Report each claw's effective context window

[teddytennant]

Toward #1.

Running the shipped example configs through a reader shows the spread the issue is asking about:

claw       context window  source
----------------------------------------------------------------
generic             unset  claw_configs/generic/mykey.py
                           note: GenericAgent exposes max_tokens (output), not a context budget
hermes            200,000  claw_configs/hermes/config.yaml
nanobot             unset  claw_configs/nanobot/config.json
                           note: unset; claw falls back to its runtime default
openclaw            unset  $OPENCLAW_STATE_DIR/openclaw.json
zeroclaw          128,000  claw_configs/zeroclaw/config.toml

So hermes runs at 200k, zeroclaw at 128k, and nanobot at whatever NanoBot defaults to. The README lists same prompt, no network answers, future-commit stripping and per-instance isolation as properties enforced for every claw, and context budget is not on that list even though it moves the result.

claw_swebench/context_window.py reads the budget out of each claw's live config under that claw's own key: custom_providers[<model.provider>].context_length for hermes, agent.max_context_tokens for zeroclaw, agents.defaults.contextWindowTokens for nanobot and openclaw. It reads rather than asserts, so the record is what the claw actually ran with instead of what the harness believes. A missing or malformed config reports a note and returns None; it never raises into a run.

Three places consume it. run_infer.py logs the budget at startup and warns when it is unset. Each instance's metadata.json gains a claw_config block with claw, model, timeout, max_turns and the context window plus the file and key it came from, so an old artifact directory can be audited later. python3 -m claw_swebench.context_window prints the table above.

I did not normalize anything. No claw's behaviour changes and no published number moves. The nanobot example still ships without contextWindowTokens; its _comment now names the key and says the budget is unset, because picking a number for it changes nanobot's score and that is a call about the benchmark, not about the plumbing. If you want the normalized version, say which budget and I will add the flag that writes it into each claw's config.

What I ran: python3 -m unittest discover -s tests, 17 tests covering the readers, the failure modes and the shipped examples. run_infer.py --claw {nanobot,hermes,zeroclaw} --dataset verified far enough to see the startup line for each, which is where the numbers above came from. I did not run inference or evaluation, so there is no before/after on any score here.

The thing this does not do is tell you what the reported nanobot run used, since that lived in a local config.json that is gitignored. That one is still yours to answer, and adding the per-claw configs behind the leaderboard runs would close the rest of #1.