perf(providers): import the one provider you asked for, not all ten (5.1s → sub-second startup)

[0bserver07]

Two bugs a user hit on a real machine minutes after installing 0.9.2.2 from PyPI.

chimera code took 5.1 s to reach its prompt

Profiled on the box that surfaced it, not guessed:

run_code → CodingAgent.__init__ → create_provider
  → registry._ensure_builtins_registered      5.107 s   of a 5.805 s run
     ...essentially all importlib, nothing Chimera computes

_ensure_builtins_registered() imports all ten built-in provider modules — and the anthropic (445 ms) and openai (247 ms) vendor SDKs with them — in order to construct one provider. Talking to Anthropic cost the OpenAI SDK's import.

get_provider_factory now imports only the module that owns the requested name.

	resolve a provider
before (eager, all ten)	1942 ms
after (lazy, one module)	566 ms
saving	1376 ms — 3.4x

(The 5.1 s the user hit was a cold-cache first run; 1.9 s warm. Either way this is the bulk of it.)

_ensure_builtins_registered() stays for callers that genuinely enumerate — list_providers in the factory's error path, chimera which — so nothing became unreachable. A test resolves every registered name from a cold interpreter with no eager registration at all.

Quitting dumped a traceback after "Bye!"

Bye!
Traceback (most recent call last):
  ...asyncio/runners.py, line 118, in run → CancelledError
During handling of the above exception, another exception occurred:
  ...chimera/cli/code.py, line 873, in run_code → KeyboardInterrupt

A Ctrl+C landing during asyncio's shutdown makes asyncio.run cancel the task and re-raise, after the REPL already printed its goodbye. A clean quit rendered as a crash. run_code now catches it — quitting is not an error.

Two things the tests get right on purpose

The behavioural assertion runs in a subprocess. In-process, sys.modules is already polluted by the suite itself, so "openai is not imported" would have passed no matter what the code did — vacuous. Falsified against the old eager path: it fails there.

The name→module table is re-derived, not hand-read. _BUILTIN_MODULES is a claim about another module's import side effects, and those rot silently. The tests import each module in a clean interpreter and compare its registrations against the map. This is not hypothetical — modal_endpoint, xai and acmecloud each also register compatible as a side effect, so a map written by reading source would have been wrong on day one.

Gates

Full suite 10,484 passed / 142 skipped, ruff clean, mypy 736 files 0 errors. The single failure is test_validation_split, documented in CLAUDE.md as env-sensitive locally and green in CI.

Also worth knowing

chimera code has no --no-mcp / --no-lsp / --no-plugins escape hatches (unlike otter), so a user with slow init had no way to bisect it. Not changed here — flagging it as a follow-up, since after this fix init is no longer dominated by one thing.

🤖 Generated with Claude Code