owenloop 0.5.23 → 0.5.25
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 +14 -0
- package/dist/packages/work/src/harness/claude.d.ts +12 -1
- package/dist/packages/work/src/harness/claude.js +165 -27
- package/dist/packages/work/src/harness/gatekeeper.d.ts +21 -0
- package/dist/packages/work/src/harness/gatekeeper.js +55 -8
- package/dist/src/defs.js +106 -11
- package/dist/src/engine.d.ts +2 -8
- package/dist/src/engine.js +40 -28
- package/dist/src/model.d.ts +12 -2
- package/dist/src/model.js +29 -22
- package/dist/src/order-resolver.js +1 -0
- package/dist/src/types.d.ts +8 -0
- package/docs/agent-runner.md +34 -3
- package/docs/authoring.md +22 -1
- package/docs/design.md +27 -2
- package/package.json +1 -1
- package/plugins/claude-code/plugin/.claude-plugin/plugin.json +1 -1
- package/plugins/claude-code/plugin/.mcp.json +1 -1
- package/plugins/codex/plugins/owenloop/.codex-plugin/plugin.json +1 -1
- package/plugins/codex/plugins/owenloop/.mcp.json +1 -1
package/docs/agent-runner.md
CHANGED
|
@@ -123,16 +123,47 @@ The Claude Code adapter refuses a `read-only` allow-list containing anything out
|
|
|
123
123
|
|
|
124
124
|
## Owenloop control-plane exception
|
|
125
125
|
|
|
126
|
-
`network: owenloop-only` means that the Step Agent may reach only the born-bound Owenloop MCP control plane for the held order. The exception exists so the Step Agent can call `get_order` and `
|
|
126
|
+
`network: owenloop-only` means that the Step Agent may reach only the born-bound Owenloop MCP control plane for the held order. The exception exists so the Step Agent can call `get_order`, `submit`, and `ask`; without those calls the Step Agent cannot inspect, complete, or honestly hold an impossible order.
|
|
127
127
|
|
|
128
|
-
For Claude Code isolation, the adapter sets `settingSources: []`, `strictMcpConfig: true`, disables skills, excludes hooks and external MCP servers, and mounts only the worker-created `owenloop` MCP server. The restricted MCP child positively registers exactly `get_order` and `
|
|
128
|
+
For Claude Code isolation, the adapter sets `settingSources: []`, `strictMcpConfig: true`, disables skills, excludes hooks and external MCP servers, and mounts only the worker-created `owenloop` MCP server. The restricted MCP child positively registers exactly `get_order`, `submit`, and `ask`; the restricted MCP child does not register `reject`. A future work-holder tool therefore does not enter restricted sessions unless the adapter explicitly adds the tool to the positive registration list. The adapter also restricts built-in tools to an audited no-network set. The same option construction runs for cold start and resume.
|
|
129
129
|
|
|
130
|
-
The born-bound Owenloop MCP server is created by the worker from the live workflow, run, origin, account, Shift, and held claim. Workflow extension data cannot replace that mount. During isolation, the adapter adds `mcp__owenloop__get_order` and `
|
|
130
|
+
The born-bound Owenloop MCP server is created by the worker from the live workflow, run, origin, account, Shift, and held claim. Workflow extension data cannot replace that mount. During isolation, the adapter adds `mcp__owenloop__get_order`, `mcp__owenloop__submit`, and `mcp__owenloop__ask` to `allowedTools` so those control calls execute without an unattended permission prompt. `allowedTools` controls permission automation; `allowedTools` does not filter the MCP server's `tools/list` response. The positive registration list on the MCP child is the visibility boundary. The adapter also adds `mcp__owenloop__reject` to `disallowedTools` as defense in depth. A direct deny or an MCP wildcard deny that blocks Owenloop `get_order`, `submit`, or `ask` is refused.
|
|
131
131
|
|
|
132
132
|
Outside Claude Code isolation, the default work-holder MCP child continues to register `get_order`, `submit`, and `reject`.
|
|
133
133
|
|
|
134
134
|
## Security boundary
|
|
135
135
|
|
|
136
|
+
### Claude exact work root
|
|
137
|
+
|
|
138
|
+
A dedicated audit Shift may enable `OWENLOOP_CLAUDE_EXACT_WORKDIR=1` in the
|
|
139
|
+
host environment. It is a host-only setting: it is removed before the Claude
|
|
140
|
+
child is launched and is not carried in prompts, events, provider options, or
|
|
141
|
+
durable session records. Absence leaves existing workflows unchanged; any other
|
|
142
|
+
present value fails configuration before a provider query starts.
|
|
143
|
+
|
|
144
|
+
This profile requires an explicit or derived built-in surface containing only
|
|
145
|
+
`Read`, `Glob`, and `Grep` (an explicit `tools: []` remains valid). It forces
|
|
146
|
+
the same isolated settings/MCP/skills profile described above, rejects an
|
|
147
|
+
incompatible authored surface rather than silently filtering it, and keeps the
|
|
148
|
+
born-bound `get_order`, `submit`, and `ask` controls headless.
|
|
149
|
+
|
|
150
|
+
Every path-bearing built-in is intercepted by a host-owned `PreToolUse` guard
|
|
151
|
+
before the SDK's bare `allowedTools` approval can run. A well-formed lexical
|
|
152
|
+
path inside the exact step directory proceeds headlessly. Outside, malformed,
|
|
153
|
+
or SDK-blocked paths are final fixed-message denials: they never create a human
|
|
154
|
+
tool approval, and the agent must work inside the directory or use `ask` when
|
|
155
|
+
the order is impossible there. The callback retains the same final-deny check
|
|
156
|
+
as defense in depth.
|
|
157
|
+
|
|
158
|
+
For a configured model, both cold and resumed strict turns require the provider
|
|
159
|
+
to confirm the identical literal model name. Missing identity, a different
|
|
160
|
+
model, an unavailable model, or a structured fallback event fails the turn
|
|
161
|
+
instead of accepting a substitute. No fallback model is configured.
|
|
162
|
+
|
|
163
|
+
The boundary is lexical containment only; it does not claim to defeat symlinks
|
|
164
|
+
or special files. Snapshot construction remains responsible for those separate
|
|
165
|
+
checks.
|
|
166
|
+
|
|
136
167
|
`x.harness` permission fields are enforced policy. Adapter preflight must prove that the selected adapter can implement each authored restriction exactly; unsupported policy fails closed.
|
|
137
168
|
|
|
138
169
|
`advisory.tools` is model guidance only. `advisory.tools` does not configure an adapter, does not remove a provider tool, and is not a security boundary.
|
package/docs/authoring.md
CHANGED
|
@@ -566,6 +566,27 @@ capability in the resulting judge list just as it does for an ordinary step
|
|
|
566
566
|
producer nor judge supplies capabilities, the judge remains capability-silent,
|
|
567
567
|
and there is nothing to compose.
|
|
568
568
|
|
|
569
|
+
### Modifier-scoped judges
|
|
570
|
+
|
|
571
|
+
Workflow `modifiers:` still controls routing for every step. A native judge
|
|
572
|
+
may additionally opt into an exact subset of that declared vocabulary with
|
|
573
|
+
`modifiers: [deep]` on the judge entry. Omitting the key is fully compatible:
|
|
574
|
+
that judge is active for every run. A scoped judge is active only when the
|
|
575
|
+
instance's stored modifier exactly matches one listed name; an instance with no
|
|
576
|
+
current modifier activates no scoped judge. Per-step escalation does not change
|
|
577
|
+
this decision.
|
|
578
|
+
|
|
579
|
+
Every scope name must be a declared top-level modifier; empty, duplicate,
|
|
580
|
+
whitespace-containing, `:`-containing, unknown, and unscoped-workflow lists
|
|
581
|
+
are definition errors. A mixed panel is allowed: its unscoped judges remain
|
|
582
|
+
required at every modifier while scoped judges join only on matching runs. An
|
|
583
|
+
inactive judge creates no order, pending approval, receipt, or completion debt,
|
|
584
|
+
but an active judge keeps normal capability composition (for example
|
|
585
|
+
`review-rigor:deep`). If the workflow binds its modifier from an artifact,
|
|
586
|
+
every producer with a scoped judge must consume a path strictly downstream of
|
|
587
|
+
that binding artifact; a sibling output on the binder cannot safely use a
|
|
588
|
+
scoped judge because it could commit before the binding settles.
|
|
589
|
+
|
|
569
590
|
Author `x:` on the producer step, not on individual judge entries. Every native
|
|
570
591
|
judge synthesized from that producer inherits the producer's complete parsed
|
|
571
592
|
`x:` map. Each judge receives an independent deep clone, so runner-side
|
|
@@ -581,7 +602,7 @@ through the normal pipeline, with its own throttles (`cadence:`,
|
|
|
581
602
|
`report`, it lands `submitted` (not `green`) instead — schema-valid, but
|
|
582
603
|
waiting on sign-off. Each judge evaluates it and calls the *same*
|
|
583
604
|
`green`/`reject` verbs you already use, targeted at `report` — no new CLI
|
|
584
|
-
surface. Once every
|
|
605
|
+
surface. Once every active judge has approved the current version, `report`
|
|
585
606
|
goes `green`. A single reject sends it straight to `rejected` and re-arms
|
|
586
607
|
`researcher`; a rebuild starts every judge's ledger fresh, so a sibling
|
|
587
608
|
judge's earlier approval never carries over to a new version.
|
package/docs/design.md
CHANGED
|
@@ -724,7 +724,7 @@ this version yet.
|
|
|
724
724
|
- **Reads as OUTSTANDING for completion** — `OUTSTANDING_STATES = DEBT_STATES
|
|
725
725
|
∪ { submitted }` (§5). A workflow is not `done` while any artifact sits in
|
|
726
726
|
`submitted`, even though the producer itself has no further debt.
|
|
727
|
-
- Artifacts whose `produces` entry declares no `judges:` never enter
|
|
727
|
+
- Artifacts whose `produces` entry declares no active `judges:` never enter
|
|
728
728
|
`submitted` — a plain commit lands `green` exactly as before. This is fully
|
|
729
729
|
backward compatible: no judges declared, zero behavior change.
|
|
730
730
|
|
|
@@ -744,6 +744,8 @@ named `${producerStep}.${producedStem}.judges.${judgeName}`. Each judge step:
|
|
|
744
744
|
- `judges: <judgedStem>` — the marker field that makes it a judge (mirrors
|
|
745
745
|
`calls:`'s marker-field pattern), read by both layers:
|
|
746
746
|
`eligibleFirings`/`applyOutcome` (model.ts) and `green()` (engine.ts).
|
|
747
|
+
- `judgeModifiers?: string[]` — the internal compiled carrier for the
|
|
748
|
+
authored judge's optional scope; it is not a general conditional-step key.
|
|
747
749
|
- The producer step's complete parsed `x:` extension carrier is deep-cloned
|
|
748
750
|
onto every synthesized judge. Judge entries do not author their own `x:`;
|
|
749
751
|
inheritance happens after the producer has been fully parsed. Every judge
|
|
@@ -773,7 +775,7 @@ sign-off ledger, present only while relevant (`undefined` once an artifact is
|
|
|
773
775
|
`green`/`rejected` cleanly, cleared on every reject/retry/fresh-submit).
|
|
774
776
|
|
|
775
777
|
- **Judge approve**: `approvals[judgeName] = artifact.version`. If every
|
|
776
|
-
|
|
778
|
+
active judge name now maps to the artifact's *current* version, the
|
|
777
779
|
artifact transitions `submitted → green`. Otherwise it stays `submitted`
|
|
778
780
|
with a partial ledger.
|
|
779
781
|
- **Judge reject**: any single reject wins immediately —
|
|
@@ -913,6 +915,7 @@ steps:
|
|
|
913
915
|
model: strong # optional, per-judge model
|
|
914
916
|
inputs: true # optional, default false — judge also
|
|
915
917
|
# reads the producer's inputs (question)
|
|
918
|
+
modifiers: [deep] # optional exact workflow-modifier scope
|
|
916
919
|
maxAttempts: 8 # optional, §6 — overrides the step default below
|
|
917
920
|
# just for `report`; absent caps still inherit
|
|
918
921
|
maxAttempts: 5 # producer's cap (default for every produce on this step)
|
|
@@ -933,6 +936,22 @@ steps:
|
|
|
933
936
|
inputs, for criteria that need "what was asked for" as context.
|
|
934
937
|
- `cadence:` / `maxRunsPerDay:` — optional throttles, same meaning as on
|
|
935
938
|
steps; firing is event-driven (on submit), the throttles just cap the rate.
|
|
939
|
+
- `modifiers:` — optional non-empty set of exact workflow modifier names. The
|
|
940
|
+
names must be declared in the workflow's top-level `modifiers:` vocabulary.
|
|
941
|
+
Unscoped judges stay unconditional; a scoped judge is inactive without a
|
|
942
|
+
matching stored instance modifier. `activeJudgesForStem(def, stem, modifier)`
|
|
943
|
+
is the shared derivation used by eligibility, pending status, producer and
|
|
944
|
+
verdict commits, rejection, groups, terminal timing, and model checking.
|
|
945
|
+
Inactive judges create no order, approval ledger slot, routing receipt, or
|
|
946
|
+
completion debt.
|
|
947
|
+
|
|
948
|
+
When a definition binds `modifier` from a produced artifact, scoped-judge
|
|
949
|
+
producers must be strictly downstream of that artifact under the same exact
|
|
950
|
+
consume/produce matching used by routing lint. This rules out the binder itself
|
|
951
|
+
and sibling branches, which might commit before the stored modifier is known.
|
|
952
|
+
Definitions with no modifier bind remain valid: their instance modifier is
|
|
953
|
+
chosen at create time (or absent), and active judge orders retain ordinary
|
|
954
|
+
capability composition.
|
|
936
955
|
|
|
937
956
|
See `examples/workflows/judged-research.yaml` for a runnable end-to-end
|
|
938
957
|
example (mirrors this shape exactly, plus `examples/workflows/judges/rigor.md`
|
|
@@ -1147,6 +1166,12 @@ at `create` time. `CheckOptions.assumeProvided` (library level, when calling
|
|
|
1147
1166
|
library default is unchanged by the CLI change below, and is what the
|
|
1148
1167
|
`modelCheck` unit tests calling it directly still rely on.
|
|
1149
1168
|
|
|
1169
|
+
`CheckOptions.modifier` selects one concrete stored-modifier variant for the
|
|
1170
|
+
pure checker. Call `modelCheck(def, { modifier: 'deep', assumeProvided: true })`
|
|
1171
|
+
for each declared variant; inactive synthesized judge steps are omitted from
|
|
1172
|
+
that variant's dead/unreached accounting, while active judges are checked
|
|
1173
|
+
normally.
|
|
1174
|
+
|
|
1150
1175
|
`owenloop check` (the CLI command), however, now defaults to
|
|
1151
1176
|
`assumeProvided: true` — seedOwed inputs are seeded green by default,
|
|
1152
1177
|
modeling "the operator already ran `provide` at `create`." This dissolves
|
package/package.json
CHANGED
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
{
|
|
2
2
|
"name": "owenloop",
|
|
3
|
-
"version": "0.5.
|
|
3
|
+
"version": "0.5.25",
|
|
4
4
|
"description": "Deterministic rails for agentic workflows. Declare steps and dependencies; the engine guarantees order, redoes what changes invalidate, and stops what keeps failing — instead of hoping the agent follows through.",
|
|
5
5
|
"license": "Apache-2.0",
|
|
6
6
|
"author": "typicalday",
|
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
{
|
|
2
2
|
"name": "owenloop",
|
|
3
|
-
"version": "0.5.
|
|
3
|
+
"version": "0.5.25",
|
|
4
4
|
"description": "Claude Code driver for owenloop — cross-harness shift and conduct skills driven by the owenloop CLI, plus one hub control-plane MCP server backing an author skill and a fallback worker, for conducting and authoring workflows against your hosted owenloop server.",
|
|
5
5
|
"author": {
|
|
6
6
|
"name": "Typical Day"
|