put_artifact cannot store JSON-object content: string arg is coerced to dict before validation

[tildesrc]

Summary

put_artifact rejects any content that is a JSON object string. The value is coerced
from a string into a Python dict somewhere in the MCP argument layer before Pydantic
validates it, so the str | None-typed content field then rejects it as the wrong type:

Error executing tool put_artifact: 1 validation error for put_artifactArguments
content
  Input should be a valid string [type=string_type,
  input_value={'started_at': '2026-09-0... {}, 'completed': False}, input_type=dict]
    For further information visit https://errors.pydantic.dev/2.13/v/string_type

Note input_type=dict — by the time validation runs the argument is no longer the string
that was sent. A string-typed field can therefore never accept JSON-object-shaped text.

The handler in src/panopticon/taskservice/mcp.py is itself correct:

async def put_artifact(
    task_id: str, name: str, content: str | None = None, content_base64: str | None = None
) -> str:
    if content is not None and content_base64 is None:
        data = content.encode()

content.encode() is exactly right for a str. The problem is upstream — the caller's
string is parsed into a dict before it ever reaches this signature.

Reproduction

put_artifact(
  task_id = "<any task id>",
  name    = ".babysit-ci-state.json",
  content = '{"started_at": "2026-09-04T17:31:00Z", "completed": false}'
)

Fails every time; any JSON object literal reproduces it.

What was ruled out

• Not name-dependent. The .json extension is irrelevant — the coercion is on content.
• Not a whitespace or parsing edge. Retried with a leading newline and with a trailing
  newline wrapping the object; identical error both times.
• Not a general content failure. The same tool accepted a large multi-line markdown
  string in the same session moments earlier. Only object-shaped content trips it.

Untested but likely affected by the same coercion: JSON array content ([...]). A JSON
scalar (bare quoted string or number) would probably survive, coercing to str/int rather
than dict.

Impact

The babysit-ci skill specifies a cross-turn state artifact named .babysit-ci-state.json
with a documented schema (started_at, head_sha, watch_bash_id, retries, completed).
That artifact is JSON by design, so the skill's own prescribed state-passing mechanism cannot
be used as written — the agent hits this error on the first write and again on every update.
Any skill or workflow that persists structured state through put_artifact hits the same wall.

Encountered while running the github-issue workflow; CI-watch state had to go to a local
/tmp file instead, which does not survive the container and is not readable via the task's
MCP resource URI.

Workarounds

• Wrap the JSON in a markdown fence so the payload is not object-shaped at the top level, and
  strip the fence on read.
• Use the REST endpoint (PUT /tasks/{id}/artifacts/{name} with raw bytes), which bypasses
  the MCP argument layer entirely.
• Base64 the JSON and send it as content_base64 — though that field is meant for binaries.

All three are worse than just writing the string.

Suggested fix

Stop parsing content before validation, or accept str | dict | list and re-serialize
non-string input to JSON text before storing. The first is preferable: artifact content
should round-trip byte-for-byte, and a permissive type would silently reformat whatever the
caller sent (key order, indentation, False vs false).

Worth checking whether the same coercion affects other string-typed MCP tool params here —
anything that might legitimately carry JSON-shaped text.