[Docs] Require artifact hygiene and full review for large code changes

.agents/skills/self-review/SKILL.md

@@ -10,6 +10,15 @@ instructions. Record the actual target branch and base/head commits, and review
 the full diff from their merge base plus relevant uncommitted changes. Distinguish
 what is in the PR from local-only work; disclose if the base could not be refreshed.
 
+Apply the [large-code-change requirements](../../../CONTRIBUTING.md#large-code-changes):
+report authored-code and total diff counts separately using the guide's counting
+convention. Above 3,000 changed code lines, verify the contributor's full
+self-review, split rationale, component map/review order, and validation across
+affected components and interfaces before recommending readiness. A quick
+precheck is insufficient; keep the PR draft until the contributor self-review is
+complete. Report missing preparation as a readiness gap; size alone is not a
+correctness finding or a reason to require expensive model/GPU runs.
+
 Check correctness, focused scope, decision-model and front contracts, fallback
 behavior, cancellation/timeouts, and resource cleanup as relevant. Verify tests
 cover the changed behavior, including failure paths and a regression case for a
@@ -23,6 +32,28 @@ This skill prepares a local contributor report. It does not itself authorize
 edits, commits, pushes, external posts, paid model calls, downloads, or changes to
 review status. Use only separately authorized execution resources and budgets.
 
+## Committed artifact hygiene
+
+Apply this check in every review, including quick prechecks. Inspect added and
+changed artifacts in the complete diff, including JSON/JSONL, CSV, logs, reports,
+source/binary hash inventories, and generated media. Classify them by purpose and
+actual consumer, rather than rejecting a file extension.
+
+- Keep necessary configuration, request examples, maintained benchmark inputs,
+  and small deterministic fixtures or reference oracles in the repository's
+  intended locations. Identify the test, tool, or documented workflow that needs
+  each retained artifact, such as labelled agent datasets or replay inputs.
+- Flag one-off run summaries, response dumps, cache statistics, profiler output,
+  agent process notes, and duplicate historical results that have no maintained
+  source-tree role. A link from PR prose or documentation alone does not justify
+  committing generated run output. Report concrete paths and consumers.
+- Preserve raw measurements, failures, and provenance in a durable artifact
+  archive or PR/CI evidence, and link the exact revision or run from the summary.
+  Do not discard evidence to reduce the diff or hide it in a committed archive.
+- When removing redundant output, check its callers, links, and reproduction
+  commands. Keep replay inputs and expected responses intact; verify their
+  hashes and rerun the affected replay or documentation checks.
+
 ## Task evidence
 
 For changes to what an agent can accomplish, show a concrete task as **input →

CONTRIBUTING.md

@@ -60,6 +60,33 @@ week for Python packages and one for the GitHub Actions.
   says how to verify. Include a **Demo / evidence** section; use the [self-review skill](.agents/skills/self-review/SKILL.md)
   and the recording guide below.
 
+Apply [committed artifact hygiene](.agents/skills/self-review/SKILL.md#committed-artifact-hygiene)
+in every review, including quick prechecks. Keep maintained fixtures and replay inputs; preserve raw run
+evidence in durable, reviewer-accessible PR/CI artifacts rather than committing redundant generated output.
+
+### Large code changes
+
+PRs with **more than 3,000 changed lines of authored code** need extra contributor
+attention before requesting review. Count additions plus deletions against the
+PR's merge base in source files, tests, and build or validation scripts. Report
+this count separately from the total diff size; exclude documentation, generated
+output, lockfiles, and static fixtures from the code count, while still reviewing
+those files for relevance and correctness.
+
+- Complete a full self-review of every affected component and its integration
+  boundaries. A quick precheck alone is insufficient; keep the PR in draft until
+  the contributor self-review is complete.
+- Consider splitting independent features, refactors, and cleanup into focused
+  PRs. If the change needs to stay together, explain why in the PR description
+  and provide a component map and suggested review order.
+- Include the code-line count and a validation summary for each affected area in
+  the PR description: commands, results, and unverified behavior with reasons.
+  Cover changed interfaces between components as well as individual components.
+
+Size signals the need for closer review; it is not itself a correctness finding.
+Choose checks based on the changed behavior and risk. Crossing this threshold
+alone does not require GPU benchmarks or other expensive experiments.
+
 ## Add an agent use-case recipe
 
 Use [recipes/README.md](recipes/README.md) and [the template](recipes/TEMPLATE.md) to document a complete task
@@ -126,8 +153,10 @@ Export a readable MP4 with H.264 where possible; aim below 10 MB. GitHub's
 [attachment guide](https://docs.github.com/en/get-started/writing-on-github/working-with-advanced-formatting/attaching-files)
 lists supported formats and current limits. Drag the clip into the PR description's **Demo / evidence** section
 or a PR comment, wait for the upload to finish, and save the resulting link. Uploading makes the file public for
-this public repository. Use attachments for videos; keep sanitized reproduction commands and small result
-records in the repository or a durable reviewer-accessible archive. Verify the uploaded video plays and that
+this public repository. Use attachments for videos; keep sanitized reproduction commands and maintained
+fixtures in the repository, and generated run records in a durable reviewer-accessible archive or PR/CI
+evidence, following [artifact hygiene](.agents/skills/self-review/SKILL.md#committed-artifact-hygiene).
+Verify the uploaded video plays and that
 reviewers can open its linked trace.
 
 Use a caption such as:

s1a/decision_models/served.py

@@ -304,6 +304,7 @@ class ServedLayaModel(DecisionModel):
 
     name = "laya-served"
     deterministic = True
+    bills_input_tokens = False  # the client's own server, not Jev's pricing
 
     def __init__(
         self,

tests/test_decision_models_base.py

@@ -23,13 +23,15 @@
     RuleModel,
     ScriptedModel,
 )
+from s1a.decision_models.served import ServedLayaModel
 
 
 class TestBillingContract(TestCase):
     def test_every_backend_declares_whether_input_tokens_use_jev_pricing(self) -> None:
         expected = {
             JevModel: True,
             LayaModel: False,
+            ServedLayaModel: False,
             CuaS1Model: False,
             RandomModel: False,
             RuleModel: False,

tests/test_tool_models.py

@@ -198,6 +198,7 @@ class ServedStub(DecisionModel):
     """Answers the way the served model does: an identity in ``model``, the server's facts in ``raw``."""
 
     name = "laya-served"
+    bills_input_tokens = False
 
     @property
     def model(self) -> str: