fix: model_id claude-sonnet-4-6 (no date suffix) + drop top_p (Sonnet 4.6 rejects both)

Two Anthropic API contract bugs surfaced by today's smoke after the
auth + env-override fixes unblocked the path:

1. model_id "claude-sonnet-4-6-20260131" doesn't exist. /v1/models
   confirmed the current frontier IDs use no date suffix: opus-4-7,
   sonnet-4-6, opus-4-6. Only older models (sonnet-4-5-20250929 etc.)
   are dated. Updated endpoints.json + smoke.yaml + mve.yaml.

2. Sonnet 4.6+ returns 400 if both `temperature` and `top_p` are sent.
   We were sending temperature=0.0 + top_p=1.0 on every call. With
   temperature=0 the top_p is a no-op anyway; removed it from the API
   call AND from SamplingParams (recording top_p=1.0 in the Trial row
   while the API used Anthropic's internal default would be a
   reproducibility lie).

After both fixes the first 5-trial smoke completed cleanly: 129 real
MCP tool calls dispatched, $0.9052 spent, all 5 Trial rows valid
(error_type=agent_gave_up — expected, git_mcp has no mounted volume).

Author: DevanshuNEU

CHANGELOG.md

@@ -4,6 +4,17 @@ All notable changes to tool-crowding are documented here. Format follows [Keep a
 
 ## [Unreleased]
 
+### Fixed — model_id `claude-sonnet-4-6-20260131` did not exist; corrected to `claude-sonnet-4-6` (2026-05-26)
+
+- `models/endpoints.json` + `configs/mve.yaml` + `configs/smoke.yaml` pinned the model as `claude-sonnet-4-6-20260131`. Anthropic returned 404 `not_found_error: model: claude-sonnet-4-6-20260131`. Probe of `https://api.anthropic.com/v1/models` confirmed the correct identifier is `claude-sonnet-4-6` (no date suffix). Sonnet 4.6 has not yet been promoted to a dated snapshot — older models like `claude-sonnet-4-5-20250929` use the dated form, current frontier models don't. Surfaced 2026-05-26 by the first live smoke after the auth + env-override fixes unblocked the path.
+- Caveat for reproducibility: `claude-sonnet-4-6` may be a floating alias rather than a frozen snapshot. The run_id chain captures the model_id string verbatim, so if Anthropic later releases `claude-sonnet-4-6-<date>`, our captured snapshot is the un-dated form. Migrate to the dated ID when one is published.
+
+### Fixed — removed `top_p` from API calls + SamplingParams (Sonnet 4.6+ rejects both `temperature` and `top_p` together) (2026-05-26)
+
+- `tcrun/agent.py::_invoke_api` sent both `temperature=0.0` and `top_p=1.0` on every API call. Sonnet 4.6 returns 400 `invalid_request_error: temperature and top_p cannot both be specified for this model`. Surfaced immediately after the model_id fix landed.
+- With `temperature=0.0` (our deterministic-dispatch default per pre-registration), `top_p` is a no-op anyway. Removed the `top_p` kwarg from the `client.messages.create` call. Removed the `top_p` field from `SamplingParams` (`tcrun/results.py`) so Trial rows don't record a value we never sent — recording `top_p=1.0` while the API used Anthropic's own internal default would be a reproducibility lie.
+- No new tests; the existing smoke + 257-test suite covers the schema migration (no tests referenced `SamplingParams.top_p`; the field was effectively dead code).
+
 ### Fixed — `.env` now overrides shell-exported credentials, with a loud warning on drift (2026-05-26)
 
 - `tcrun/cli.py::_load_env` previously called `load_dotenv(dotenv_path=env_path)` with `override=False` (python-dotenv's default). A stale `export ANTHROPIC_API_KEY=...` line in the user's shell rc (`~/.zshrc`, `~/.bashrc`) would silently shadow the value in `.env` — every subprocess inherits the shell env, dotenv refuses to overwrite, and Anthropic returns 401 against a key the dashboard swears is live. Surfaced 2026-05-26: a freshly-rotated key in `harness/.env` produced 401s for 30 minutes of debugging because `~/.zshrc:11` exported a long-revoked key into every shell. The harness's reproducibility chain depends on `.env` being the source of truth for credentials — silent shell-leak means two machines with different `.zshrc` files would produce different `run_id`s for the same checked-in config.

harness/configs/mve.yaml

@@ -35,7 +35,7 @@ distractors:
 
 N: [1, 5, 10]
 runs_per_cell: 3
-model: claude-sonnet-4-6-20260131
+model: claude-sonnet-4-6
 host: claude-desktop
 seed: 42
 

harness/configs/smoke.yaml

@@ -30,7 +30,7 @@ distractors: []
 
 N: [1]
 runs_per_cell: 1
-model: claude-sonnet-4-6-20260131
+model: claude-sonnet-4-6
 host: claude-desktop
 seed: 42
 

harness/models/endpoints.json

@@ -2,10 +2,10 @@
   "schema_version": "1.0",
   "models": [
     {
-      "model_id": "claude-sonnet-4-6-20260131",
+      "model_id": "claude-sonnet-4-6",
       "provider": "anthropic",
       "api_url": "https://api.anthropic.com/v1/messages",
-      "checkpoint_identifier": "claude-sonnet-4-6-20260131",
+      "checkpoint_identifier": "claude-sonnet-4-6",
       "default_temperature": 0.0,
       "default_max_tokens": 4096,
       "system_prefix_template": "You are a code-retrieval assistant. [nonce: {nonce}]",

harness/tcrun/agent.py

@@ -643,7 +643,6 @@ async def _invoke_api(
                 model=inputs.model_snapshot_id,
                 max_tokens=inputs.sampling_params.max_tokens,
                 temperature=inputs.sampling_params.temperature,
-                top_p=inputs.sampling_params.top_p,
                 system=system_prompt,
                 messages=messages,
                 tools=tools,

harness/tcrun/results.py

@@ -77,10 +77,16 @@ class ServerEntry(BaseModel):
 
 
 class SamplingParams(BaseModel):
-    """Model sampling parameters (SPEC.md §4)."""
+    """Model sampling parameters (SPEC.md §4).
+
+    `top_p` was removed 2026-05-26: Sonnet 4.6+ rejects requests that specify
+    both `temperature` and `top_p` (400 invalid_request_error). With
+    `temperature=0.0` (our deterministic default), `top_p` has no observable
+    effect anyway. Recording a `top_p` value in the Trial row that we never
+    actually sent to the API would mislead downstream reproductions.
+    """
 
     temperature: float = 0.0
-    top_p: float = 1.0
     max_tokens: int = 4096