1 commit

Author	SHA1	Message	Date
gglucass	dee3500db6	feat(proxy): per-provider Kompress enable/disable (#1119) ... ## Description Adds per-provider control of Kompress (the lossy ML text compressor) via two new `ProxyConfig` fields, `disable_kompress_anthropic` and `disable_kompress_openai`. The global `--disable-kompress` / `HEADROOM_DISABLE_KOMPRESS` remains the baseline for **all** providers; the per-provider flags optionally override it for one provider (`None` inherits the global; `True`/`False` force-disable/enable). **Motivation.** In token mode, older excluded-tool results (`Read`/`Bash`/`Grep`/...) fall outside the recent-read protection window and become Kompress-eligible. On Anthropic that content is typically already in the cached prefix (0.1x cache-read discount), so recompressing it saves little, risks busting the prefix cache (1.25x writes), and lossily corrupts exact command/file output. This makes it possible to disable Kompress for the Anthropic pipeline while keeping it for OpenAI/Codex — **without changing any routing, tool-exclusion, or read-protection logic**. Structural compressors (SmartCrusher, log/search/diff, schema compaction) keep running for the disabled provider. Closes # N/A ## Type of Change - [ ] Bug fix (non-breaking change that fixes an issue) - [x] New feature (non-breaking change that adds functionality) - [ ] Breaking change (fix or feature that would cause existing functionality to change) - [ ] Documentation update - [ ] Performance improvement - [ ] Code refactoring (no functional changes) ## Changes Made - `ProxyConfig`: add `disable_kompress_anthropic: bool | None` and `disable_kompress_openai: bool | None` (default `None` = inherit global `disable_kompress`). - `HeadroomProxy.__init__`: resolve Kompress on/off per provider and build each pipeline's `ContentRouter` accordingly. When both providers resolve identically, **one `ContentRouter` instance is reused** so the Kompress model still loads once (startup warmup dedupes by `id()`); a second instance is created only when they differ. - Wiring: Click CLI (`--disable-kompress-anthropic/--enable-kompress-anthropic`, and `-openai`), argparse entrypoint, env builder (`HEADROOM_DISABLE_KOMPRESS_ANTHROPIC` / `_OPENAI`, tristate via new `_get_env_optional_bool`), and the `/config` debug payload. - Tests: `tests/test_proxy_per_provider_kompress.py`. ## Testing - [x] Unit tests pass (`pytest`) - [x] Linting passes (`ruff check .`) - [x] Type checking passes (`mypy headroom`) - [x] New tests added for new functionality - [x] Manual testing performed ### Test Output ```text $ uv run --extra dev pytest tests/test_proxy_warmup.py tests/test_proxy_disable_kompress.py \ tests/test_cli_proxy_env.py tests/test_proxy_per_provider_kompress.py -q collected 60 items tests/test_proxy_warmup.py ......... [ 15%] tests/test_proxy_disable_kompress.py .. [ 18%] tests/test_cli_proxy_env.py ............................................ [ 91%] tests/test_proxy_per_provider_kompress.py ..... [100%] ============================== 60 passed in 6.05s ============================== $ uvx ruff check headroom/proxy/server.py headroom/proxy/models.py headroom/cli/proxy.py All checks passed! $ uvx ruff format --check headroom/proxy/server.py headroom/proxy/models.py headroom/cli/proxy.py \ tests/test_proxy_per_provider_kompress.py 4 files already formatted $ uv run --extra dev mypy headroom/proxy/models.py headroom/proxy/server.py headroom/cli/proxy.py Success: no issues found in 3 source files ``` ## Real Behavior Proof - Environment: fresh clone of `chopratejas/headroom` @ `main` (8894ee0c); `uv run --extra dev`; Python 3.10.18 on macOS. - Exact command / steps: set `HEADROOM_DISABLE_KOMPRESS_ANTHROPIC=1`, build the proxy from env, and inspect each pipeline's `ContentRouter`: ``` HEADROOM_DISABLE_KOMPRESS_ANTHROPIC=1 uv run --extra dev python - <<'PY' import dataclasses from headroom.proxy.server import _proxy_config_from_env, HeadroomProxy cfg = _proxy_config_from_env() print("config.disable_kompress_anthropic =", cfg.disable_kompress_anthropic) print("config.disable_kompress_openai =", cfg.disable_kompress_openai) cfg = dataclasses.replace(cfg, optimize=False, cache_enabled=False, rate_limit_enabled=False, cost_tracking_enabled=False, code_aware_enabled=False) p = HeadroomProxy(cfg) a = p.anthropic_pipeline.transforms[-1]; o = p.openai_pipeline.transforms[-1] print("anthropic enable_kompress =", a.config.enable_kompress) print("openai enable_kompress =", o.config.enable_kompress) print("separate router instances =", a is not o) PY ``` - Observed result: `config.disable_kompress_anthropic = True`, `config.disable_kompress_openai = None`; `anthropic enable_kompress = False`, `openai enable_kompress = True`, `separate router instances = True`. `headroom proxy --help` lists `--disable-kompress-anthropic / --enable-kompress-anthropic` and `--disable-kompress-openai / --enable-kompress-openai`. - Not tested: end-to-end token-savings impact on live Anthropic traffic (reasoned from the 0.9 cache-read discount, not measured here); a follow-up could instrument per-provider/per-transform savings. ## Review Readiness - [x] I have performed a self-review - [x] This PR is ready for human review ## Checklist - [x] My code follows the project's style guidelines - [x] I have performed a self-review of my code - [x] I have commented my code, particularly in hard-to-understand areas - [ ] I have made corresponding changes to the documentation - [x] My changes generate no new warnings - [x] I have added tests that prove my fix is effective or that my feature works - [x] New and existing unit tests pass locally with my changes - [ ] I have updated the CHANGELOG.md if applicable ## Screenshots (if applicable) N/A ## Additional Notes - Opened as a draft pending maintainer review. - Backward compatible: with no per-provider override set, behavior is identical to today (both pipelines follow the global `disable_kompress`, sharing a single `ContentRouter`). - Tradeoff: disabling Kompress for a provider forgoes only *lossy text* savings; structural compressors still run. On Anthropic the forgone slice is low-value due to the 0.9 cache-read discount and avoids cache-bust risk on already-cached tool output. - CHANGELOG.md / docs not updated here — happy to add if desired. --------- Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>	2026-06-22 15:07:10 -05:00