Run agents on a local model

[arjunvlama]

Charter couldn't run on Ollama. Two call arguments that Ollama's client rejects stopped every run, both fixed in BoundFlow 0.7.4 (#114); this floors the pin there and covers the path from Charter's side.

	When	Error
Output cap	every call	TypeError: AsyncClient.chat() got an unexpected keyword argument 'max_tokens'
Forced finalizer	the last permitted call, once max_llm_calls is spent	TypeError: ... unexpected keyword argument 'tool_choice'

tests/e2e/test_ollama.py

The agent is charter init's own scaffold, so this is the README's quickstart with the model line changed. Nothing is faked — including the provider, which Charter builds from worker.yaml's provider: ollama rather than being handed by the test.

• test_a_local_model_completes_a_task — the run succeeds and the injected response_format comes back filled.
• test_a_run_out_of_calls_ends_on_the_finalizer — max_llm_calls: 1, so the first call is also the forced ending. Ollama can't force a tool at all, so the ending holds because the finalizer is the only tool left on offer.

Verified against Ollama 0.34.0, qwen2.5:7b, CPU only:

SDK	Result
boundflow 0.7.4	2 passed
boundflow 0.7.0	both fail with TypeError: ... unexpected keyword argument 'max_tokens', the original bug
OLLAMA_TEST_MODEL unset	2 skipped

They skip without OLLAMA_TEST_MODEL, and carry their own timeout mark because the suite's 180s is shorter than a single call on CPU.

Not in CI, deliberately. A model small enough to pull on every run is too small to drive the harness — qwen2.5:3b answered the ticket by writing a file instead of calling submit_result. The SDK-level contract with Ollama is covered by BoundFlow's Python SDK (Ollama) job, which can use a 0.5B model because it asserts on the call rather than on the agent.

A failure reason that lost its number

0.7.4 offers only submit_result on the last call a budget allows. A model that answers in text instead ends the run there, rather than on a further call — and that further call was where BoundFlow raised reached max_llm_calls=N. So the reason read only "the agent stopped without calling submit_result", which doesn't tell an operator whether to raise the ceiling or fix the agent.

_publish now names the ceiling when the budget is spent. New test, test_a_budget_spent_on_an_answer_in_prose_still_names_the_ceiling, which fails without it.

This also turned tests/e2e/test_failures.py::test_a_spent_budget_says_which_ceiling_it_hit red — not for that reason. The scripted model tells a subagent's call from the parent's by the absence of task in the offered tools, so it read the narrowed finalize offer as a subagent's and answered prose a turn early. Fixed in the fake: a bind of the finalizer alone is the last call, not a subagent.

Worth noting main would fail the same way — CI installs boundflow>=0.7.0, which now resolves to 0.7.4.

Docs

A new docs/local-models.md, linked from the README's documentation list. Notes from the runs rather than prose: the config, and the three things measured against Ollama 0.34.0 on CPU.

• max_call_seconds defaults to 60; a single quickstart call took about 170s.
• An unpriced model reports no cost, so max_cost_usd never binds — max_llm_calls is the ceiling that does. Usage is still reported, so runs are metered at $0.
• qwen2.5:7b completes the quickstart; qwen2.5:3b answered by writing a file instead of calling submit_result.

Plus the context window being the server's (OLLAMA_CONTEXT_LENGTH), and that forcing the ending is weaker on Ollama, which can't force a tool at all.

The README itself is unchanged except for one line in the documentation list.