feat(skill): object-store names for skill inputs and CLI flags; drop two S3-form tombstones

[earakely-scale]

Summary

Skill.object_url replaces Skill.s3_url. It's an object-store prefix like any other.

• Reading stored tasks: stored add_skills steps carry s3_url, so from_dict reads it permanently. It reads the legacy key first, as the artifact twins do.
• Writing: to_dict writes both keys, because a worker still on an older version reads only s3_url.
• Old spelling: the s3_url= keyword still constructs a Skill, warning through warn_deprecated, so its callers are counted. skill.s3_url still reads as object_url. A Skill copied with dataclasses.replace gets the same s3_url value it already has, which isn't counted.
• deploy_agent's skills entries: read object_url, keeping skill_s3_url and s3_uri as read aliases.

CLI.

• New flags: a2a-agent add-skill --skill-object-url, artifact skill put --prefix-url and artifact file-artifact-universe put-bundled --prefix-url.
• --skill-s3-url and --s3-url remain as hidden flags that print their replacement on stderr. Giving both spellings is a usage error.
• Output labels now read skill_object_url=, bundle_object_url= and cli_object_url=.

Skill helpers. parse_skill_md and fetch_skill_md are public. Their private names stay as aliases, since callers outside agent-env import and patch them.

Validator. The object-store skill fixture is renamed to validator-test-object / VALIDATOR-OBJ-99, with criterion secret_code_object. Only secret_code_inline is read back.

Two refusals left over from removing the S3 transfer forms are dropped:

• A stored deploy step with agent_changelog_s3_prefix now loads without it, like the other S3-only fields. No stored task carries one.
• A trajectory get answered with only trajectory_s3_prefix now yields no trajectory instead of a named error. That refusal hasn't fired in 14 days.

Testing

• Unit:
  • A skill named by object_url carries both spellings and doesn't warn.
  • The s3_url keyword warns, is counted and still works.
  • Conflicting spellings raise TypeError; a copy doesn't.
  • Stored skills load keyed legacy, neutral, or both (legacy wins).
  • An inline skill reads s3_url as None.
  • deploy_agent's skills dict negotiates a bundle from object_url, skill_s3_url or s3_uri.
  • put-bundled and skill put work under either flag, with a stderr notice for the old one; giving both is exit 2.
  • The helper aliases are the public functions.
  • The tests that pinned the two refusals now pin the new behaviour.
  • Suite: 6175 passed, 13 skipped.
• Mutants (out-of-tree): 12 of 12 caught. Each of these fails a test:
  • to_dict writing one key;
  • from_dict reading neutral-first, or legacy-only;
  • the keyword not warning;
  • copies conflicting;
  • dropping the s3_url property;
  • deploy_agent not reading object_url;
  • the old flag not noting its replacement;
  • both flags accepted;
  • dropping a private alias;
  • restoring either refusal.
• Dev stage (Mongo):
  • A task whose add_skills step names a skill by object_url is stored with both keys, equal.
  • 0.9.1282 loads it through s3_url.
  • This branch loads it, and also a stored dev task whose skill has only s3_url.

Labelled bump-version.

🤖 Generated with Claude Code

Confidence Score: 3/5

The PR is not ready to merge until copied skills can change URLs and mixed deploy skill entries keep their valid URL.

Findings

1.  Copied skill URL cannot change ▶
2.  Legacy skill URL gets dropped ▶
3.  Old helper patches stop working ▶

Fix with agent prompt

### Issue 1
src/agent_env/task_step/task_steps/add_skills.py:47-50
For a `Skill` with an existing `object_url`, `dataclasses.replace(skill, object_url=new_url)` also passes the old URL through `s3_url`. This check only handles an unchanged copy, so `renamed_keyword` raises `TypeError` instead of creating the updated skill. It also treats an explicit `s3_url=None` as omitted, so that use is not counted. Separate copies from explicit uses of the old keyword.

### Issue 2
src/agent_env/task_step/task_steps/deploy_agent.py:79-80
If a deploy skill has `object_url: None` alongside a valid `skill_s3_url` or `s3_uri`, `dict.get` picks the null value. `register_skill` then sends only the name and description, not the skill bundle. Fall back to the older URL when `object_url` is null.

### Issue 3
src/agent_env/artifact/artifacts/skill.py:75-76
The private names `_parse_skill_md` and `_fetch_skill_md` still import, but skill validation and the CLI now call `parse_skill_md` and `fetch_skill_md` directly. A caller that patches either old name gets the real helper instead. Keep those patches working, or make the change clear and test it.

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.

Summary

Skills and CLI commands now use object_url names for object-store prefixes while keeping older names usable. Stored skills keep both key spellings for compatibility, and two refusals for old S3-form data are removed.

• Skills accept object_url while old saved names still work.
• Commands use object-store names for skill and bundle prefixes.
• A trajectory reply with only an S3 prefix no longer stops the fetch.
• Stored deploy steps load past the old changelog prefix field.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart LR
  A["Skill object_url"] --> B["to_dict: object_url + s3_url"]
  B --> C["Stored add_skills step"]
  C --> D["from_dict: legacy key first"]
  D --> E["A2AAgent.add_skill"]
  F["Deploy skills dict"] --> G["_skill_fields"]
  G --> H["A2AAgent.register_skill"]

Loading

_{Reviews (1) · Last reviewed commit: "feat(skill): object-store names for skil..." · Reviewed by Greptile}

[greptile-apps]

Copied skill URL cannot change

For a Skill with an existing object_url, dataclasses.replace(skill, object_url=new_url) also passes the old URL through s3_url. This check only handles an unchanged copy, so renamed_keyword raises TypeError instead of creating the updated skill. It also treats an explicit s3_url=None as omitted, so that use is not counted. Separate copies from explicit uses of the old keyword.

Prompt To Fix With AI

This is a comment left during a code review.
Path: src/agent_env/task_step/task_steps/add_skills.py
Line: 47-50

Comment:
**Copied skill URL cannot change**

For a `Skill` with an existing `object_url`, `dataclasses.replace(skill, object_url=new_url)` also passes the old URL through `s3_url`. This check only handles an unchanged copy, so `renamed_keyword` raises `TypeError` instead of creating the updated skill. It also treats an explicit `s3_url=None` as omitted, so that use is not counted. Separate copies from explicit uses of the old keyword.

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.

[greptile-apps]

Legacy skill URL gets dropped

If a deploy skill has object_url: None alongside a valid skill_s3_url or s3_uri, dict.get picks the null value. register_skill then sends only the name and description, not the skill bundle. Fall back to the older URL when object_url is null.

Prompt To Fix With AI

This is a comment left during a code review.
Path: src/agent_env/task_step/task_steps/deploy_agent.py
Line: 79-80

Comment:
**Legacy skill URL gets dropped**

If a deploy skill has `object_url: None` alongside a valid `skill_s3_url` or `s3_uri`, `dict.get` picks the null value. `register_skill` then sends only the name and description, not the skill bundle. Fall back to the older URL when `object_url` is null.

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.

[greptile-apps]

Old helper patches stop working

The private names _parse_skill_md and _fetch_skill_md still import, but skill validation and the CLI now call parse_skill_md and fetch_skill_md directly. A caller that patches either old name gets the real helper instead. Keep those patches working, or make the change clear and test it.

Prompt To Fix With AI

This is a comment left during a code review.
Path: src/agent_env/artifact/artifacts/skill.py
Line: 75-76

Comment:
**Old helper patches stop working**

The private names `_parse_skill_md` and `_fetch_skill_md` still import, but skill validation and the CLI now call `parse_skill_md` and `fetch_skill_md` directly. A caller that patches either old name gets the real helper instead. Keep those patches working, or make the change clear and test it.

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.