fix(v1): discover the context window under the model card's specs

[parkerpettit]

Prime Inference reports a model's context window as specs.context_window on its /v1/models cards rather than at the top level, so discover_threshold returned None there and Bash auto-compaction (compaction = {} with no summarize_at_tokens) never triggered: no proactive compaction, and an overflow was re-raised instead of compacted. Read the nested field too.

Checked against Prime Inference with the bundled Bash program behind a request-logging proxy. Before: discover_threshold("openai/gpt-4o-mini") was None, and a tool-calling run in auto mode made no checkpoint request while the same run with an explicit --summarize-at-tokens compacted. After: discovery returns 111616 (128000 − 16384) for gpt-4o-mini and 43616 for a 60k-window model, and in auto mode the program issues the checkpoint-summary request once a tool turn crosses the threshold.

Note

Fix discover_threshold to read context-window fields from model card specs metadata

Updates discover_threshold in compaction.py to also inspect a dictionary-valued specs extension when resolving configured context-window fields. Top-level metadata remains preferred, and positive integer validation still applies before deriving the threshold.

^{Macroscope summarized 7dfbfae.}

[macroscopeapp]

Approvability

Verdict: Approved at 7dfbfae

Macroscope's review found this PR approvable — This one-file bug fix extends existing context-window lookup to nested model-card specs metadata while preserving current top-level provider behavior. Its runtime impact is limited to allowing the existing compactor to discover thresholds for affected provider responses, with no schema, security, or deployment changes.

^{You can add or adjust custom eligibility rules. Learn more.}