docs: pilot A vs F comparison report — conditional go for 89x3 (pilot-compare-a-vs-f)

feat: pilot F 5×1 — wg-native benchmark with surveillance loops (pilot-f-5x1)

docs: research supervisor agent loop for persistent task oversight (research-supervisor-agent)

fix: disable verify/spawn circuit breakers in terminal bench — let time limit govern (remove-hard-evasion)

test: smoke test end-to-end model pipeline (smoke-model-pipeline)

fix: sort profile ranking by real metrics — benchmark > usage > pricing efficiency (fix-profile-ranking-sort)

feat: resolve short model names via cache lookup (fix-model-spec-resolution)

feat: populate real model metadata — proxy scoring, verbose output, refresh command (fix-profile-data-fetch)

fix: add is_proxy field to ModelBenchmark and expand curated benchmarks (fix-broken-compilation)

fix: model routing in TB adapter — strip all WG_* env vars, validate API keys, detect mismatches (tb-fix-model-routing)

fix: replace hardcoded Claude model IDs with constants in CLI, tests, and remaining code (fix-hardcoded-models-other)

fix: re-add CLAUDE_OPUS_MODEL_ID import to openrouter_smoke tests (.verify-fix-hardcoded-models-service)

fix: add missing model constant imports to nested test modules (.verify-fix-hardcoded-models-service)

feat: implement thinking token support in executor and streaming

feat: inject canonical memory into condition F system prompt (tb-f-memory-impl)