Fix ColonyChat.register() (two-step migration) + version drift

[ColonistOne]

Two independent defects, one per commit so either can be dropped.

1. ColonyChat.register() is broken, and the suite is already red

It calls ColonyClient.register, which colony-sdk has removed. The pin was colony-sdk>=1.18.0,<2, so every install resolves to a version without it.

Unlike the sibling repos — where a bare MagicMock invented the missing method and hid the break — the three register tests here patched that exact attribute, so patch() raises AttributeError and they fail. The suite fails on a clean checkout of main today (3 failed, 107 passed against colony-sdk 1.32.0).

Nobody saw it because CI last ran 2026-06-09, against an older SDK, and no commit has landed since. A dormant repo's last green tick says nothing about the present.

Found while fixing the same SDK removal in sentinel#20; the sibling fix is colony-agent-template#26.

The migration

Added register_begin() / register_confirm() as the pair callers should prefer. register_begin reserves the handle and returns api_key, claim_token and a convenience key_fingerprint, leaving the account pending — it holds the handle but cannot send or read. register_confirm activates it by proving the key was kept.

Splitting them is the point, and it's specifically why a library differs from a CLI here. sentinel, colony-agent-template and crewai-colony can do begin → persist → read back → confirm internally because they own a config file. colony-chat doesn't own the caller's storage, so collapsing the two halves would confirm before anything was written — defeating the gate entirely. Exposing both is what lets storage sit in between. Same shape pydantic-ai-colony already uses, for the same reason.

begun = ColonyChat.register_begin(handle="...", display_name="...")
secrets_store.put("COLONY_CHAT_API_KEY", begun["api_key"])
saved = secrets_store.get("COLONY_CHAT_API_KEY")   # read it BACK
ColonyChat.register_confirm(claim_token=begun["claim_token"], key_fingerprint=saved[-6:])
client = ColonyChat(api_key=saved)

register() is kept and works again, implemented as both halves back to back, with a docstring that is explicit about giving up that guarantee. Keeping it avoids turning a bug fix into an API removal — and since it was already non-functional, nothing that currently works changes shape. Happy to remove it instead if you'd rather not ship a one-shot at all — that's a genuine design call and I'd take your steer.

Also fixed: base_url now reaches confirm, not just begin. Confirm is a separate HTTP request, so a self-hosted Colony would have activated against production.

Tests

Nine new. register_begin raises ColonyChatError when the response has no api_key or claim_token — returning the partial dict would hand back something that looks like a successful registration and is permanently pending. REGISTER_ALREADY_ACTIVE is tolerated as the documented idempotent guard, paired with a control that a different error still propagates.

Mutation-tested rather than eyeballed: swapping the code check for a bare except ColonyAPIError: return active turns exactly that one control red and nothing else.

2. Version drift — the package has been lying about its own version

Separate commit, unrelated to the above.

pyproject.toml said 0.2.0, colony_chat/_version.py said 0.1.3, nothing compared them. This is already published — verified by downloading the artifact rather than reading the repo:

wheel metadata  Version: 0.2.0
runtime         __version__ = "0.1.3"

Neither direction of this drift announces itself: _version.py ahead makes a tag unpublishable under its own number; pyproject ahead publishes cleanly and then lies. It was the quiet one.

Fixed _version.py to 0.2.0 and added tests/test_version_consistency.py, which compares the two on every push — a unit test rather than a release-workflow step, because by tag time the fix means retagging something already published. test_version_exported no longer hardcodes the literal; two places to update per release is what caused this, and that assertion was the one that got missed.

The guard ships with its own control — a check that reads one source twice, or a regex that matches nothing and compares None to None, passes forever and certifies nothing. Mutation-tested by restoring _version.py to 0.1.3: the guard and both controls fail, then pass again once corrected.

tomllib is 3.11+ and CI runs 3.10, so pyproject is read with a regex anchored to the first version key after [project].

Also

chat.thecolony.cc → chat.thecolony.ai in the description, Homepage, Documentation and docs. Both hosts probed first — root and /skill.md return 200 on each, skill.md identical in size. Not the author email — thecolony.ai publishes no MX record, so mail to the .ai form bounces.

Verification

123 tests pass (110 → 123, and 3 that were failing now pass). ruff check, ruff format --check and mypy colony_chat all clean — I ran the exact CI gates locally.

No version bump or release here — the CHANGELOG entry sits under Unreleased. Releasing is yours to call.

Leaving this for @jackparnell — I don't self-merge in TheColonyCC/*.

🤖 Generated with Claude Code

https://claude.ai/code/session_01TRn9SBFGaxRwZbwRsKNJ7b