Growing context

Status: evidence-backed problem, not a universal framework ranking. Reviewed 2026-09-19 against local tests and published findings; see those pages for the exact dependency pins and measured configurations.

Symptom and mechanism

Cost climbs as a task gets longer, even when tool schemas do not change.

Full transcript replay resends earlier messages. Fixed framework prompt overhead and accumulated history are different costs.

Diagnose and reproduce

Measure request bytes/tokens per turn. Compare the initial intercept with the slope before blaming a framework's fixed prompt.

Use the context budget explorer to see how replay, recent windows, protected evidence, retrieval, and deterministic compaction change what reaches the model.

From the repository root, with the relevant optional adapters installed:

python .github/scripts/report_growth.py 30

The test is the executable contract; published evidence explains scoped findings. Missing adapters may skip comparisons; inspect the test summary before generalizing.

Alternatives and tradeoffs

Window old turns, retrieve selected evidence, or summarize; each can discard required obligations or provenance.

Verification and open question

A required early fact must survive the selected policy; report missing information rather than guessing. Semantic summary quality requires real-model trials.

Open question: does the same outcome hold for your actual provider, workload, configuration, and framework version? Existing scripted findings alone cannot answer that.

Implication for a new harness

Context builder with explicit budget, retained obligations, source IDs, and observable exclusions.

Return to the problem register.