agents-relay 2.0.249 → 2.0.250
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/README.md
CHANGED
|
@@ -5,6 +5,8 @@ Agents Relay is a TypeScript runtime for durable asynchronous agent Jobs driven
|
|
|
5
5
|
V2 uses a read-only local SQLite Job Registry for discovery. The immutable jobId selects the Job Store even when prNumber is zero; lifecycle transitions are atomic across processes. GitHub is **audit-only**: its PR/head/publish events are persisted for reporting, never consulted to start, stop, or complete local Workflows. Local Git delivery and its confirmed merge receipt—not a remote PR MERGED fact—gate Graph completion. Job creation and local Task operations require no GitHub authentication or PR.
|
|
6
6
|
|
|
7
7
|
|
|
8
|
+
Agent frontmatter hints (`skills`, `model`, `reasoning`) are described in [Agent Definition Contract](skills/agents-relay/contracts/agent-contract.md). They are advisory only and do not alter Skill resolution or Router constraints.
|
|
9
|
+
|
|
8
10
|
The canonical V2 architecture contract is
|
|
9
11
|
[DESIGN-PRINCIPLES.md](DESIGN-PRINCIPLES.md). README sections below describe
|
|
10
12
|
the runtime; they do not replace that contract.
|
package/package.json
CHANGED
|
@@ -31,7 +31,7 @@ Read `contracts/task-job-contract.md` before authoring or changing Jobs or Tasks
|
|
|
31
31
|
|
|
32
32
|
Usually let `job create --initial-plan` establish the first Planner round; use `plan` for an existing create-only Job. Direct `task create` is for deliberately orchestrator-authored work. Tasks describe verifiable outcomes, **not shell commands**. Declare exactly one durable result (`task_pr` or `file`), meaningful dependencies, and `storySize`; Relay chooses worker/model routes unless the task truly needs a specific capability.
|
|
33
33
|
|
|
34
|
-
Agent
|
|
34
|
+
Agent `skills` suggests Worker skills; `model`/`reasoning` suggest Router preferences. Neither is mandatory (see `contracts/agent-contract.md`).
|
|
35
35
|
|
|
36
36
|
An assigned **execution Prompt** pins a specific Skill version and supplies Role, Bootstrap, Contracts, Contexts, and Task/Plan identity. Preserve these references. Resolve the matching CLI as `../../dist/cli.js` relative to the pinned Skill directory; do **not** substitute an unrelated globally installed CLI. Before acting, load the exact execution-bound context:
|
|
37
37
|
|
|
@@ -0,0 +1,64 @@
|
|
|
1
|
+
# Agents Relay Agent Definition Contract
|
|
2
|
+
|
|
3
|
+
This is the canonical contract for optional Agent frontmatter hints and their
|
|
4
|
+
effect on execution. It applies to project, shared, and Relay-built-in Agents,
|
|
5
|
+
including merged Agents. Agent role instructions, Task content, and Graph-owned
|
|
6
|
+
workflow/lifecycle behavior remain unchanged.
|
|
7
|
+
|
|
8
|
+
## Optional frontmatter
|
|
9
|
+
|
|
10
|
+
Use simple top-level YAML scalar fields and inline comma-separated lists.
|
|
11
|
+
These fields are **advisory**, never requirements:
|
|
12
|
+
|
|
13
|
+
```yaml
|
|
14
|
+
---
|
|
15
|
+
name: vlog-editor
|
|
16
|
+
extends: shared/vlog-editor
|
|
17
|
+
skills: [markcut, shotcraft]
|
|
18
|
+
model: openai/gpt-6
|
|
19
|
+
reasoning: high
|
|
20
|
+
---
|
|
21
|
+
```
|
|
22
|
+
|
|
23
|
+
| Field | Consumer | Semantics |
|
|
24
|
+
| --- | --- | --- |
|
|
25
|
+
| `skills` | Worker User Assignment (including Planner) | Optional skill suggestions, e.g. `Suggested skills ...: markcut, shotcraft`. Only use relevant, available skills. |
|
|
26
|
+
| `model` | Unified Execution Router decision prompt | Preferred provider/model name, e.g. `openai/gpt-6`, for advisory ranking only. |
|
|
27
|
+
| `reasoning` | Unified Execution Router decision prompt | Preferred reasoning effort, e.g. `high`, for advisory ranking only. |
|
|
28
|
+
|
|
29
|
+
A field may be omitted. `reasoning` can be specified without `model`.
|
|
30
|
+
These optional hints do not auto-install/discover/select/require skills, do not
|
|
31
|
+
make a Task depend on a skill, do not create or enable model candidates, and do
|
|
32
|
+
not override the Task's explicit routing constraints, available routes, resource
|
|
33
|
+
or runtime gates, user routing instructions, or deterministic fallback.
|
|
34
|
+
Unknown/unavailable suggestions are not execution errors.
|
|
35
|
+
|
|
36
|
+
## Resolution and precedence
|
|
37
|
+
|
|
38
|
+
The effective Agent inherits frontmatter metadata using the existing Agent
|
|
39
|
+
resolution and `extends` merge rules; a more specific override wins.
|
|
40
|
+
The selected Agent's `skills` hint appears in the Worker/Planner **user
|
|
41
|
+
prompt**, not as an enforced bootstrap/contract requirement. The selected
|
|
42
|
+
Agent's `model` and `reasoning` hints go to the Router **decision prompt**,
|
|
43
|
+
not to the Worker prompt.
|
|
44
|
+
|
|
45
|
+
User-authored Router `[routing-instruction]` guidance has priority over Agent
|
|
46
|
+
preferences, and authoritative hard constraints still apply before selection.
|
|
47
|
+
Only configured/default or explicitly listed `[choices]` routes are eligible;
|
|
48
|
+
the Agent hint does not widen that choice set. Existing Router selection,
|
|
49
|
+
fallback, review, and lifecycle behavior stays the same.
|
|
50
|
+
|
|
51
|
+
When a field is missing, Relay does not add the corresponding hint. If no safe,
|
|
52
|
+
read-only Agent path can be resolved for routing, Relay skips the optional
|
|
53
|
+
Router advice rather than preparing a Task worktree or blocking a valid route.
|
|
54
|
+
The frontmatter is not durable Job/Task state and does not change CLI or
|
|
55
|
+
SQLite schemas.
|
|
56
|
+
|
|
57
|
+
## Related contracts
|
|
58
|
+
|
|
59
|
+
- [Task / Job Contract](task-job-contract.md) remains authoritative for durable
|
|
60
|
+
Task/Job inputs, dependencies, and lifecycle facts.
|
|
61
|
+
- [Prompt Builder](../../../docs/agent-prompt-builder.md) documents the actual
|
|
62
|
+
Worker User Assignment and the merged Agent snapshot.
|
|
63
|
+
- [Execution Router](../../../docs/worker-model-routers.md) documents choice
|
|
64
|
+
discovery, hard gates, decision priorities, and fallback.
|
|
@@ -17,6 +17,10 @@ When creating a new Job from an active discussion, the orchestrator must consoli
|
|
|
17
17
|
|
|
18
18
|
Use `job create` with standalone objective Markdown. Existing PRs are not imported as Jobs. Use Relay CLI for subsequent Job reads and updates; do not infer Job state or objective from GitHub.
|
|
19
19
|
|
|
20
|
+
## Agent definitions and advisory frontmatter
|
|
21
|
+
|
|
22
|
+
Agent frontmatter `skills`, `model`, and `reasoning` are optional **hints**, not Task requirements. See the canonical [Agent Definition Contract](agent-contract.md) for syntax, inheritance, the separate Worker/Router prompt consumers, and the unchanged hard-gate/Skill resolution semantics. Task producers must not turn those hints into durable Task routing constraints or required Skills.
|
|
23
|
+
|
|
20
24
|
## Task producers
|
|
21
25
|
|
|
22
26
|
Any task producer — orchestrator, planner, or nested task-producing agent — must follow this contract.
|