Skip to content

Compress tool outputs, logs, files, and RAG chunks before they reach the LLM. 20% fewer tokens for coding agents, 60-95% fewer tokens for JSON, same answers. Library, proxy, MCP server.

fix(models): version-boundary longest-prefix match in ModelRegistry.get (#1658)

Abhay Singhb699bedf9528
diff --git a/CHANGELOG.md b/CHANGELOG.md
index 9b3630e1e..b91c005f8 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -43,6 +43,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
* **memory/sync:** stop the Claude Code sync adapter from clobbering distinct memories that share a first line. `write_memories` derived each file name from the first line of the content only (`headroom_{slug}.md`), so two different DB memories whose first lines slugify identically wrote to the same file and the second silently overwrote the first — and because the loser never landed on disk, the next sync re-exported it, ping-ponging the pair forever. When the slug is already taken by a *different* memory (distinct `headroom_id`) the file name is now disambiguated with a content-hash suffix; an update to the same memory still rewrites its slug file in place, so existing file names are unchanged.
* **cli/proxy:** honor `HEADROOM_MIN_TOKENS=0` / `HEADROOM_MAX_ITEMS=0`. The Click `proxy` command built these with `_get_env_int_optional(name) or 500`/`or 50`, so an explicit `0` — a legitimate value (`min_tokens_to_crush=0` means "crush every item") — was treated as falsy and silently replaced with the default. The `headroom proxy` argparse path already preserved `0` via `_get_env_int`, so the two entry points disagreed. The Click path now uses the same None-checking helper.
* **proxy:** strip the inbound `Content-Encoding`/`Transfer-Encoding` request headers on the Anthropic `/v1/messages` and OpenAI `/v1/chat/completions` paths before forwarding upstream. `read_request_json_with_bytes` already decompresses the inbound body (zstd/gzip/deflate/br), so the bytes forwarded upstream are plain JSON — but these two handlers left the original `content-encoding` header in place, so a client (or an edge proxy like a Cloudflare Worker) that sent a compressed body got its request rejected with upstream HTTP 400 because the provider tried to decompress already-decoded JSON. The `/v1/responses` handler already carried this fix (#1542); it is now applied to the messages and chat paths too.
+* **models:** fix the model registry's prefix fallback silently returning the wrong context window. `ModelRegistry.get` accepted any registered name as a `str.startswith` prefix and returned the *first* match, so `gpt-4-32k-0613` resolved to `gpt-4` (8192) instead of `gpt-4-32k` (32768), and unregistered ids like `gpt-4.1`/`gpt-4.5` inherited `gpt-4`'s 8192-token window — making the proxy think a nearly-empty context was almost full and compress far too aggressively. The fallback now requires the registered name to end at a version boundary in the query (so `gpt-4.1` no longer matches `gpt-4`) and picks the longest qualifying name (so `gpt-4-32k-0613` → `gpt-4-32k`).
* **proxy:** include the system prompt, tools, and the response-shaping request fields in the SemanticCache key. `_compute_key` hashed only `{model, messages}`, so two non-streaming requests with identical messages but a different top-level `system` prompt, tool set, sampling config, or output-shaping field collided on one key and the second caller was served the first's cached response — generated under different request semantics, in the default config (`cache_enabled` defaults on). The key now folds the request fields that shape generation — `temperature`/`top_p`/`top_k`/`max_tokens`/`stop`, plus OpenAI `tool_choice`/`response_format`/`parallel_tool_calls`/`seed`/`presence_penalty`/`frequency_penalty`/`logit_bias`/`n`/`logprobs`/`top_logprobs`/`reasoning_effort`/`verbosity`/`modalities` and Anthropic `thinking`/`tool_choice`/`output_config` — canonicalizing `system`/`tools` so a moved `cache_control` breakpoint does not fragment it, and the handlers snapshot the fields once at the cache read and reuse them at write so a body mutated by the pipeline cannot diverge the key. Non-streaming path only.
* **learn (verbosity):** `--verbosity --apply --all` now aggregates the savings baseline across every project instead of overwriting it per project (last-project-wins), which previously left the output shaper with a tiny, unrepresentative baseline. The applied verbosity level comes from the project with the most samples ([#1288](https://github.com/headroomlabs-ai/headroom/pull/1288)).
* **proxy/anthropic:** restore token-mode compression on continued Claude Code turns with a frozen prefix and deferred CCR tool injection. Token mode now runs request-side compression even when the client did not pre-register `headroom_retrieve`, relying on the existing marker-triggered injection override to keep emitted CCR markers redeemable ([#1487](https://github.com/headroomlabs-ai/headroom/issues/1487)).
diff --git a/headroom/models/registry.py b/headroom/models/registry.py
index 4a0863bf2..6526c9c2a 100644
--- a/headroom/models/registry.py
+++ b/headroom/models/registry.py
@@ -631,12 +631,29 @@ class ModelRegistry:
if model_lower in _ALIASES:
return _MODELS[_ALIASES[model_lower]]
- # Prefix matching
+ # Prefix matching. Two rules keep this from silently mis-resolving
+ # newer or larger variants:
+ # 1. The registered name must end at a version boundary in the query
+ # (next char is a separator like ``-``/``/``/``:``/``@``/``_``), so
+ # ``gpt-4.1`` — a distinct model, not a variant of ``gpt-4`` — does
+ # not inherit gpt-4's 8192-token window (the ``.`` is not a
+ # boundary, so it no longer matches).
+ # 2. When several names qualify, the LONGEST wins, so
+ # ``gpt-4-32k-0613`` resolves to ``gpt-4-32k`` (32768) rather than
+ # the shorter ``gpt-4`` (8192) that happens to be registered first.
+ best: ModelInfo | None = None
+ best_len = -1
for name, info in _MODELS.items():
- if model_lower.startswith(name):
- return info
+ if not model_lower.startswith(name):
+ continue
+ rest = model_lower[len(name) :]
+ if rest and rest[0] not in ("-", "/", ":", "@", "_"):
+ continue
+ if len(name) > best_len:
+ best = info
+ best_len = len(name)
- return None
+ return best
@classmethod
def resolve(
diff --git a/tests/test_models.py b/tests/test_models.py
index 578e3674a..69b3e0c85 100644
--- a/tests/test_models.py
+++ b/tests/test_models.py
@@ -90,6 +90,24 @@ class TestModelRegistry:
assert info is not None
assert info.name == "gpt-4o"
+ def test_get_prefix_matching_prefers_longest_registered_name(self):
+ """`gpt-4-32k-0613` must resolve to `gpt-4-32k` (32768), not the
+ shorter `gpt-4` (8192) that is registered first."""
+ info = ModelRegistry.get("gpt-4-32k-0613")
+ assert info is not None
+ assert info.name == "gpt-4-32k"
+ assert ModelRegistry.get_context_limit("gpt-4-32k-0613") == 32768
+
+ def test_get_prefix_matching_requires_version_boundary(self):
+ """`gpt-4.1`/`gpt-4.5` are distinct models, not variants of `gpt-4`.
+ A `.`-separated suffix must not match `gpt-4`, so they no longer
+ inherit gpt-4's 8192-token window (they fall back to the default)."""
+ assert ModelRegistry.get("gpt-4.1") is None
+ assert ModelRegistry.get("gpt-4.5-preview") is None
+ # Not silently reported as an 8192-token model:
+ assert ModelRegistry.get_context_limit("gpt-4.1") != 8192
+ assert ModelRegistry.get_context_limit("gpt-4.1", default=100) == 100
+
def test_resolve_future_google_family_fallback(self):
"""Resolve should return provider-scoped fallbacks for plausible future models."""
with patch("headroom.models.registry.get_model_pricing", return_value=None):