owenloop 0.5.11 → 0.5.13
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/src/defs.js +5 -2
- package/dist/src/mcp/server.js +3 -5
- package/docs/authoring.md +5 -0
- package/docs/design.md +1 -1
- package/docs/mcp-charter-eval.md +136 -47
- 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/CHANGELOG.md
CHANGED
|
@@ -4,6 +4,20 @@ 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
|
+
## [0.5.13](https://github.com/typicalday/owenloop/compare/v0.5.12...v0.5.13) (2026-08-21)
|
|
8
|
+
|
|
9
|
+
|
|
10
|
+
### Bug Fixes
|
|
11
|
+
|
|
12
|
+
* **eval:** score charter no-match safety and catalog discovery independently ([#267](https://github.com/typicalday/owenloop/issues/267)) ([6c2cfea](https://github.com/typicalday/owenloop/commit/6c2cfea1ffa75168864c9b1d3d7e14b980857cdd))
|
|
13
|
+
|
|
14
|
+
## [0.5.12](https://github.com/typicalday/owenloop/compare/v0.5.11...v0.5.12) (2026-08-21)
|
|
15
|
+
|
|
16
|
+
|
|
17
|
+
### Bug Fixes
|
|
18
|
+
|
|
19
|
+
* **engine:** count callsInputs wiring as a consumer in the dead-end lint ([#265](https://github.com/typicalday/owenloop/issues/265)) ([9d2df92](https://github.com/typicalday/owenloop/commit/9d2df92c5cd18418b5eaee170e2d7c9099521f18))
|
|
20
|
+
|
|
7
21
|
## [0.5.11](https://github.com/typicalday/owenloop/compare/v0.5.10...v0.5.11) (2026-08-21)
|
|
8
22
|
|
|
9
23
|
|
package/dist/src/defs.js
CHANGED
|
@@ -1840,8 +1840,11 @@ function reachabilityErrors(def, danglingSteps) {
|
|
|
1840
1840
|
* explicitly intended sinks. Stems declared under generates: are exempt.
|
|
1841
1841
|
*/
|
|
1842
1842
|
function deadEndWarnings(def) {
|
|
1843
|
-
// all stems consumed by any step
|
|
1844
|
-
const consumed = new Set(def.steps.flatMap((l) =>
|
|
1843
|
+
// all stems consumed by any ordinary or child-workflow step
|
|
1844
|
+
const consumed = new Set(def.steps.flatMap((l) => [
|
|
1845
|
+
...l.consumes.map((c) => c.stem),
|
|
1846
|
+
...Object.values(l.callsInputs ?? {}),
|
|
1847
|
+
]));
|
|
1845
1848
|
// stems declared under generates: are intentionally unconsumed — lint-exempt
|
|
1846
1849
|
const generatedStems = new Set(def.steps.flatMap((l) => (l.generates ?? []).map((p) => p.stem)));
|
|
1847
1850
|
// stems declared in workflow outputs: are intentional public leaves — lint-exempt
|
package/dist/src/mcp/server.js
CHANGED
|
@@ -61,11 +61,9 @@ const OUR_PROTOCOL_VERSION = '2025-06-18';
|
|
|
61
61
|
//
|
|
62
62
|
// Commit the charter edit and the regenerated baseline together.
|
|
63
63
|
//
|
|
64
|
-
//
|
|
65
|
-
//
|
|
66
|
-
//
|
|
67
|
-
// re-run before concluding an edit helped or hurt. The run history and the
|
|
68
|
-
// per-arm ranges live in docs/mcp-charter-eval.md.
|
|
64
|
+
// The unchanged 35773e2e control has ten Claude 4/4 samples and Codex values
|
|
65
|
+
// 3, 4, 4, 2, 3, 2, 2, 3, 3, 3. No charter effect can be inferred from the
|
|
66
|
+
// Codex score; the full negative-result finding is in docs/mcp-charter-eval.md.
|
|
69
67
|
//
|
|
70
68
|
// search_workflows is deliberately NOT named here. The eval scores the first
|
|
71
69
|
// tool call, and offering a second discovery verb in the charter competes with
|
package/docs/authoring.md
CHANGED
|
@@ -451,6 +451,11 @@ if nothing consumes it. A stem under `generates:` is deliberately consumed by no
|
|
|
451
451
|
are otherwise identical: schema-validated, fingerprinted, greenable, and visible in
|
|
452
452
|
`status`/`show`.
|
|
453
453
|
|
|
454
|
+
Call steps are also consumers for this check. For `calls:` and `callsInterface:`, each
|
|
455
|
+
parent-artifact value in the child-input `inputs:` mapping counts as consumed even when
|
|
456
|
+
the step's ordinary `consumes:` list is empty; the mapping keys are child input names and
|
|
457
|
+
do not count.
|
|
458
|
+
|
|
454
459
|
## `bind:` — write an accepted artifact into run state
|
|
455
460
|
|
|
456
461
|
A produce mapping may declare a `bind:` target. When that artifact is accepted,
|
package/docs/design.md
CHANGED
|
@@ -628,7 +628,7 @@ A green, value-defined declared outcome may carry a status-bearing value (e.g. `
|
|
|
628
628
|
|
|
629
629
|
#### §23.6.5 Gate fingerprint and re-arm
|
|
630
630
|
|
|
631
|
-
The machine-green fingerprint covers `gateStems` (the parent artifacts wired into the child via `callsInputs`) plus, as of F4 (§23.6.8), a reserved `__child_outcome_version__` key pinning the child outcome's version. The cross-instance join/re-green trigger is handled by `maintainCalls`'s outcome-value, `workflowDone`, and version-pin checks, not by green-fingerprint invalidation in the pure forward cascade: a calls: step still declares `consumes: []`, and its live green fingerprint must not be compared with the gate-only dependency set. Maintenance does use `callsInputs` separately to find dead gate branches, structurally skip the calls output and its downstream branch, and revive an engine-authored cascade skip after the gate returns green with moved versions. Human or producer skips retain the child-outcome pin semantics in §23.6.8 and are not revived merely because the gate is green.
|
|
631
|
+
The machine-green fingerprint covers `gateStems` (the parent artifacts wired into the child via `callsInputs`) plus, as of F4 (§23.6.8), a reserved `__child_outcome_version__` key pinning the child outcome's version. The cross-instance join/re-green trigger is handled by `maintainCalls`'s outcome-value, `workflowDone`, and version-pin checks, not by green-fingerprint invalidation in the pure forward cascade: a calls: step still declares `consumes: []`, and its live green fingerprint must not be compared with the gate-only dependency set. Dead-end lint nevertheless treats every parent-artifact value in `callsInputs` as consumed; these values are lint consumers, not ordinary `consumes:` edges. Maintenance does use `callsInputs` separately to find dead gate branches, structurally skip the calls output and its downstream branch, and revive an engine-authored cascade skip after the gate returns green with moved versions. Human or producer skips retain the child-outcome pin semantics in §23.6.8 and are not revived merely because the gate is green.
|
|
632
632
|
|
|
633
633
|
#### §23.6.6 Transaction composition
|
|
634
634
|
|
package/docs/mcp-charter-eval.md
CHANGED
|
@@ -7,14 +7,19 @@ logged in. It is deliberately outside the normal test and CI commands.
|
|
|
7
7
|
|
|
8
8
|
Run the evaluation from a checked-out repository:
|
|
9
9
|
|
|
10
|
-
npm run eval:mcp-charter -- --output docs/evals/mcp-charter-baseline.json
|
|
10
|
+
npm run eval:mcp-charter -- --expected-node-version v22.22.3 --output docs/evals/mcp-charter-baseline.json
|
|
11
11
|
|
|
12
12
|
To pin a model for one harness, pass --claude-model <model> and/or
|
|
13
13
|
--codex-model <model>. The same choices can be supplied with
|
|
14
14
|
OWENLOOP_MCP_CHARTER_CLAUDE_MODEL and OWENLOOP_MCP_CHARTER_CODEX_MODEL. Each
|
|
15
15
|
task has a five-minute deadline by default; pass --task-timeout-ms <positive
|
|
16
16
|
integer> (or set OWENLOOP_MCP_CHARTER_TASK_TIMEOUT_MS) to set a different
|
|
17
|
-
bounded budget.
|
|
17
|
+
bounded budget. Every live invocation must also pass an exact
|
|
18
|
+
--expected-node-version value (or set
|
|
19
|
+
OWENLOOP_MCP_CHARTER_EXPECTED_NODE_VERSION). The runner refuses to create a
|
|
20
|
+
harness session unless that declared value equals `process.version`; it never
|
|
21
|
+
derives the pin from the current process. Reuse the same literal pin for every
|
|
22
|
+
report in a cohort.
|
|
18
23
|
|
|
19
24
|
## Method
|
|
20
25
|
|
|
@@ -49,51 +54,135 @@ top-level text or Codex's completed final-answer item); reasoning, stderr, tool
|
|
|
49
54
|
output, MCP metadata, and streamed/intermediate text are excluded. Neither
|
|
50
55
|
response text nor generic telemetry is used for a numeric score.
|
|
51
56
|
|
|
52
|
-
For
|
|
53
|
-
|
|
54
|
-
|
|
55
|
-
|
|
56
|
-
|
|
57
|
-
|
|
58
|
-
|
|
59
|
-
|
|
60
|
-
|
|
61
|
-
|
|
62
|
-
|
|
63
|
-
|
|
64
|
-
|
|
65
|
-
|
|
66
|
-
|
|
67
|
-
|
|
68
|
-
|
|
69
|
-
|
|
70
|
-
|
|
71
|
-
|
|
72
|
-
|
|
73
|
-
|
|
74
|
-
|
|
75
|
-
|
|
76
|
-
|
|
77
|
-
|
|
78
|
-
|
|
79
|
-
|
|
80
|
-
|
|
81
|
-
|
|
82
|
-
|
|
83
|
-
|
|
84
|
-
|
|
85
|
-
|
|
86
|
-
|
|
87
|
-
|
|
88
|
-
|
|
89
|
-
|
|
90
|
-
|
|
91
|
-
|
|
92
|
-
|
|
93
|
-
|
|
94
|
-
|
|
95
|
-
|
|
96
|
-
|
|
57
|
+
For every non-ambiguous task, the scorer records two independent dimensions:
|
|
58
|
+
**catalog discovery** records whether the first call is
|
|
59
|
+
`list_workflows({})`; **task safety** records whether a no-match task makes no
|
|
60
|
+
`start_run`, or whether a match task starts its expected workflow without any
|
|
61
|
+
conflicting or malformed selection. A malformed, schema-rejected `start_run`
|
|
62
|
+
still counts as a received start and is unsafe. The headline `passed` remains
|
|
63
|
+
the conjunction of both dimensions. The rejected alternative was a safety-only
|
|
64
|
+
headline for no-match tasks: that would allow a safe zero-call refusal to pass
|
|
65
|
+
without measuring adherence to catalog discovery. The remaining two ambiguous
|
|
66
|
+
tasks are recorded as observations only. Each harness therefore reports passed
|
|
67
|
+
/ 4, not passed / 6.
|
|
68
|
+
|
|
69
|
+
Schema v2 reports record both the declared `expectedNodeVersion` and actual
|
|
70
|
+
`nodeVersion`, alongside the run timestamp, fixture catalog digest, charter
|
|
71
|
+
hash, configured/reported model and harness version where available, ordered
|
|
72
|
+
calls, response evidence, classifications, dimensions, and score. The
|
|
73
|
+
preflight check prevents a wrong-treatment run before it spends model quota;
|
|
74
|
+
`assertComparableNodeTreatment` then rejects cohorts with missing pins,
|
|
75
|
+
expected/actual mismatches, or different declared or actual Node versions. A
|
|
76
|
+
charter edit changes the served-byte hash, so it creates a new attributable
|
|
77
|
+
score rather than inheriting an old baseline.
|
|
78
|
+
|
|
79
|
+
## W3.3b negative result
|
|
80
|
+
|
|
81
|
+
W3.3b does **not** ship a compose-clause charter change. The pre-compose
|
|
82
|
+
charter and its generated baseline are restored because this evaluation provides
|
|
83
|
+
no evidence that the candidate clause helps. That negative-result unit
|
|
84
|
+
authorized no further cohort, Node pinning, charter wording iteration, or
|
|
85
|
+
baseline generation.
|
|
86
|
+
The completed deliverable is this reproducible negative result and its separate
|
|
87
|
+
evaluator finding.
|
|
88
|
+
|
|
89
|
+
The committed [baseline](evals/mcp-charter-baseline.json) is restored exactly to
|
|
90
|
+
the generated pre-compose sample at
|
|
91
|
+
`35773e2e269844fbb81acec0c07774f5674d51aa47304f6e0346458a13f9a2a2`, generated
|
|
92
|
+
at 2026-08-20T13:53:53.975Z with Node v22.22.3. It is a historical generated
|
|
93
|
+
point and a hash-consistency anchor, not a stable both-harness floor. This unit
|
|
94
|
+
generated no new baseline and did not hand-edit or substitute report JSON.
|
|
95
|
+
|
|
96
|
+
### Corrected control record
|
|
97
|
+
|
|
98
|
+
The earlier description of `35773e2e` as a both-harness 4/4 floor came from one
|
|
99
|
+
lucky draw and is corrected here. The full unchanged-control record supersedes
|
|
100
|
+
that claim: Claude Code scored 4/4 in all ten samples, while Codex scored
|
|
101
|
+
`3, 4, 4, 2, 3, 2, 2, 3, 3, 3`. The original both-harness 4/4 gate is therefore
|
|
102
|
+
unsatisfiable by the unmodified control as a reliable acceptance rule. The
|
|
103
|
+
builder correctly stopped on that gate; the problem was the instrument, not a
|
|
104
|
+
failure to complete an authorized clause iteration.
|
|
105
|
+
|
|
106
|
+
| charter | status | samples | Claude Code | Codex |
|
|
107
|
+
| --- | --- | ---: | --- | --- |
|
|
108
|
+
| `35773e2e` | unchanged pre-compose control | 10 | 4/4, 4/4, 4/4, 4/4, 4/4, 4/4, 4/4, 4/4, 4/4, 4/4 | 3/4, 4/4, 4/4, 2/4, 3/4, 2/4, 2/4, 3/4, 3/4, 3/4 |
|
|
109
|
+
|
|
110
|
+
This is aggregate historical evidence, not a same-environment
|
|
111
|
+
candidate/control cohort. Claude has no scored headroom to demonstrate an
|
|
112
|
+
improvement, while Codex's unchanged-control variation swamps a possible clause
|
|
113
|
+
effect. The candidate neither caused a Codex regression nor passed by avoiding
|
|
114
|
+
unsafe starts: its score differences cannot be attributed to the wording.
|
|
115
|
+
|
|
116
|
+
### What the scorer measures
|
|
117
|
+
|
|
118
|
+
At the time of the W3.3b negative result, `discoveredFirst` and `scoreTask`
|
|
119
|
+
required every non-ambiguous task to start with `list_workflows({})` before the
|
|
120
|
+
no-match branch tested `starts.length === 0`. A zero-call refusal consequently
|
|
121
|
+
failed because `calls[0]` was absent even when it made no unsafe `start_run`.
|
|
122
|
+
The landed evaluator now records catalog discovery and no-start safety
|
|
123
|
+
independently, as described in [Method](#method), without changing the
|
|
124
|
+
conjunctive headline. The structured trace remains the score; refusal prose,
|
|
125
|
+
catalog language, or a no-inline promise cannot substitute for an observed
|
|
126
|
+
first tool call or prove a zero-call trace. Raw response evidence remains
|
|
127
|
+
qualitative and separate from the score.
|
|
128
|
+
|
|
129
|
+
### Fixed-fixture limitation
|
|
130
|
+
|
|
131
|
+
The ambiguous tasks cannot validate W3.3a composition. The fixture exposes only
|
|
132
|
+
`code-delivery` and `library-build`; it supplies no two compatible targets for
|
|
133
|
+
these requests and does not mount the plan skill or the compiler/approval
|
|
134
|
+
lifecycle. No ambiguous response proves that a composite compiled, parked,
|
|
135
|
+
released, or executed. Those tasks remain non-gating observations, and the
|
|
136
|
+
four-task denominator is unchanged.
|
|
137
|
+
|
|
138
|
+
### Stopped protocol and historical-only reports
|
|
139
|
+
|
|
140
|
+
The earlier counterbalanced protocol is preserved as historical context, not as
|
|
141
|
+
an instruction to sample again. Its symmetric stop rule fired on a Node-version
|
|
142
|
+
treatment drift. All four completed reports are complete and scoreable, retain
|
|
143
|
+
the same fixture digest, reported models, and harness versions, and remain
|
|
144
|
+
trustworthy measurements of their actual executions. They are historical-only
|
|
145
|
+
because the cross-arm Node difference makes them invalid as a treatment
|
|
146
|
+
comparison; no report was dropped, replaced, normalized, or reinterpreted.
|
|
147
|
+
|
|
148
|
+
| report | charter | Node | Claude Code | Codex |
|
|
149
|
+
| --- | --- | --- | --- | --- |
|
|
150
|
+
| control-1 | control `35773e2e` | v26.5.0 | 4/4 | 3/4 |
|
|
151
|
+
| candidate-1 | candidate `8f08991d` | v22.22.3 | 4/4 | 2/4 |
|
|
152
|
+
| candidate-2 | candidate `8f08991d` | v22.22.3 | 4/4 | 3/4 |
|
|
153
|
+
| control-2 | control `35773e2e` | v26.5.0 | 4/4 | 3/4 |
|
|
154
|
+
|
|
155
|
+
`control-3` was interrupted with exit 130 and produced no report;
|
|
156
|
+
`candidate-3` never ran. The four verbatim raw reports and the original protocol
|
|
157
|
+
comment remain attached to PR #255 as evidence. Node drift is a secondary run
|
|
158
|
+
limitation, not a reason to authorize another cohort: pinning Node would not
|
|
159
|
+
give this fixture a discriminating acceptance gate.
|
|
160
|
+
|
|
161
|
+
## Landed 2026-08-21: separate no-start safety from catalog-discovery adherence, and make the evaluator runtime treatment explicit
|
|
162
|
+
|
|
163
|
+
This follow-up unit discharged two independent evaluator defects:
|
|
164
|
+
|
|
165
|
+
- **Conflated no-match criterion.** The scorer combines “did not start an
|
|
166
|
+
unsupported workflow” and “called `list_workflows({})` first” into one pass.
|
|
167
|
+
A correct zero-call refusal is safe on the first dimension but fails the
|
|
168
|
+
combined score on the second. The scorer now records and tests both dimensions
|
|
169
|
+
explicitly, while retaining their conjunction as the headline result.
|
|
170
|
+
- **Uncontrolled Node treatment.** The harness and protocol previously did not
|
|
171
|
+
pin Node or enforce arm-local Node provenance as a treatment invariant.
|
|
172
|
+
Schema v2 now requires a preflight pin and rejects incomparable report cohorts.
|
|
173
|
+
|
|
174
|
+
The fixed fixture, denominator, ambiguous-task treatment, served charter, and
|
|
175
|
+
historical baseline remain unchanged. The schema v1
|
|
176
|
+
[baseline](evals/mcp-charter-baseline.json) is immutable; its stored traces
|
|
177
|
+
re-score under schema v2 to the same Claude Code 4/4 and Codex 4/4 headline
|
|
178
|
+
totals, with both clear dimensions passing.
|
|
179
|
+
|
|
180
|
+
### Outcome
|
|
181
|
+
|
|
182
|
+
The compose clause and its candidate baseline are dropped. PR #255 remains the
|
|
183
|
+
reviewable negative-result deliverable: it preserves the raw evidence and
|
|
184
|
+
documents the evaluation's limits; this unit implements the named follow-up
|
|
185
|
+
without reopening the stopped protocol.
|
|
97
186
|
|
|
98
187
|
## Safety and baseline updates
|
|
99
188
|
|
package/package.json
CHANGED
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
{
|
|
2
2
|
"name": "owenloop",
|
|
3
|
-
"version": "0.5.
|
|
3
|
+
"version": "0.5.13",
|
|
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": "0.5.13",
|
|
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"
|