owenloop 0.5.0 → 0.5.2
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 +67 -0
- package/README.md +144 -35
- package/bin/owenloop.mjs +49 -1
- package/dist/packages/work/src/agent/brief.d.ts +29 -4
- package/dist/packages/work/src/agent/brief.js +49 -4
- package/dist/packages/work/src/agent/capability-model.d.ts +123 -0
- package/dist/packages/work/src/agent/capability-model.js +148 -0
- package/dist/packages/work/src/agent/loop.d.ts +56 -7
- package/dist/packages/work/src/agent/loop.js +166 -9
- package/dist/packages/work/src/bundle/cache.d.ts +7 -5
- package/dist/packages/work/src/bundle/cache.js +20 -6
- package/dist/packages/work/src/bundle/fetch.d.ts +7 -5
- package/dist/packages/work/src/bundle/fetch.js +27 -7
- package/dist/packages/work/src/bundle/types.d.ts +8 -4
- package/dist/packages/work/src/consumed-verifier.d.ts +45 -0
- package/dist/packages/work/src/consumed-verifier.js +205 -0
- package/dist/packages/work/src/exec/instructions.d.ts +99 -0
- package/dist/packages/work/src/exec/instructions.js +323 -0
- package/dist/packages/work/src/exec/loop.d.ts +29 -1
- package/dist/packages/work/src/exec/loop.js +313 -35
- package/dist/packages/work/src/exec/payload.d.ts +29 -0
- package/dist/packages/work/src/exec/payload.js +54 -0
- package/dist/packages/work/src/exec/receipt.d.ts +9 -6
- package/dist/packages/work/src/exec/receipt.js +12 -1
- package/dist/packages/work/src/exec/runner.d.ts +7 -1
- package/dist/packages/work/src/exec/runner.js +84 -3
- package/dist/packages/work/src/harness/child-env.js +7 -0
- package/dist/packages/work/src/harness/claude.d.ts +20 -8
- package/dist/packages/work/src/harness/claude.js +260 -58
- package/dist/packages/work/src/harness/codex.d.ts +5 -4
- package/dist/packages/work/src/harness/codex.js +169 -41
- package/dist/packages/work/src/harness/contract.d.ts +36 -11
- package/dist/packages/work/src/harness/fake.js +1 -0
- package/dist/packages/work/src/harness/jsonrpc-stdio.js +7 -0
- package/dist/packages/work/src/harness/permissions.d.ts +16 -3
- package/dist/packages/work/src/harness/permissions.js +102 -13
- package/dist/packages/work/src/harness/session-store.d.ts +30 -10
- package/dist/packages/work/src/harness/session-store.js +317 -89
- package/dist/packages/work/src/hold/mcp.d.ts +25 -5
- package/dist/packages/work/src/hold/mcp.js +134 -12
- package/dist/packages/work/src/hub/client.d.ts +8 -1
- package/dist/packages/work/src/hub/client.js +15 -1
- package/dist/packages/work/src/hub/types.d.ts +137 -33
- package/dist/packages/work/src/hub/types.js +8 -10
- package/dist/packages/work/src/mcp/server.js +1 -1
- package/dist/packages/work/src/owenloop-bin.d.ts +1 -0
- package/dist/packages/work/src/owenloop-bin.js +18 -0
- package/dist/packages/work/src/roles/agent-run.d.ts +12 -6
- package/dist/packages/work/src/roles/agent-run.js +115 -34
- package/dist/packages/work/src/roles/exec.d.ts +8 -0
- package/dist/packages/work/src/roles/exec.js +27 -2
- package/dist/packages/work/src/roles/hold.d.ts +6 -0
- package/dist/packages/work/src/roles/hold.js +83 -13
- package/dist/packages/work/src/roles/lint.d.ts +3 -4
- package/dist/packages/work/src/roles/lint.js +41 -7
- package/dist/packages/work/src/roles/settings.js +19 -2
- package/dist/packages/work/src/roles/shift.d.ts +1 -0
- package/dist/packages/work/src/roles/shift.js +36 -2
- package/dist/packages/work/src/settings/settings.d.ts +70 -1
- package/dist/packages/work/src/settings/settings.js +90 -4
- package/dist/packages/work/src/shift/logretention.d.ts +253 -0
- package/dist/packages/work/src/shift/logretention.js +407 -0
- package/dist/packages/work/src/shift/logsink.d.ts +14 -0
- package/dist/packages/work/src/shift/logsink.js +98 -0
- package/dist/packages/work/src/shift/loop.d.ts +32 -4
- package/dist/packages/work/src/shift/loop.js +687 -166
- package/dist/packages/work/src/shift/protocol.d.ts +128 -1
- package/dist/packages/work/src/shift/protocol.js +14 -0
- package/dist/packages/work/src/shift/runtime.d.ts +16 -0
- package/dist/packages/work/src/shift/runtime.js +291 -5
- package/dist/packages/work/src/shift/server.d.ts +13 -0
- package/dist/packages/work/src/shift/server.js +61 -36
- package/dist/packages/work/src/shift/spawn.d.ts +115 -19
- package/dist/packages/work/src/shift/spawn.js +234 -30
- package/dist/packages/work/src/shift/state.d.ts +76 -49
- package/dist/packages/work/src/shift/state.js +350 -103
- package/dist/packages/work/src/shift/truncate.d.ts +35 -0
- package/dist/packages/work/src/shift/truncate.js +73 -0
- package/dist/packages/work/src/submit-proof.d.ts +61 -0
- package/dist/packages/work/src/submit-proof.js +119 -0
- package/dist/packages/work/src/submit-value.d.ts +48 -0
- package/dist/packages/work/src/submit-value.js +118 -0
- package/dist/packages/work/src/usage.d.ts +1 -1
- package/dist/packages/work/src/usage.js +11 -0
- package/dist/src/add.d.ts +64 -259
- package/dist/src/add.js +81 -615
- package/dist/src/archive.d.ts +97 -0
- package/dist/src/archive.js +591 -0
- package/dist/src/bundle/index.d.ts +64 -0
- package/dist/src/bundle/index.js +508 -0
- package/dist/src/bundle/manifest.d.ts +50 -0
- package/dist/src/bundle/manifest.js +435 -0
- package/dist/src/bundle/runtime.d.ts +26 -0
- package/dist/src/bundle/runtime.js +64 -0
- package/dist/src/bundle/tar.d.ts +62 -0
- package/dist/src/bundle/tar.js +238 -0
- package/dist/src/bundle/types.d.ts +105 -0
- package/dist/src/bundle/types.js +21 -0
- package/dist/src/capabilities.d.ts +93 -0
- package/dist/src/capabilities.js +101 -0
- package/dist/src/cli.d.ts +90 -20
- package/dist/src/cli.js +2421 -374
- package/dist/src/credentials.d.ts +4 -3
- package/dist/src/credentials.js +21 -19
- package/dist/src/crypto/allowed-signers.d.ts +68 -0
- package/dist/src/crypto/allowed-signers.js +307 -0
- package/dist/src/crypto/canonical.d.ts +8 -0
- package/dist/src/crypto/canonical.js +92 -0
- package/dist/src/crypto/chain.d.ts +65 -0
- package/dist/src/crypto/chain.js +377 -0
- package/dist/src/crypto/dsse.d.ts +209 -0
- package/dist/src/crypto/dsse.js +282 -0
- package/dist/src/crypto/enrollment.d.ts +78 -0
- package/dist/src/crypto/enrollment.js +186 -0
- package/dist/src/crypto/index.d.ts +48 -0
- package/dist/src/crypto/index.js +44 -0
- package/dist/src/crypto/keys.d.ts +258 -0
- package/dist/src/crypto/keys.js +908 -0
- package/dist/src/crypto/org-root.d.ts +33 -0
- package/dist/src/crypto/org-root.js +78 -0
- package/dist/src/crypto/origin-rules.d.ts +56 -0
- package/dist/src/crypto/origin-rules.js +143 -0
- package/dist/src/crypto/policy-floor.d.ts +98 -0
- package/dist/src/crypto/policy-floor.js +353 -0
- package/dist/src/crypto/records.d.ts +319 -0
- package/dist/src/crypto/records.js +168 -0
- package/dist/src/crypto/scope.d.ts +49 -0
- package/dist/src/crypto/scope.js +83 -0
- package/dist/src/crypto/ssh.d.ts +171 -0
- package/dist/src/crypto/ssh.js +396 -0
- package/dist/src/crypto/submission.d.ts +33 -0
- package/dist/src/crypto/submission.js +47 -0
- package/dist/src/crypto/trust-roots.d.ts +29 -0
- package/dist/src/crypto/trust-roots.js +39 -0
- package/dist/src/crypto/verify-consumed.d.ts +67 -0
- package/dist/src/crypto/verify-consumed.js +259 -0
- package/dist/src/crypto/verify-origin.d.ts +50 -0
- package/dist/src/crypto/verify-origin.js +151 -0
- package/dist/src/crypto/verify-publication.d.ts +60 -0
- package/dist/src/crypto/verify-publication.js +150 -0
- package/dist/src/defs.d.ts +24 -1
- package/dist/src/defs.js +335 -27
- package/dist/src/engine.d.ts +211 -6
- package/dist/src/engine.js +459 -59
- package/dist/src/factory.d.ts +16 -0
- package/dist/src/factory.js +16 -4
- package/dist/src/global-config.d.ts +59 -0
- package/dist/src/global-config.js +84 -0
- package/dist/src/hub.d.ts +102 -32
- package/dist/src/hub.js +147 -58
- package/dist/src/index.d.ts +19 -2
- package/dist/src/index.js +23 -2
- package/dist/src/install.d.ts +538 -0
- package/dist/src/install.js +1981 -0
- package/dist/src/lock.d.ts +15 -57
- package/dist/src/lock.js +251 -143
- package/dist/src/mcp/serve.d.ts +46 -22
- package/dist/src/mcp/serve.js +174 -44
- package/dist/src/mcp/server.js +12 -1
- package/dist/src/model.d.ts +51 -0
- package/dist/src/model.js +157 -1
- package/dist/src/order-resolver.d.ts +184 -0
- package/dist/src/order-resolver.js +252 -0
- package/dist/src/package-version.d.ts +10 -0
- package/dist/src/package-version.js +42 -0
- package/dist/src/paths.d.ts +14 -1
- package/dist/src/paths.js +18 -0
- package/dist/src/schemas/enrollment-grant.v1.schema.json +90 -0
- package/dist/src/schemas/index.d.ts +1433 -0
- package/dist/src/schemas/index.js +31 -0
- package/dist/src/schemas/order.v1.schema.json +120 -0
- package/dist/src/schemas/origin.v1.schema.json +68 -0
- package/dist/src/schemas/policy-floor.v1.schema.json +29 -0
- package/dist/src/schemas/publication.v1.schema.json +31 -0
- package/dist/src/schemas/revocation.v1.schema.json +39 -0
- package/dist/src/schemas/submission.v1.schema.json +61 -0
- package/dist/src/store/def-source.d.ts +52 -0
- package/dist/src/store/def-source.js +335 -0
- package/dist/src/store/index-file.d.ts +69 -0
- package/dist/src/store/index-file.js +151 -0
- package/dist/src/store/index.d.ts +27 -0
- package/dist/src/store/index.js +22 -0
- package/dist/src/store/ingestor.d.ts +43 -0
- package/dist/src/store/ingestor.js +211 -0
- package/dist/src/store/install.d.ts +228 -0
- package/dist/src/store/install.js +555 -0
- package/dist/src/store/instruction-source.d.ts +49 -0
- package/dist/src/store/instruction-source.js +298 -0
- package/dist/src/store/origin-evidence.d.ts +21 -0
- package/dist/src/store/origin-evidence.js +66 -0
- package/dist/src/store/pre-commit-verifier.d.ts +68 -0
- package/dist/src/store/pre-commit-verifier.js +418 -0
- package/dist/src/store/resolve.d.ts +138 -0
- package/dist/src/store/resolve.js +363 -0
- package/dist/src/store/types.d.ts +288 -0
- package/dist/src/store/types.js +356 -0
- package/dist/src/store.d.ts +33 -1
- package/dist/src/store.js +57 -4
- package/dist/src/types.d.ts +146 -12
- package/dist/src/untar.d.ts +8 -23
- package/dist/src/untar.js +10 -136
- package/docs/agent-runner.md +86 -0
- package/docs/authoring.md +64 -6
- package/docs/bundles.md +350 -0
- package/docs/cli.md +1298 -106
- package/docs/crypto.md +832 -0
- package/docs/design.md +342 -20
- package/docs/embedding.md +132 -25
- package/docs/shift-logs.md +642 -0
- package/docs/wire-contracts.md +395 -0
- package/package.json +6 -1
- package/plugins/claude-code/.claude-plugin/marketplace.json +14 -0
- package/plugins/claude-code/plugin/.claude-plugin/plugin.json +8 -0
- package/plugins/claude-code/plugin/.mcp.json +14 -0
- package/plugins/claude-code/plugin/hooks/hooks.json +26 -0
- package/plugins/claude-code/plugin/hooks/session-end.sh +40 -0
- package/plugins/claude-code/plugin/hooks/session-start.sh +52 -0
- package/plugins/claude-code/plugin/skills/author/SKILL.md +181 -0
- package/plugins/claude-code/plugin/skills/conduct/SKILL.md +106 -0
- package/plugins/claude-code/plugin/skills/shift/SKILL.md +105 -0
- package/plugins/codex/.agents/plugins/marketplace.json +21 -0
- package/plugins/codex/plugins/owenloop/.codex-plugin/plugin.json +7 -0
- package/plugins/codex/plugins/owenloop/.mcp.json +20 -0
- package/plugins/codex/plugins/owenloop/hooks/hooks.json +26 -0
- package/plugins/codex/plugins/owenloop/hooks/session-end.sh +40 -0
- package/plugins/codex/plugins/owenloop/hooks/session-start.sh +52 -0
- package/plugins/codex/plugins/owenloop/skills/author/SKILL.md +181 -0
- package/plugins/codex/plugins/owenloop/skills/conduct/SKILL.md +106 -0
- package/plugins/codex/plugins/owenloop/skills/shift/SKILL.md +105 -0
|
@@ -0,0 +1,123 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* Resolve an order's COMPOSED CAPABILITY to the model and reasoning effort this
|
|
3
|
+
* machine runs it at.
|
|
4
|
+
*
|
|
5
|
+
* The hub decides WHAT grade of work an order is (`builder` composed with the
|
|
6
|
+
* run's `deep` modifier → `builder:deep`). It never decides which model serves
|
|
7
|
+
* that grade. That decision is local to the operator's machine, because it is a
|
|
8
|
+
* statement about accounts, quotas and taste — three things a shared workflow
|
|
9
|
+
* definition has no business asserting.
|
|
10
|
+
*
|
|
11
|
+
* So the settings file carries a flat map from capability to `{model, effort}`:
|
|
12
|
+
*
|
|
13
|
+
* ```json
|
|
14
|
+
* {
|
|
15
|
+
* "capabilityModels": {
|
|
16
|
+
* "wise:deep": { "model": "<model-id>", "effort": "xhigh" },
|
|
17
|
+
* "build:deep": { "model": "<model-id>", "effort": "xhigh" },
|
|
18
|
+
* "wise": { "model": "<model-id>", "effort": "high" }
|
|
19
|
+
* }
|
|
20
|
+
* }
|
|
21
|
+
* ```
|
|
22
|
+
*
|
|
23
|
+
* `<model-id>` is whatever string the shift's harness accepts as a model. This
|
|
24
|
+
* module never enumerates real model ids: it ships as neutral runtime, and
|
|
25
|
+
* `test/vendor-gate.test.ts` forbids a harness vendor's names here. `README.md`
|
|
26
|
+
* and `docs/` carry worked examples with real ids.
|
|
27
|
+
*
|
|
28
|
+
* THIS REPLACES `tierMap`/`tierProfiles` OUTRIGHT. A tier was an abstract rung
|
|
29
|
+
* (`fast`/`standard`/`strong`/`strongest`) that a machine then mapped to a
|
|
30
|
+
* model, and the tier ladder could not express the thing the operator actually
|
|
31
|
+
* wanted: "Luna at low effort" and "Fable at max effort" are different routes,
|
|
32
|
+
* but the ladder welded model choice and depth into one ordered axis, so one
|
|
33
|
+
* could not vary without the other. A capability is just a name. What serves it
|
|
34
|
+
* is a local lookup, and depth rides on the capability's modifier suffix rather
|
|
35
|
+
* than on a rung.
|
|
36
|
+
*
|
|
37
|
+
* The one behavior worth carrying out of the tier code is EFFORT VALIDITY
|
|
38
|
+
* CHECKING, which now happens at settings load rather than at resolution time.
|
|
39
|
+
* It checks the effort against `EFFORT_LADDER` — the neutral start contract's
|
|
40
|
+
* own five rungs — and nothing else.
|
|
41
|
+
*
|
|
42
|
+
* IT DELIBERATELY DOES NOT CHECK EFFORT AGAINST THE MODEL. Doing that would
|
|
43
|
+
* require a table of real model ids in this file, which the vendor gate forbids
|
|
44
|
+
* and which this repository has no ground truth for anyway. Look at what the
|
|
45
|
+
* harness adapters in `src/harness/` actually enforce: one validates effort
|
|
46
|
+
* against a closed union that is a property of THE HARNESS — every model that
|
|
47
|
+
* harness runs accepts the same rungs — and the other passes effort through
|
|
48
|
+
* without validating it at all. Neither knows a per-model constraint, so a
|
|
49
|
+
* table here would be either a restatement of `EFFORT_LADDER` under
|
|
50
|
+
* model-shaped keys, or an invention that blocks a valid operator config. The
|
|
51
|
+
* real per-harness check belongs on the adapter contract; see the follow-up
|
|
52
|
+
* noted in `validateCapabilityModels`.
|
|
53
|
+
*
|
|
54
|
+
* The old snap-up/ceiling-refusal logic in `resolveEffort` is gone with the
|
|
55
|
+
* ladder it resolved through: there is no abstract rung left to snap.
|
|
56
|
+
*/
|
|
57
|
+
/** Reasoning rungs the neutral start contract accepts, weakest to strongest. */
|
|
58
|
+
export declare const EFFORT_LADDER: readonly ["low", "medium", "high", "xhigh", "max"];
|
|
59
|
+
export type Effort = (typeof EFFORT_LADDER)[number];
|
|
60
|
+
/**
|
|
61
|
+
* What separates a capability's NAME PART from its MODIFIER: `wise:deep` is the
|
|
62
|
+
* name `wise` at the modifier `deep`. Split on the FIRST separator, matching
|
|
63
|
+
* the engine's own `capabilityName` — a capability name may not contain `:`
|
|
64
|
+
* (install-time rule), so anything after the first one is modifier territory.
|
|
65
|
+
*/
|
|
66
|
+
export declare const MODIFIER_SEPARATOR = ":";
|
|
67
|
+
/** The bare capability name inside a composed capability. `wise:deep` → `wise`. */
|
|
68
|
+
export declare function capabilityNamePart(capability: string): string;
|
|
69
|
+
/** One settings row: the model that serves a capability, and how hard it thinks. */
|
|
70
|
+
export interface CapabilityModelRow {
|
|
71
|
+
model: string;
|
|
72
|
+
effort: string;
|
|
73
|
+
}
|
|
74
|
+
/** The settings map, keyed by composed capability (`wise:deep`) or bare name (`wise`). */
|
|
75
|
+
export type CapabilityModelMap = Readonly<Record<string, CapabilityModelRow>>;
|
|
76
|
+
/** Thrown when the capability map is unusable. Never silently repaired. */
|
|
77
|
+
export declare class CapabilityModelError extends Error {
|
|
78
|
+
}
|
|
79
|
+
/**
|
|
80
|
+
* Validate the map, throwing on the first unusable row. Returns nothing: every
|
|
81
|
+
* fault this function can see is certainly wrong.
|
|
82
|
+
*
|
|
83
|
+
* WHAT IT CATCHES — a capability key that is empty, a row that is not an object,
|
|
84
|
+
* a missing or empty `model`, and an `effort` that is not one of
|
|
85
|
+
* `EFFORT_LADDER`. Each is wrong under every harness and every model, and a
|
|
86
|
+
* shift that started anyway would refuse or misroute its first matching order
|
|
87
|
+
* hours later, far from the typo.
|
|
88
|
+
*
|
|
89
|
+
* WHAT IT DOES NOT CATCH — an effort a particular MODEL or HARNESS rejects, and
|
|
90
|
+
* a model id that does not exist. Neither is knowable here (see this module's
|
|
91
|
+
* header). FOLLOW-UP: `HarnessAdapter` is the right owner for the first one —
|
|
92
|
+
* an adapter that validates effort against a closed union could declare that
|
|
93
|
+
* union, and a composition root could check every row against every registered
|
|
94
|
+
* adapter at shift start. Until then those faults surface at the vendor API on
|
|
95
|
+
* the first order that lands on the bad row.
|
|
96
|
+
*/
|
|
97
|
+
export declare function validateCapabilityModels(map: Readonly<Record<string, unknown>>, ctx?: string): void;
|
|
98
|
+
/** Which row served the order: its exact compound, or the bare name fallback. */
|
|
99
|
+
export type CapabilityMatch = 'exact' | 'bare';
|
|
100
|
+
export interface CapabilityResolution {
|
|
101
|
+
/** The capability the winning row was keyed by (`wise:deep` or `wise`). */
|
|
102
|
+
capability: string;
|
|
103
|
+
match: CapabilityMatch;
|
|
104
|
+
model: string;
|
|
105
|
+
effort: string;
|
|
106
|
+
}
|
|
107
|
+
/**
|
|
108
|
+
* Resolve an order's capabilities against the map: exact compound row first,
|
|
109
|
+
* then the bare name-part row, else `undefined` (the caller REFUSES the order —
|
|
110
|
+
* never a default model).
|
|
111
|
+
*
|
|
112
|
+
* TWO PASSES ACROSS ALL CAPABILITIES, NOT ONE PASS PER CAPABILITY. A step may
|
|
113
|
+
* author several capabilities, and the hub's claim gate is itself exact-first
|
|
114
|
+
* across the set (§4/§5: "lookup mirrors the claim rule"). Resolving capability
|
|
115
|
+
* by capability instead would let a bare row on the FIRST capability beat an
|
|
116
|
+
* exact row on the second — the shift would run a `builder:deep` order at the
|
|
117
|
+
* bare `builder` grade purely because of authoring order.
|
|
118
|
+
*
|
|
119
|
+
* The bare row is what makes a NAME-MATCH FALLBACK order resolvable at all: the
|
|
120
|
+
* hub stamps `wise:deep` on an order that a crew bound only to `wise:standard`
|
|
121
|
+
* then claims, and that shift has no `wise:deep` row by construction.
|
|
122
|
+
*/
|
|
123
|
+
export declare function resolveCapabilityModel(map: CapabilityModelMap, capabilities: readonly string[]): CapabilityResolution | undefined;
|
|
@@ -0,0 +1,148 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* Resolve an order's COMPOSED CAPABILITY to the model and reasoning effort this
|
|
3
|
+
* machine runs it at.
|
|
4
|
+
*
|
|
5
|
+
* The hub decides WHAT grade of work an order is (`builder` composed with the
|
|
6
|
+
* run's `deep` modifier → `builder:deep`). It never decides which model serves
|
|
7
|
+
* that grade. That decision is local to the operator's machine, because it is a
|
|
8
|
+
* statement about accounts, quotas and taste — three things a shared workflow
|
|
9
|
+
* definition has no business asserting.
|
|
10
|
+
*
|
|
11
|
+
* So the settings file carries a flat map from capability to `{model, effort}`:
|
|
12
|
+
*
|
|
13
|
+
* ```json
|
|
14
|
+
* {
|
|
15
|
+
* "capabilityModels": {
|
|
16
|
+
* "wise:deep": { "model": "<model-id>", "effort": "xhigh" },
|
|
17
|
+
* "build:deep": { "model": "<model-id>", "effort": "xhigh" },
|
|
18
|
+
* "wise": { "model": "<model-id>", "effort": "high" }
|
|
19
|
+
* }
|
|
20
|
+
* }
|
|
21
|
+
* ```
|
|
22
|
+
*
|
|
23
|
+
* `<model-id>` is whatever string the shift's harness accepts as a model. This
|
|
24
|
+
* module never enumerates real model ids: it ships as neutral runtime, and
|
|
25
|
+
* `test/vendor-gate.test.ts` forbids a harness vendor's names here. `README.md`
|
|
26
|
+
* and `docs/` carry worked examples with real ids.
|
|
27
|
+
*
|
|
28
|
+
* THIS REPLACES `tierMap`/`tierProfiles` OUTRIGHT. A tier was an abstract rung
|
|
29
|
+
* (`fast`/`standard`/`strong`/`strongest`) that a machine then mapped to a
|
|
30
|
+
* model, and the tier ladder could not express the thing the operator actually
|
|
31
|
+
* wanted: "Luna at low effort" and "Fable at max effort" are different routes,
|
|
32
|
+
* but the ladder welded model choice and depth into one ordered axis, so one
|
|
33
|
+
* could not vary without the other. A capability is just a name. What serves it
|
|
34
|
+
* is a local lookup, and depth rides on the capability's modifier suffix rather
|
|
35
|
+
* than on a rung.
|
|
36
|
+
*
|
|
37
|
+
* The one behavior worth carrying out of the tier code is EFFORT VALIDITY
|
|
38
|
+
* CHECKING, which now happens at settings load rather than at resolution time.
|
|
39
|
+
* It checks the effort against `EFFORT_LADDER` — the neutral start contract's
|
|
40
|
+
* own five rungs — and nothing else.
|
|
41
|
+
*
|
|
42
|
+
* IT DELIBERATELY DOES NOT CHECK EFFORT AGAINST THE MODEL. Doing that would
|
|
43
|
+
* require a table of real model ids in this file, which the vendor gate forbids
|
|
44
|
+
* and which this repository has no ground truth for anyway. Look at what the
|
|
45
|
+
* harness adapters in `src/harness/` actually enforce: one validates effort
|
|
46
|
+
* against a closed union that is a property of THE HARNESS — every model that
|
|
47
|
+
* harness runs accepts the same rungs — and the other passes effort through
|
|
48
|
+
* without validating it at all. Neither knows a per-model constraint, so a
|
|
49
|
+
* table here would be either a restatement of `EFFORT_LADDER` under
|
|
50
|
+
* model-shaped keys, or an invention that blocks a valid operator config. The
|
|
51
|
+
* real per-harness check belongs on the adapter contract; see the follow-up
|
|
52
|
+
* noted in `validateCapabilityModels`.
|
|
53
|
+
*
|
|
54
|
+
* The old snap-up/ceiling-refusal logic in `resolveEffort` is gone with the
|
|
55
|
+
* ladder it resolved through: there is no abstract rung left to snap.
|
|
56
|
+
*/
|
|
57
|
+
/** Reasoning rungs the neutral start contract accepts, weakest to strongest. */
|
|
58
|
+
export const EFFORT_LADDER = ['low', 'medium', 'high', 'xhigh', 'max'];
|
|
59
|
+
/**
|
|
60
|
+
* What separates a capability's NAME PART from its MODIFIER: `wise:deep` is the
|
|
61
|
+
* name `wise` at the modifier `deep`. Split on the FIRST separator, matching
|
|
62
|
+
* the engine's own `capabilityName` — a capability name may not contain `:`
|
|
63
|
+
* (install-time rule), so anything after the first one is modifier territory.
|
|
64
|
+
*/
|
|
65
|
+
export const MODIFIER_SEPARATOR = ':';
|
|
66
|
+
/** The bare capability name inside a composed capability. `wise:deep` → `wise`. */
|
|
67
|
+
export function capabilityNamePart(capability) {
|
|
68
|
+
const at = capability.indexOf(MODIFIER_SEPARATOR);
|
|
69
|
+
return at === -1 ? capability : capability.slice(0, at);
|
|
70
|
+
}
|
|
71
|
+
/** Thrown when the capability map is unusable. Never silently repaired. */
|
|
72
|
+
export class CapabilityModelError extends Error {
|
|
73
|
+
}
|
|
74
|
+
function isEffort(value) {
|
|
75
|
+
return EFFORT_LADDER.includes(value);
|
|
76
|
+
}
|
|
77
|
+
/**
|
|
78
|
+
* Validate the map, throwing on the first unusable row. Returns nothing: every
|
|
79
|
+
* fault this function can see is certainly wrong.
|
|
80
|
+
*
|
|
81
|
+
* WHAT IT CATCHES — a capability key that is empty, a row that is not an object,
|
|
82
|
+
* a missing or empty `model`, and an `effort` that is not one of
|
|
83
|
+
* `EFFORT_LADDER`. Each is wrong under every harness and every model, and a
|
|
84
|
+
* shift that started anyway would refuse or misroute its first matching order
|
|
85
|
+
* hours later, far from the typo.
|
|
86
|
+
*
|
|
87
|
+
* WHAT IT DOES NOT CATCH — an effort a particular MODEL or HARNESS rejects, and
|
|
88
|
+
* a model id that does not exist. Neither is knowable here (see this module's
|
|
89
|
+
* header). FOLLOW-UP: `HarnessAdapter` is the right owner for the first one —
|
|
90
|
+
* an adapter that validates effort against a closed union could declare that
|
|
91
|
+
* union, and a composition root could check every row against every registered
|
|
92
|
+
* adapter at shift start. Until then those faults surface at the vendor API on
|
|
93
|
+
* the first order that lands on the bad row.
|
|
94
|
+
*/
|
|
95
|
+
export function validateCapabilityModels(map, ctx = 'capabilityModels') {
|
|
96
|
+
for (const [capability, raw] of Object.entries(map)) {
|
|
97
|
+
if (capability.trim() === '') {
|
|
98
|
+
throw new CapabilityModelError(`${ctx}: a capability key may not be empty`);
|
|
99
|
+
}
|
|
100
|
+
if (typeof raw !== 'object' || raw === null || Array.isArray(raw)) {
|
|
101
|
+
throw new CapabilityModelError(`${ctx}['${capability}'] must be an object with 'model' and 'effort', got ${JSON.stringify(raw)}`);
|
|
102
|
+
}
|
|
103
|
+
const row = raw;
|
|
104
|
+
const model = row['model'];
|
|
105
|
+
if (typeof model !== 'string' || model.trim() === '') {
|
|
106
|
+
throw new CapabilityModelError(`${ctx}['${capability}'].model must be a non-empty string`);
|
|
107
|
+
}
|
|
108
|
+
const effort = row['effort'];
|
|
109
|
+
if (typeof effort !== 'string' || !isEffort(effort)) {
|
|
110
|
+
throw new CapabilityModelError(`${ctx}['${capability}'].effort must be one of ${EFFORT_LADDER.join(', ')}, got ${JSON.stringify(effort)}`);
|
|
111
|
+
}
|
|
112
|
+
}
|
|
113
|
+
}
|
|
114
|
+
/**
|
|
115
|
+
* Resolve an order's capabilities against the map: exact compound row first,
|
|
116
|
+
* then the bare name-part row, else `undefined` (the caller REFUSES the order —
|
|
117
|
+
* never a default model).
|
|
118
|
+
*
|
|
119
|
+
* TWO PASSES ACROSS ALL CAPABILITIES, NOT ONE PASS PER CAPABILITY. A step may
|
|
120
|
+
* author several capabilities, and the hub's claim gate is itself exact-first
|
|
121
|
+
* across the set (§4/§5: "lookup mirrors the claim rule"). Resolving capability
|
|
122
|
+
* by capability instead would let a bare row on the FIRST capability beat an
|
|
123
|
+
* exact row on the second — the shift would run a `builder:deep` order at the
|
|
124
|
+
* bare `builder` grade purely because of authoring order.
|
|
125
|
+
*
|
|
126
|
+
* The bare row is what makes a NAME-MATCH FALLBACK order resolvable at all: the
|
|
127
|
+
* hub stamps `wise:deep` on an order that a crew bound only to `wise:standard`
|
|
128
|
+
* then claims, and that shift has no `wise:deep` row by construction.
|
|
129
|
+
*/
|
|
130
|
+
export function resolveCapabilityModel(map, capabilities) {
|
|
131
|
+
for (const capability of capabilities) {
|
|
132
|
+
const row = map[capability];
|
|
133
|
+
if (row !== undefined)
|
|
134
|
+
return { capability, match: 'exact', model: row.model, effort: row.effort };
|
|
135
|
+
}
|
|
136
|
+
for (const capability of capabilities) {
|
|
137
|
+
const name = capabilityNamePart(capability);
|
|
138
|
+
// Skip a capability that IS its own name part — the first pass already
|
|
139
|
+
// tried that key, and reporting it as a `bare` match would misreport an
|
|
140
|
+
// exact hit as a fallback in the resolution record.
|
|
141
|
+
if (name === capability)
|
|
142
|
+
continue;
|
|
143
|
+
const row = map[name];
|
|
144
|
+
if (row !== undefined)
|
|
145
|
+
return { capability: name, match: 'bare', model: row.model, effort: row.effort };
|
|
146
|
+
}
|
|
147
|
+
return undefined;
|
|
148
|
+
}
|
|
@@ -1,19 +1,21 @@
|
|
|
1
1
|
import type { HubClient } from '../hub/client.ts';
|
|
2
2
|
import type { ContactHolder, OrderPacket } from '../hub/types.ts';
|
|
3
|
+
import type { ConsumedVerifier } from '../consumed-verifier.ts';
|
|
3
4
|
import type { NormalizedStepSpec } from '../bundle/types.ts';
|
|
4
5
|
import type { HarnessAdapter } from '../harness/contract.ts';
|
|
5
6
|
import { type SessionRecord } from '../harness/session-store.ts';
|
|
7
|
+
import { type CapabilityModelMap, type CapabilityResolution } from './capability-model.ts';
|
|
6
8
|
/** Discriminated result of one worker life; the role maps these to exit codes. */
|
|
7
|
-
export type AgentRunOutcome = 'submitted' | 'completed' | 'misroute' | 'no-template' | 'no-harness' | 'no-submit' | 'killed' | 'lease-lost' | 'ownership-error' | 'hub-unreachable' | 'stopped';
|
|
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';
|
|
8
10
|
/**
|
|
9
11
|
* Locate the normalized step spec for an order. Injected because the lookup is
|
|
10
|
-
* I/O (the
|
|
11
|
-
*
|
|
12
|
-
* the
|
|
12
|
+
* I/O (the verified workflow store) and because resolving the local definition
|
|
13
|
+
* is the role's business, not the loop's. `null` means instruction resolution
|
|
14
|
+
* refused; the loop releases and leaves the order for the pickup window.
|
|
13
15
|
*
|
|
14
|
-
* The spec arrives
|
|
15
|
-
* `StepPermissions`.
|
|
16
|
-
* reaches into a step
|
|
16
|
+
* The spec arrives already normalized from the verified `StepDef` — brief,
|
|
17
|
+
* harness id, and `StepPermissions`. The loop stays vendor-neutral: it never
|
|
18
|
+
* reaches into a step definition's extension bag, so there is no key to name.
|
|
17
19
|
*/
|
|
18
20
|
export type StepLoader = (order: OrderPacket) => Promise<NormalizedStepSpec | null>;
|
|
19
21
|
/**
|
|
@@ -52,6 +54,20 @@ export interface AgentRunLoopOptions {
|
|
|
52
54
|
cwd: string;
|
|
53
55
|
loadStep: StepLoader;
|
|
54
56
|
resolveAdapter: AdapterResolver;
|
|
57
|
+
/**
|
|
58
|
+
* THIS MACHINE'S capability → `{model, effort}` map, straight from the
|
|
59
|
+
* settings file (plan §5). The loop resolves the order's composed capabilities
|
|
60
|
+
* against it before dispatch and REFUSES the order when nothing matches —
|
|
61
|
+
* there is no default model here, by design.
|
|
62
|
+
*
|
|
63
|
+
* Absent (or empty) means the shift declared no map at all, which makes every
|
|
64
|
+
* capability-bearing order unresolvable. That is deliberate: a shift that
|
|
65
|
+
* forgot its map should say so on the first order, not run everything on a
|
|
66
|
+
* hardcoded model.
|
|
67
|
+
*/
|
|
68
|
+
capabilityModels?: CapabilityModelMap;
|
|
69
|
+
/** Gate dynamic values and rejection reasons before any prompt rendering. */
|
|
70
|
+
consumedVerifier?: ConsumedVerifier;
|
|
55
71
|
/** Append one session record. Wired to `appendSession` by the role. */
|
|
56
72
|
appendSession: (rec: SessionRecord) => void;
|
|
57
73
|
/** The attempt number for this `(workflow, run, step)`. Wired to `latestFor`. */
|
|
@@ -113,6 +129,39 @@ export interface ConfirmOptions {
|
|
|
113
129
|
*/
|
|
114
130
|
cancelled?: () => boolean;
|
|
115
131
|
}
|
|
132
|
+
/**
|
|
133
|
+
* How this order's model was decided.
|
|
134
|
+
*
|
|
135
|
+
* `resolved` — a settings row matched; `resolution` names it, plus the model
|
|
136
|
+
* and effort the harness starts with.
|
|
137
|
+
* `refused` — the order named capabilities and NONE of them, exact or bare,
|
|
138
|
+
* has a row. The order is handed back; nothing launches. Plan
|
|
139
|
+
* §5: "never a silent default model".
|
|
140
|
+
* `unrouted` — the order named NO capabilities at all, so there is nothing to
|
|
141
|
+
* look up. This is not a failure: a def that declares no
|
|
142
|
+
* `modifiers:` may author capability-silent steps, and those
|
|
143
|
+
* have always run at the harness's own default model. The loop
|
|
144
|
+
* passes no model and no effort, exactly as it did for a step
|
|
145
|
+
* with no authored tier before this scheme existed.
|
|
146
|
+
*/
|
|
147
|
+
export type OrderRouting = {
|
|
148
|
+
kind: 'resolved';
|
|
149
|
+
resolution: CapabilityResolution;
|
|
150
|
+
} | {
|
|
151
|
+
kind: 'refused';
|
|
152
|
+
capabilities: readonly string[];
|
|
153
|
+
} | {
|
|
154
|
+
kind: 'unrouted';
|
|
155
|
+
};
|
|
156
|
+
/**
|
|
157
|
+
* Resolve an order's composed capabilities against this machine's map.
|
|
158
|
+
*
|
|
159
|
+
* ONLY `packet.capabilities` is consulted. `packet.model` — the def's authored
|
|
160
|
+
* tier name under the retired scheme — is deliberately ignored: a tier is an
|
|
161
|
+
* abstract rung with no meaning left, and reading it as a model id would put a
|
|
162
|
+
* literal `strong` on the wire to a vendor.
|
|
163
|
+
*/
|
|
164
|
+
export declare function resolveOrderRouting(packet: OrderPacket, map: CapabilityModelMap | undefined): OrderRouting;
|
|
116
165
|
/**
|
|
117
166
|
* Poll `get_order` at `intervalMs` for at most `graceMs`, answering ONE
|
|
118
167
|
* question: did the hub see the agent's `submit`?
|
|
@@ -59,8 +59,10 @@
|
|
|
59
59
|
import { existsSync } from 'node:fs';
|
|
60
60
|
import { createLeaseLoop } from "../lease/loop.js";
|
|
61
61
|
import { isResumeUnavailable } from "../harness/contract.js";
|
|
62
|
+
import { preflightStepPermissions } from "../harness/permissions.js";
|
|
62
63
|
import { orderId } from "../harness/session-store.js";
|
|
63
64
|
import { buildOwenloopMcp, renderBrief, renderRejection, renderReplayBrief, } from "./brief.js";
|
|
65
|
+
import { resolveCapabilityModel, } from "./capability-model.js";
|
|
64
66
|
/** Default confirm-phase cadence: one `get_order` per second. */
|
|
65
67
|
export const DEFAULT_CONFIRM_INTERVAL_MS = 1_000;
|
|
66
68
|
/** Default confirm-phase bound: give the hub 15s to show the submit's outcome. */
|
|
@@ -68,6 +70,21 @@ export const DEFAULT_SUBMIT_GRACE_MS = 15_000;
|
|
|
68
70
|
function errMsg(e) {
|
|
69
71
|
return e instanceof Error ? e.message : String(e);
|
|
70
72
|
}
|
|
73
|
+
/**
|
|
74
|
+
* Resolve an order's composed capabilities against this machine's map.
|
|
75
|
+
*
|
|
76
|
+
* ONLY `packet.capabilities` is consulted. `packet.model` — the def's authored
|
|
77
|
+
* tier name under the retired scheme — is deliberately ignored: a tier is an
|
|
78
|
+
* abstract rung with no meaning left, and reading it as a model id would put a
|
|
79
|
+
* literal `strong` on the wire to a vendor.
|
|
80
|
+
*/
|
|
81
|
+
export function resolveOrderRouting(packet, map) {
|
|
82
|
+
const capabilities = packet.capabilities ?? [];
|
|
83
|
+
if (capabilities.length === 0)
|
|
84
|
+
return { kind: 'unrouted' };
|
|
85
|
+
const resolution = map === undefined ? undefined : resolveCapabilityModel(map, capabilities);
|
|
86
|
+
return resolution === undefined ? { kind: 'refused', capabilities } : { kind: 'resolved', resolution };
|
|
87
|
+
}
|
|
71
88
|
/**
|
|
72
89
|
* Poll `get_order` at `intervalMs` for at most `graceMs`, answering ONE
|
|
73
90
|
* question: did the hub see the agent's `submit`?
|
|
@@ -128,6 +145,8 @@ export function createAgentRunLoop(opts) {
|
|
|
128
145
|
* `SessionRecord.deliveredReasonAt`. Carried forward from the prior record
|
|
129
146
|
* when this firing delivered no new reasons, so it can never regress. */
|
|
130
147
|
let deliveredReasonAt;
|
|
148
|
+
/** Captures a thrown safety-gate append from the synchronous `started` event. */
|
|
149
|
+
let activePersistenceFailure;
|
|
131
150
|
let resolveOrder;
|
|
132
151
|
const orderReady = new Promise((r) => {
|
|
133
152
|
resolveOrder = r;
|
|
@@ -230,9 +249,17 @@ export function createAgentRunLoop(opts) {
|
|
|
230
249
|
case 'started':
|
|
231
250
|
sessionRef = e.ref;
|
|
232
251
|
opts.err(`owenloop work agent-run: session started for ${order} (harness '${e.ref.harness}')`);
|
|
233
|
-
//
|
|
234
|
-
//
|
|
235
|
-
|
|
252
|
+
// This synchronous event is the cold-start gate. The adapter must not
|
|
253
|
+
// begin provider work until the callback returns. A thrown durable append
|
|
254
|
+
// therefore aborts the turn before delivery and is handled separately
|
|
255
|
+
// from an ordinary harness failure below.
|
|
256
|
+
try {
|
|
257
|
+
record('active');
|
|
258
|
+
}
|
|
259
|
+
catch (error) {
|
|
260
|
+
activePersistenceFailure = error;
|
|
261
|
+
throw error;
|
|
262
|
+
}
|
|
236
263
|
return;
|
|
237
264
|
case 'progress':
|
|
238
265
|
opts.err(`owenloop work agent-run: ${e.text}`);
|
|
@@ -248,6 +275,40 @@ export function createAgentRunLoop(opts) {
|
|
|
248
275
|
return;
|
|
249
276
|
}
|
|
250
277
|
}
|
|
278
|
+
/**
|
|
279
|
+
* Tell the hub what this shift resolved the order to (plan §6), before the
|
|
280
|
+
* harness launches.
|
|
281
|
+
*
|
|
282
|
+
* BEST EFFORT, ALWAYS. A hub that is unreachable, older than the verb, or
|
|
283
|
+
* simply erroring must never be able to stop work the shift is otherwise
|
|
284
|
+
* ready to do — the report is observability, not a gate, and the hub's own
|
|
285
|
+
* verb says so. A failure is logged and swallowed.
|
|
286
|
+
*
|
|
287
|
+
* An `unrouted` order is not reported at all: the payload's `capability` is
|
|
288
|
+
* required and non-empty on the hub side, and there is no capability to name.
|
|
289
|
+
*/
|
|
290
|
+
async function reportResolution(routing, harness) {
|
|
291
|
+
if (routing.kind === 'unrouted')
|
|
292
|
+
return;
|
|
293
|
+
const payload = routing.kind === 'resolved'
|
|
294
|
+
? {
|
|
295
|
+
capability: routing.resolution.capability,
|
|
296
|
+
match: routing.resolution.match,
|
|
297
|
+
model: routing.resolution.model,
|
|
298
|
+
effort: routing.resolution.effort,
|
|
299
|
+
harness,
|
|
300
|
+
}
|
|
301
|
+
: // The refusal names the capability the order was OFFERED under — the
|
|
302
|
+
// first one — because that is the row an operator has to add. Model
|
|
303
|
+
// and effort are absent: nothing was selected.
|
|
304
|
+
{ capability: routing.capabilities[0] ?? '', match: 'refused', harness };
|
|
305
|
+
try {
|
|
306
|
+
await opts.hub.reportResolution({ workflow, run: runId, resolution: payload });
|
|
307
|
+
}
|
|
308
|
+
catch (e) {
|
|
309
|
+
opts.err(`owenloop work agent-run: reporting the resolution for ${order} failed (continuing): ${errMsg(e)}`);
|
|
310
|
+
}
|
|
311
|
+
}
|
|
251
312
|
/** Release and finish. Used by every path that hands the order back. */
|
|
252
313
|
async function releaseWith(reason, outcome) {
|
|
253
314
|
lease.stop(reason);
|
|
@@ -267,11 +328,39 @@ export function createAgentRunLoop(opts) {
|
|
|
267
328
|
]);
|
|
268
329
|
if (first.t === 'lease')
|
|
269
330
|
return mapNoHold(first.o);
|
|
270
|
-
|
|
271
|
-
if (packet === null || packet.
|
|
331
|
+
let packet = first.res.order;
|
|
332
|
+
if (packet === null || packet.worker === 'command') {
|
|
272
333
|
opts.err(`owenloop work agent-run: ${order} is not an agent order (misroute) — releasing`);
|
|
273
334
|
return releaseWith('misroute', 'misroute');
|
|
274
335
|
}
|
|
336
|
+
// Consume-side verification is the prompt boundary. Do this before loading
|
|
337
|
+
// or rendering any step brief: `owes[].reasons` is dynamic text supplied by
|
|
338
|
+
// the transport, and an unverified rejection thread must never reach a
|
|
339
|
+
// provider session. The same whole-order gate also protects `consumes`.
|
|
340
|
+
const hasConsumedData = Object.keys(packet.consumes).length > 0
|
|
341
|
+
|| packet.owes.some((owed) => owed.reasons.length > 0 || owed.proof !== undefined);
|
|
342
|
+
if (hasConsumedData) {
|
|
343
|
+
if (opts.consumedVerifier === undefined) {
|
|
344
|
+
const detail = `consume-side verifier is not configured; dynamic values cannot be admitted to an agent prompt`;
|
|
345
|
+
opts.err(`owenloop work agent-run: consumed artifact refusal (prerequisite) for ${order}: ${detail}`);
|
|
346
|
+
return releaseWith('unverified-consumed', 'unverified-consumed');
|
|
347
|
+
}
|
|
348
|
+
try {
|
|
349
|
+
const checked = await opts.consumedVerifier(packet, { hardRule: false });
|
|
350
|
+
if (!checked.ok) {
|
|
351
|
+
opts.err(`owenloop work agent-run: ${checked.reason} — releasing ${order}`);
|
|
352
|
+
return releaseWith('unverified-consumed', 'unverified-consumed');
|
|
353
|
+
}
|
|
354
|
+
packet = checked.order;
|
|
355
|
+
for (const warning of checked.warnings)
|
|
356
|
+
opts.err(warning);
|
|
357
|
+
}
|
|
358
|
+
catch (error) {
|
|
359
|
+
const detail = error instanceof Error ? error.message : String(error);
|
|
360
|
+
opts.err(`owenloop work agent-run: consumed artifact refusal (prerequisite) for ${order}: ${detail} — releasing`);
|
|
361
|
+
return releaseWith('unverified-consumed', 'unverified-consumed');
|
|
362
|
+
}
|
|
363
|
+
}
|
|
275
364
|
stepName = packet.step;
|
|
276
365
|
recordCwd = packet.workdir ?? opts.cwd;
|
|
277
366
|
// The step spec. No bundle / no spec ⇒ we cannot brief anybody; release so
|
|
@@ -309,12 +398,60 @@ export function createAgentRunLoop(opts) {
|
|
|
309
398
|
origin: opts.origin,
|
|
310
399
|
account: opts.account,
|
|
311
400
|
...(opts.shiftId !== undefined ? { shiftId: opts.shiftId } : {}),
|
|
401
|
+
// Straight from the packet, not from settings: the modifier is the hub's
|
|
402
|
+
// statement about the RUN, and a shift has no say in it.
|
|
403
|
+
...(packet.modifier !== undefined ? { modifier: packet.modifier } : {}),
|
|
404
|
+
...(packet.escalated === true ? { escalated: true } : {}),
|
|
312
405
|
};
|
|
313
406
|
// Permissions arrive PRE-NORMALIZED in the spec. `prepare` ran
|
|
314
407
|
// `normalizeStepPermissions` over the step's `x.harness` options at cache
|
|
315
408
|
// time, so this loop never looks inside an extension bag and never has a bag
|
|
316
409
|
// key — vendor-keyed lookup does not exist in neutral code any more.
|
|
317
410
|
const permissions = step.permissions;
|
|
411
|
+
// Mandatory final-boundary preflight. Common checks run first, followed by
|
|
412
|
+
// the selected adapter's exact capability check. A refusal starts neither a
|
|
413
|
+
// cold session nor a resumed provider process; the held claim is released.
|
|
414
|
+
const policyIssues = [
|
|
415
|
+
...preflightStepPermissions(permissions),
|
|
416
|
+
...active.preflight(permissions),
|
|
417
|
+
];
|
|
418
|
+
if (policyIssues.length > 0) {
|
|
419
|
+
for (const issue of policyIssues) {
|
|
420
|
+
opts.err(`owenloop work agent-run: harness policy refusal for ${order} on '${resolution.id}'` +
|
|
421
|
+
`${issue.field !== undefined ? ` (${issue.field})` : ''}: ${issue.message}`);
|
|
422
|
+
}
|
|
423
|
+
return releaseWith('incompatible-harness-policy', 'incompatible-harness-policy');
|
|
424
|
+
}
|
|
425
|
+
// ---- ROUTING: which model serves this order, and telling the hub --------
|
|
426
|
+
//
|
|
427
|
+
// DELIBERATELY THE LAST GATE BEFORE DISPATCH, and deliberately AFTER every
|
|
428
|
+
// launch gate above (misroute, consumed verification, step spec, adapter,
|
|
429
|
+
// harness policy). Plan §3: the launch gates block on a malformed order
|
|
430
|
+
// before model selection is even attempted, so a refusal never has to be
|
|
431
|
+
// explained in terms of a model the order was never going to reach.
|
|
432
|
+
const routing = resolveOrderRouting(packet, opts.capabilityModels);
|
|
433
|
+
// Plan §6: the record must exist BEFORE tokens are spent, which is why this
|
|
434
|
+
// is a dedicated verb and not a field on the next heartbeat. It is reported
|
|
435
|
+
// on the refusal path too — a refusal is a true observation of what this
|
|
436
|
+
// shift resolved, and the whole point of the report is that the hub can say
|
|
437
|
+
// why an order did or did not run.
|
|
438
|
+
await reportResolution(routing, resolution.id);
|
|
439
|
+
if (routing.kind === 'refused') {
|
|
440
|
+
opts.err(`owenloop work agent-run: no settings row for ${order} — the order's capabilities ` +
|
|
441
|
+
`[${routing.capabilities.join(', ')}] match neither an exact row nor a bare name row in ` +
|
|
442
|
+
`'capabilityModels'. Add a row for one of them (a bare row such as ` +
|
|
443
|
+
`'${routing.capabilities[0]?.split(':')[0] ?? ''}' covers every modifier) and restart the shift. ` +
|
|
444
|
+
`Releasing — this shift will not pick a model it was not told to use.`);
|
|
445
|
+
return releaseWith('unresolvable-capability', 'unresolvable-capability');
|
|
446
|
+
}
|
|
447
|
+
if (routing.kind === 'unrouted') {
|
|
448
|
+
opts.err(`owenloop work agent-run: ${order} carries no capabilities, so no settings row applies — ` +
|
|
449
|
+
`running on '${resolution.id}''s own default model.`);
|
|
450
|
+
}
|
|
451
|
+
else {
|
|
452
|
+
opts.out(`owenloop work agent-run: ${order} resolved '${routing.resolution.capability}' ` +
|
|
453
|
+
`(${routing.resolution.match} row) → ${routing.resolution.model} at ${routing.resolution.effort}`);
|
|
454
|
+
}
|
|
318
455
|
// ---- PHASE 4: resume or cold replay? -----------------------------------
|
|
319
456
|
//
|
|
320
457
|
// The delta first, because "is there anything new to say?" is one of the
|
|
@@ -376,11 +513,14 @@ export function createAgentRunLoop(opts) {
|
|
|
376
513
|
deliveredReasonAt = delta.deliveredReasonAt;
|
|
377
514
|
}
|
|
378
515
|
const owenloopMcp = buildOwenloopMcp(spec);
|
|
516
|
+
const resolvedModel = routing.kind === 'resolved'
|
|
517
|
+
? { model: routing.resolution.model, effort: routing.resolution.effort }
|
|
518
|
+
: undefined;
|
|
379
519
|
const deliverArgs = {
|
|
380
520
|
cwd: recordCwd,
|
|
381
521
|
owenloopMcp,
|
|
382
522
|
permissions,
|
|
383
|
-
...(
|
|
523
|
+
...(resolvedModel ?? {}),
|
|
384
524
|
};
|
|
385
525
|
/** Built lazily: a cold start after a refused resume needs a FRESH one. */
|
|
386
526
|
const coldArgs = () => ({
|
|
@@ -395,12 +535,13 @@ export function createAgentRunLoop(opts) {
|
|
|
395
535
|
cwd: recordCwd,
|
|
396
536
|
owenloopMcp,
|
|
397
537
|
permissions,
|
|
398
|
-
...(
|
|
538
|
+
...(resolvedModel ?? {}),
|
|
399
539
|
});
|
|
400
540
|
const path = resumable ? 'resume' : delta.message !== '' ? 'cold replay' : 'cold start';
|
|
401
541
|
opts.out(`owenloop work agent-run: hosting ${order} (step '${packet.step}', harness '${resolution.id}', attempt ${attempt}, ${path})`);
|
|
402
542
|
/** Cold-start this firing. Shared by the ordinary path and the fallback. */
|
|
403
543
|
async function coldStart() {
|
|
544
|
+
activePersistenceFailure = undefined;
|
|
404
545
|
try {
|
|
405
546
|
const ref = await active.start(coldArgs(), onEvent);
|
|
406
547
|
sessionRef = ref;
|
|
@@ -413,6 +554,10 @@ export function createAgentRunLoop(opts) {
|
|
|
413
554
|
return { t: 'turn' };
|
|
414
555
|
}
|
|
415
556
|
catch (e) {
|
|
557
|
+
if (activePersistenceFailure !== undefined) {
|
|
558
|
+
await teardown();
|
|
559
|
+
return { t: 'turn', persistenceFailure: activePersistenceFailure };
|
|
560
|
+
}
|
|
416
561
|
return { t: 'turn', failure: e };
|
|
417
562
|
}
|
|
418
563
|
}
|
|
@@ -434,8 +579,15 @@ export function createAgentRunLoop(opts) {
|
|
|
434
579
|
opts.err(`owenloop work agent-run: resuming session ${prev.token} for ${order} with ${delta.count} new rejection reason${delta.count === 1 ? '' : 's'} (no brief re-sent)`);
|
|
435
580
|
// Deliberately written with the PRIOR watermark: nothing has been delivered
|
|
436
581
|
// yet, and this row is what the next firing reads if this turn never
|
|
437
|
-
// completes. See `markDelivered`.
|
|
438
|
-
|
|
582
|
+
// completes. See `markDelivered`. This durable write is also the resume
|
|
583
|
+
// delivery gate; failure stops the old provider session without delivery.
|
|
584
|
+
try {
|
|
585
|
+
record('active');
|
|
586
|
+
}
|
|
587
|
+
catch (error) {
|
|
588
|
+
await teardown();
|
|
589
|
+
return { t: 'turn', persistenceFailure: error };
|
|
590
|
+
}
|
|
439
591
|
try {
|
|
440
592
|
await active.deliver(sessionRef, delta.message, deliverArgs, onEvent);
|
|
441
593
|
markDelivered();
|
|
@@ -473,6 +625,11 @@ export function createAgentRunLoop(opts) {
|
|
|
473
625
|
opts.err(`owenloop work agent-run: lease ${raced.o} mid-turn — tore the session down, no outcome observed`);
|
|
474
626
|
return mapLeaseDuringTurn(raced.o);
|
|
475
627
|
}
|
|
628
|
+
if (raced.persistenceFailure !== undefined) {
|
|
629
|
+
opts.err(`owenloop work agent-run: durable active-session persistence failed before provider delivery for ${order}: ` +
|
|
630
|
+
`${errMsg(raced.persistenceFailure)} — releasing`);
|
|
631
|
+
return releaseWith('session-store-failed', 'session-store-failed');
|
|
632
|
+
}
|
|
476
633
|
// TURN END — NOT task end. Log the failure shape for humans, then confirm.
|
|
477
634
|
if (raced.failure !== undefined) {
|
|
478
635
|
const why = isResumeUnavailable(raced.failure)
|
|
@@ -30,11 +30,13 @@ export declare function readLatestBundle(cacheDir: string, name: string): Cached
|
|
|
30
30
|
* absent, unreadable, or not valid JSON. Layout mirrors `writeBundle`:
|
|
31
31
|
* `<cacheDir>/bundles/<name>/<hash>/steps/<step>.json`.
|
|
32
32
|
*
|
|
33
|
-
*
|
|
34
|
-
*
|
|
35
|
-
*
|
|
36
|
-
* `
|
|
37
|
-
*
|
|
33
|
+
* A missing or corrupt spec returns `null` — a command step has no spec by
|
|
34
|
+
* design, and a hash dir prepared before Phase 5 has none at all. The one
|
|
35
|
+
* deliberate exception is a structurally valid old normalized spec that leaves
|
|
36
|
+
* `filesystem` or `network` buried under `permissions.extensions`: that cache
|
|
37
|
+
* throws with an explicit `owenloop work prepare` repair because returning it
|
|
38
|
+
* would silently discard an authored restriction. There is no fallback to the
|
|
39
|
+
* legacy `templates/<step>.md`.
|
|
38
40
|
*/
|
|
39
41
|
export declare function readStepSpec(cacheDir: string, name: string, hash: string, step: string): NormalizedStepSpec | null;
|
|
40
42
|
/**
|