // the order lapses back through the hub's pickup window.
368
393
  let material;
@@ -402,6 +427,9 @@ export function createAgentRunLoop(opts) {
402
427
  // statement about the RUN, and a shift has no say in it.
403
428
  ...(packet.modifier !== undefined ? { modifier: packet.modifier } : {}),
404
429
  ...(packet.escalated === true ? { escalated: true } : {}),
430
+ // The hub's own answer to "what does this order still owe", so the submit
431
+ // contract names the right paths on a re-offer where some are already paid.
432
+ owes: packet.owes.map((owed) => owed.path),
405
433
  };
406
434
  // Permissions arrive PRE-NORMALIZED in the spec. `prepare` ran
407
435
  // `normalizeStepPermissions` over the step's `x.harness` options at cache
@@ -22,6 +22,28 @@ export declare function resolveWorkRepo(env: Record<string, string | undefined>,
22
22
  * run's directory and must never be removed.
23
23
  */
24
24
  export declare function isUnderWorkRoot(dir: string, workRoot: string): boolean;
25
+ /**
26
+ * Resolve the declared roots: `OWENLOOP_ALLOWED_WORKDIR_ROOTS` >
27
+ * `settings.allowedWorkdirRoots` > none. Mirrors `resolveWorkRoot`'s precedence
28
+ * shape, so there is one rule to remember across this file.
29
+ *
30
+ * The env var is PATH-style `:`-separated because it names a list of
31
+ * directories and that is the convention every operator already knows for one.
32
+ * It REPLACES the settings list rather than adding to it — a narrowing override
33
+ * that could only ever widen would not be a usable safety control.
34
+ *
35
+ * Every entry is resolved to an absolute path against `cwd`, so a relative
36
+ * declaration means what the operator saw when they typed it.
37
+ */
38
+ export declare function resolveAllowedWorkdirRoots(env: Record<string, string | undefined>, settingsRoots: string[] | undefined, cwd: string): string[];
39
+ /**
40
+ * Is `workdir` inside (or equal to) one of `roots`?
41
+ *
42
+ * An EMPTY `roots` returns true for every path — that is the unrestricted
43
+ * default, not an accident, and it is why callers can pass the resolved list
44
+ * unconditionally instead of branching on whether the operator declared one.
45
+ */
46
+ export declare function isWorkdirAllowed(workdir: string, roots: string[]): boolean;
25
47
  export interface EnsureWorkDirOptions {
26
48
  workRoot: string;
27
49
  workflow: string;
@@ -81,6 +81,86 @@ export function isUnderWorkRoot(dir, workRoot) {
81
81
  const rel = relative(resolve(workRoot), resolve(dir));
82
82
  return rel !== '' && !rel.startsWith('..') && !isAbsolute(rel);
83
83
  }
84
+ /**
85
+ * ── OPERATOR-DECLARED WORKDIR ROOTS ─────────────────────────────────────────
86
+ *
87
+ * A SEPARATE MECHANISM FROM `workRoot` ABOVE, and the two are easy to confuse,
88
+ * so state the difference once:
89
+ *
90
+ * | | `workRoot` (singular) | `allowedWorkdirRoots` (plural) |
91
+ * |---|---|---|
92
+ * | what it is | a directory owenloop WRITES INTO | a list of directories orders may NAME |
93
+ * | who supplies the path | this machine's operator | the run, via `OrderPacket.workdir` |
94
+ * | what it does | holds `<workflow>/<run>/` per-run dirs | permits or refuses a cwd |
95
+ * | unset means | default to `<cacheDir>/work` | NO RESTRICTION |
96
+ *
97
+ * WHY THIS EXISTS. `OrderPacket.workdir` is an opaque location hint the HUB
98
+ * supplies and the worker obeys (`src/exec/loop.ts`, `src/agent/loop.ts`). Since
99
+ * a def may set it from a declared INPUT — `workdirFrom: target.path`, the
100
+ * grammar owenloop 0.5.4 added — the string can originate with whoever started
101
+ * the run, and a shift is not project-bound. So without a local check, a run on
102
+ * the hub picks any absolute path on the operator's machine as a working
103
+ * directory.
104
+ *
105
+ * The boundary belongs to the MACHINE'S OWNER, not to the hub: the person who
106
+ * starts a shift decides which directories that shift may work in. That is why
107
+ * the declaration is a `shift start` flag / settings key / env var and never a
108
+ * def field, a hub policy, or anything a run can influence.
109
+ *
110
+ * DEFAULT-OPEN, DELIBERATELY. An empty list means every path is allowed, which
111
+ * preserves the behaviour of every shift running today. Making it default-closed
112
+ * would break every existing shift on upgrade, and — worse — would do so as a
113
+ * silent refusal of work the hub believes was accepted.
114
+ *
115
+ * INCLUSIVE OF THE ROOT ITSELF, unlike `isUnderWorkRoot`. An operator who
116
+ * declares `/Users/alex/code/dev` means "this project", and `target.path` for
117
+ * that project IS that exact directory. `isUnderWorkRoot` excludes `dir ===
118
+ * root` because there it guards a DELETION and the root must survive; here it
119
+ * guards a cwd, and there is nothing to protect the root from.
120
+ */
121
+ /** Split a PATH-style (`:`-separated) list of directories, dropping empties. */
122
+ function splitRootList(raw) {
123
+ return raw
124
+ .split(':')
125
+ .map((entry) => entry.trim())
126
+ .filter((entry) => entry !== '');
127
+ }
128
+ /**
129
+ * Resolve the declared roots: `OWENLOOP_ALLOWED_WORKDIR_ROOTS` >
130
+ * `settings.allowedWorkdirRoots` > none. Mirrors `resolveWorkRoot`'s precedence
131
+ * shape, so there is one rule to remember across this file.
132
+ *
133
+ * The env var is PATH-style `:`-separated because it names a list of
134
+ * directories and that is the convention every operator already knows for one.
135
+ * It REPLACES the settings list rather than adding to it — a narrowing override
136
+ * that could only ever widen would not be a usable safety control.
137
+ *
138
+ * Every entry is resolved to an absolute path against `cwd`, so a relative
139
+ * declaration means what the operator saw when they typed it.
140
+ */
141
+ export function resolveAllowedWorkdirRoots(env, settingsRoots, cwd) {
142
+ const override = env['OWENLOOP_ALLOWED_WORKDIR_ROOTS'];
143
+ const raw = override !== undefined && override.trim() !== ''
144
+ ? splitRootList(override)
145
+ : (settingsRoots ?? []).map((entry) => entry.trim()).filter((entry) => entry !== '');
146
+ return raw.map((entry) => resolve(cwd, entry));
147
+ }
148
+ /**
149
+ * Is `workdir` inside (or equal to) one of `roots`?
150
+ *
151
+ * An EMPTY `roots` returns true for every path — that is the unrestricted
152
+ * default, not an accident, and it is why callers can pass the resolved list
153
+ * unconditionally instead of branching on whether the operator declared one.
154
+ */
155
+ export function isWorkdirAllowed(workdir, roots) {
156
+ if (roots.length === 0)
157
+ return true;
158
+ const target = resolve(workdir);
159
+ return roots.some((root) => {
160
+ const rel = relative(resolve(root), target);
161
+ return rel === '' || (!rel.startsWith('..') && !isAbsolute(rel));
162
+ });
163
+ }
84
164
  function defaultRunGit(args, cwd) {
85
165
  const r = spawnSync('git', args, { cwd, encoding: 'utf8' });
86
166
  return { status: r.status, output: `${r.stdout ?? ''}${r.stderr ?? ''}`.trim() };
@@ -22,7 +22,7 @@ import type { SshProcessAdapter } from '../../../../src/crypto/ssh.ts';
22
22
  */
23
23
  export declare const CONSUMES_INLINE_MAX_BYTES = 65536;
24
24
  /** Discriminated result of an exec run; the role maps these to exit codes. */
25
- export type ExecOutcome = 'submitted' | 'completed' | 'misroute' | 'unresolved-instructions' | 'killed' | 'lease-lost' | 'ownership-error' | 'hub-unreachable' | 'submit-rejected' | 'submit-failed' | 'rejected' | 'judge-rejected' | 'judge-no-verdict' | 'reject-failed' | 'stopped';
25
+ export type ExecOutcome = 'submitted' | 'completed' | 'misroute' | 'workdir-denied' | 'unresolved-instructions' | 'killed' | 'lease-lost' | 'ownership-error' | 'hub-unreachable' | 'submit-rejected' | 'submit-failed' | 'rejected' | 'judge-rejected' | 'judge-no-verdict' | 'reject-failed' | 'stopped';
26
26
  export interface ExecLoopOptions {
27
27
  hub: HubClient;
28
28
  runner: CommandRunner;
@@ -40,6 +40,18 @@ export interface ExecLoopOptions {
40
40
  instructions: InstructionResolver;
41
41
  /** cwd for the command when the order packet carries no `workdir`. */
42
42
  cwd: string;
43
+ /**
44
+ * The directories this MACHINE's operator declared as places work may happen,
45
+ * already resolved to absolute paths by `resolveAllowedWorkdirRoots`.
46
+ *
47
+ * NOT `workRoot`. `workRoot` is the single directory owenloop CREATES worktrees
48
+ * under; this is the set of directories an ORDER is permitted to name as its
49
+ * cwd. They answer different questions and neither derives from the other.
50
+ *
51
+ * UNSET or EMPTY means NO RESTRICTION — the pre-existing behaviour, and the
52
+ * only default that does not break every shift already running.
53
+ */
54
+ allowedWorkdirRoots?: string[];
43
55
  sleep: (ms: number) => Promise<void>;
44
56
  now: () => number;
45
57
  random?: () => number;
@@ -20,9 +20,12 @@
20
20
  * 3. RUN + RACE — the runner shells the command out while the lease loop runs.
21
21
  * Whichever settles first wins (plan decision 9):
22
22
  * - an ordinary command settles (any exit code, or a machinery error) ⇒
23
- * build a receipt and `submit` it to every owed path. A payload reject
24
- * follows those submits. Exit 0 means the receipt/reject delivery won;
25
- * the receipt's exit code still carries the command result.
23
+ * build a receipt. A payload reject is delivered FIRST (the hub refuses
24
+ * a reject once the claim has closed, and the last owed submit is what
25
+ * closes it); then `submit` the receipt to every owed path, unless that
26
+ * reject closed the run, in which case the owed paths stay debts. Exit 0
27
+ * means the receipt/reject delivery won; the receipt's exit code still
28
+ * carries the command result.
26
29
  * - a judge command exits 0 ⇒ submit its receipt; a non-zero exit ⇒ send
27
30
  * `reject` for `order.judge` without a receipt; signal or machinery
28
31
  * failure ⇒ no verdict, leave the claim for the reap path.
@@ -42,6 +45,7 @@ import { createHash } from 'node:crypto';
42
45
  import { mkdtempSync, rmSync, writeFileSync } from 'node:fs';
43
46
  import { tmpdir } from 'node:os';
44
47
  import { join, resolve } from 'node:path';
48
+ import { isWorkdirAllowed } from "../agent/workdir.js";
45
49
  import { createLeaseLoop } from "../lease/loop.js";
46
50
  import { parsePayloadLine } from "./payload.js";
47
51
  import { buildReceipt } from "./receipt.js";
@@ -243,6 +247,102 @@ export function createExecLoop(opts) {
243
247
  return 'reject-failed';
244
248
  }
245
249
  }
250
+ /**
251
+ * Deliver a PAYLOAD reject before the owed submits, and say whether this
252
+ * worker is finished.
253
+ *
254
+ * The distinction from `issueReject` is the third outcome. A judge reject is
255
+ * always the end of the run; a payload reject may leave the claim open, and
256
+ * when it does the owed receipts still have to land. So this returns either a
257
+ * terminal `ExecOutcome` or the sentinel `'continue'`.
258
+ *
259
+ * refused / threw → terminal. `issueReject` has already logged it and
260
+ * settled the lease. Nothing is submitted: a reject that
261
+ * did not land must not be followed by a receipt that
262
+ * greens the path the reject was protesting.
263
+ * landed, closed → terminal `'rejected'`. The rejected path was consumed
264
+ * by this firing, the hub closed the run `no_work`, and
265
+ * the owed paths stay debts for the next firing.
266
+ * landed, still open → `'continue'`. The rejected path was not consumed
267
+ * here, so the claim and the consume fingerprint are
268
+ * both intact and every owed submit proceeds normally.
269
+ *
270
+ * `closed` is read from the hub's own response rather than re-derived from the
271
+ * order, because the consumed-input test lives in the hub
272
+ * (`reject-artifact.ts` checks `runRow.fingerprint`, which is what the engine
273
+ * recorded at claim time — not what the def declares the step consumes).
274
+ */
275
+ async function relayPayloadReject(path, text, owed) {
276
+ let res;
277
+ try {
278
+ res = await hub.reject({ workflow, run: runId, path, text });
279
+ }
280
+ catch (e) {
281
+ opts.err(`owenloop work exec: reject of ${path} failed: ${errMsg(e)}`);
282
+ // Nothing has been submitted yet, so unlike the post-submit case there
283
+ // are no committed receipts to reason about. Release and let the step be
284
+ // re-offered.
285
+ lease.stop('reject-failed');
286
+ await leasePromise;
287
+ return 'reject-failed';
288
+ }
289
+ if (res.ok !== true) {
290
+ opts.err(`owenloop work exec: reject of ${path} was refused: ${res.text}`);
291
+ lease.stop('reject-failed', res.closed === true ? { release: false } : undefined);
292
+ await leasePromise;
293
+ return 'reject-failed';
294
+ }
295
+ if (res.closed === true) {
296
+ opts.out(`owenloop work exec: rejected ${path} — run closed, ${owed} owed path(s) left as debts`);
297
+ lease.stop('rejected', { release: false });
298
+ await leasePromise;
299
+ return 'rejected';
300
+ }
301
+ opts.out(`owenloop work exec: rejected ${path} (not consumed by this firing) — submitting owed receipts`);
302
+ return 'continue';
303
+ }
304
+ /**
305
+ * Relay the child's own output to this process's stderr when the command did
306
+ * not succeed.
307
+ *
308
+ * ## Why this exists
309
+ *
310
+ * Until this was added, a command step that failed produced a worker log that
311
+ * named the failure and said NOTHING about its cause. The runner captures the
312
+ * last 4 KiB of combined stdout+stderr as `outputTail` and puts it in the
313
+ * receipt — but a receipt only reaches a human if the submit is ACCEPTED, and
314
+ * the most common command-step failure is precisely the one where it is not:
315
+ * the script dies before printing a payload line, so the receipt has no
316
+ * `payload` key and `merge`'s schema rejects it. The one artifact carrying the
317
+ * diagnosis was discarded by the same event that made it necessary.
318
+ *
319
+ * Measured on `wf_40bd0c3f6783f9d31291d74d`: the `merger` step failed four
320
+ * times across two shifts and one foreground `owenloop work exec`, and every
321
+ * one of those logs contained the same three lines — holding, running,
322
+ * schema-rejected — with no trace of what the child said. Reproducing the
323
+ * script by hand outside the engine was the only way to see its stderr.
324
+ *
325
+ * `outputTail` is already capped at 4 KiB by the runner, so this cannot flood
326
+ * a log. Each line is prefixed so the child's words are never mistaken for
327
+ * exec's own.
328
+ */
329
+ function relayChildOutput(result, step) {
330
+ if (result.exitCode === 0 && result.error === undefined)
331
+ return;
332
+ const how = result.error !== undefined
333
+ ? `could not be run (${result.error})`
334
+ : result.signal !== undefined
335
+ ? `was killed by ${result.signal}`
336
+ : `exited ${result.exitCode}`;
337
+ opts.err(`owenloop work exec: the command for step '${step}' ${how}; its last output follows`);
338
+ const tail = result.outputTail.replace(/\n+$/, '');
339
+ if (tail === '') {
340
+ opts.err(' (the command produced no output)');
341
+ return;
342
+ }
343
+ for (const line of tail.split('\n'))
344
+ opts.err(` | ${line}`);
345
+ }
246
346
  /** Build the receipt and submit it to every owed path. */
247
347
  async function submitReceipt(result, order, resolvedCommand) {
248
348
  if (signalled) {
@@ -255,6 +355,9 @@ export function createExecLoop(opts) {
255
355
  await leasePromise;
256
356
  return 'killed';
257
357
  }
358
+ // Before any of the branches below decide what to do about the failure.
359
+ // Every one of them is reachable with a useless log otherwise.
360
+ relayChildOutput(result, order.step);
258
361
  const parsedPayload = parsePayloadLine(result.payloadLine, result.payloadOverCap);
259
362
  if (order.judge !== undefined) {
260
363
  if (result.exitCode === null) {
@@ -286,6 +389,54 @@ export function createExecLoop(opts) {
286
389
  run: runId,
287
390
  step: order.step,
288
391
  }, parsedPayload);
392
+ // ---- the payload reject goes FIRST, and it is not an ordering nicety ----
393
+ //
394
+ // It used to run after every owed submit, on the reasoning that "all owed
395
+ // submissions must land before invalidating a consumed input; rejection
396
+ // moves the consume fingerprint and would born-reject later submits from
397
+ // this run". The second half of that sentence is true. The first half made
398
+ // the reject UNDELIVERABLE, because the submits it waited on are what kill
399
+ // the claim it needs.
400
+ //
401
+ // The hub refuses a reject from a run whose claim has closed
402
+ // (`hub-core/src/verbs/reject-artifact.ts`: `held` requires
403
+ // `runRow.outcome === undefined`, and the refusal is explicit —
404
+ // "is not currently held by an open claim — nothing was rejected"). A step's
405
+ // last owed submit closes the run. So for the single-owed-path case — which
406
+ // is every gate in the delivery workflow — submit-then-reject could only
407
+ // ever produce the pair observed on `run_ecfedb23a84194e446159e67`:
408
+ //
409
+ // submitted receipt to mergeable (green)
410
+ // reject of pr was refused: ... is not currently held by an open claim
411
+ //
412
+ // The gate greened its own output and its reject evaporated, so `merger`
413
+ // proceeded on a PR the gate had explicitly refused to confirm. The def-side
414
+ // cascade that was supposed to un-green `mergeable` never fired, because it
415
+ // hangs off a `pr` rejection that never happened.
416
+ //
417
+ // Rejecting first is correct in both directions, and which one applies is
418
+ // the hub's answer, not a guess:
419
+ //
420
+ // `closed: true` — the rejected path WAS one of this firing's consumed
421
+ // inputs, so the hub closed the run `no_work` (not `failed`: no attempt
422
+ // is burned). The owed paths stay debts and the step re-fires against the
423
+ // rebuilt input. Skipping the submits here is not a loss — an artifact
424
+ // derived from an input this same command just declared bad is exactly
425
+ // what the engine's dead-input cascade would have thrown away anyway.
426
+ // This is also the branch that makes the old born-reject warning moot:
427
+ // there are no later submits to born-reject.
428
+ //
429
+ // `closed: false` — the rejected path was NOT consumed by this firing, so
430
+ // the claim is still open and the fingerprint has not moved. Fall through
431
+ // and submit every owed path exactly as before.
432
+ //
433
+ // A REFUSED reject now submits nothing. That is the point: previously the
434
+ // receipt had already landed and greened a path whose gate had failed.
435
+ if (order.judge === undefined && parsedPayload.reject !== undefined) {
436
+ const rejected = await relayPayloadReject(parsedPayload.reject.path, parsedPayload.reject.text, order.owes.length);
437
+ if (rejected !== 'continue')
438
+ return rejected;
439
+ }
289
440
  for (const owe of order.owes) {
290
441
  let res;
291
442
  try {
@@ -332,12 +483,6 @@ export function createExecLoop(opts) {
332
483
  }
333
484
  opts.out(`owenloop work exec: submitted receipt to ${owe.path} (${outcome})`);
334
485
  }
335
- if (order.judge === undefined && parsedPayload.reject !== undefined) {
336
- // All owed submissions must land before invalidating a consumed input;
337
- // rejection moves the consume fingerprint and would born-reject later
338
- // submits from this run.
339
- return issueReject(parsedPayload.reject.path, parsedPayload.reject.text, 'rejected');
340
- }
341
486
  // Every owed path landed — the run has closed, so stop WITHOUT releasing.
342
487
  lease.stop('submitted', { release: false });
343
488
  await leasePromise;
@@ -363,6 +508,36 @@ export function createExecLoop(opts) {
363
508
  await leasePromise;
364
509
  return 'misroute';
365
510
  }
511
+ // MACHINE POLICY, checked before any instruction is resolved and long before
512
+ // anything is spawned. The operator who started this shift named the
513
+ // directories work may happen in (`owenloop shift start --work-root`, or
514
+ // `allowedWorkdirRoots` in settings); an order naming anything else is
515
+ // refused HERE, in the worker, because the worker is the only process that
516
+ // ever sees `OrderPacket.workdir` — the shift's `whats_next` sweep receives
517
+ // a `WorkOrder`, which has no such field.
518
+ //
519
+ // RELEASED, not failed. This is not a defect in the order and not a failure
520
+ // of the work: it is one machine declining work it was not configured to
521
+ // host. A targeted release returns the order to the hub's pickup window so a
522
+ // differently-configured machine can take it, exactly like `misroute`.
523
+ // Submitting a failure receipt instead would kill a run that is perfectly
524
+ // valid somewhere else.
525
+ //
526
+ // ONLY an order-NAMED workdir is checked. An order that declares none
527
+ // inherits this worker's own launch directory, which the operator chose
528
+ // themselves when they started the shift — bounding an operator's own
529
+ // choice by the operator's own roots protects nobody and would deny every
530
+ // step that legitimately declares no workdir.
531
+ if (order.workdir !== undefined &&
532
+ !isWorkdirAllowed(order.workdir, opts.allowedWorkdirRoots ?? [])) {
533
+ const roots = (opts.allowedWorkdirRoots ?? []).join(', ');
534
+ opts.err(`owenloop work exec: step '${order.step}' (${workflow}/${runId}) names workdir ` +
535
+ `'${order.workdir}', which is outside every work root this machine declared (${roots}) — ` +
536
+ 'releasing for the pickup window');
537
+ lease.stop('workdir-denied'); // targeted release — local policy, not a failure
538
+ await leasePromise;
539
+ return 'workdir-denied';
540
+ }
366
541
  let resolvedCommand;
367
542
  let resolvedBundleDir;
368
543
  try {