cawdex 1.35.74 → 1.35.76

package/resources/hal/cawdex_agent/README.md

@@ -1,24 +1,24 @@
-# Cawdex HAL adapter
-
-This directory is a HAL-style custom agent package. It exposes `run(input, **kwargs)` in `main.py` and shells out to the installed `cawdex` CLI in non-interactive benchmark mode.
-
-Defaults:
-
-- SWE-bench-like tasks return a git patch string.
-- ScienceAgentBench-like tasks return a solution/trajectory string.
-- AppWorld-like tasks return `Completed` after a successful Cawdex run.
-- WebDevBench-like tasks are routed to the `webdevbench` benchmark profile, which keeps canary requirements plus frontend-backend and production/security validation evidence visible.
-- SWE-Cycle-like tasks are routed to the `swe-cycle` benchmark profile, which keeps lifecycle phase, environment setup, implementation, verification-test generation, and static/dynamic judge evidence visible.
-- SWE-CI-like tasks are routed to the `swe-ci` benchmark profile, which keeps current/target commits, test gaps, inferred requirements, and CI-loop verifier deltas visible.
-- SWE-PRBench-like tasks are routed to the `swe-prbench` benchmark profile, which reviews PR metadata and diff first, expands only for concrete suspected findings, and returns severity-rated file/line findings unless patches are explicitly requested.
-- TML-Bench/Kaggle-style tabular ML tasks are routed to the `tml-bench` benchmark profile, which extracts the data contract, avoids hidden-label leakage, trains an honest baseline, and validates the generated submission schema before completion.
-- Pi-Bench-style proactive personal assistant tasks are routed to the `pi-bench` benchmark profile, which builds a user/workspace/app context contract, tracks hidden-intent hypotheses, asks focused clarification only when needed, and verifies observable state after proactive actions.
-- USACO and other text-response tasks return the original task dict with a `response` field.
-- Oracle-like fields such as `patch`, `test_patch`, `solution`, `answer`, `gold`, `FAIL_TO_PASS`, and `PASS_TO_PASS` are omitted from the prompt unless `CAWDEX_HAL_INCLUDE_ORACLE_FIELDS=1` is set.
-- Traces and logs are written under `.cawdex/hal-trace/` unless `CAWDEX_HAL_TRACE_DIR` is set.
-
-Useful overrides:
-
-- `CAWDEX_HAL_COMMAND` or `CAWDEX_HAL_COMMAND`: command used to invoke Cawdex, default `cawdex`.
-- `CAWDEX_HAL_TIMEOUT_SEC`: per-task timeout, default `1800`.
-- HAL `-A model_name=...`, `-A provider=...`, `-A max_turns=...`, and `-A output_format=...` are forwarded to Cawdex CLI flags when present.
+# Cawdex HAL adapter
+
+This directory is a HAL-style custom agent package. It exposes `run(input, **kwargs)` in `main.py` and shells out to the installed `cawdex` CLI in non-interactive benchmark mode.
+
+Defaults:
+
+- SWE-bench-like tasks return a git patch string.
+- ScienceAgentBench-like tasks return a solution/trajectory string.
+- AppWorld-like tasks return `Completed` after a successful Cawdex run.
+- WebDevBench-like tasks are routed to the `webdevbench` benchmark profile, which keeps canary requirements plus frontend-backend and production/security validation evidence visible.
+- SWE-Cycle-like tasks are routed to the `swe-cycle` benchmark profile, which keeps lifecycle phase, environment setup, implementation, verification-test generation, and static/dynamic judge evidence visible.
+- SWE-CI-like tasks are routed to the `swe-ci` benchmark profile, which keeps current/target commits, test gaps, inferred requirements, and CI-loop verifier deltas visible.
+- SWE-PRBench-like tasks are routed to the `swe-prbench` benchmark profile, which reviews PR metadata and diff first, expands only for concrete suspected findings, and returns severity-rated file/line findings unless patches are explicitly requested.
+- TML-Bench/Kaggle-style tabular ML tasks are routed to the `tml-bench` benchmark profile, which extracts the data contract, avoids hidden-label leakage, trains an honest baseline, and validates the generated submission schema before completion.
+- Pi-Bench-style proactive personal assistant tasks are routed to the `pi-bench` benchmark profile, which builds a user/workspace/app context contract, tracks hidden-intent hypotheses, asks focused clarification only when needed, and verifies observable state after proactive actions.
+- USACO and other text-response tasks return the original task dict with a `response` field.
+- Oracle-like fields such as `patch`, `test_patch`, `solution`, `answer`, `gold`, `FAIL_TO_PASS`, and `PASS_TO_PASS` are omitted from the prompt unless `CAWDEX_HAL_INCLUDE_ORACLE_FIELDS=1` is set.
+- Traces and logs are written under `.cawdex/hal-trace/` unless `CAWDEX_HAL_TRACE_DIR` is set.
+
+Useful overrides:
+
+- `CAWDEX_HAL_COMMAND` or `CAWDEX_HAL_COMMAND`: command used to invoke Cawdex, default `cawdex`.
+- `CAWDEX_HAL_TIMEOUT_SEC`: per-task timeout, default `1800`.
+- HAL `-A model_name=...`, `-A provider=...`, `-A max_turns=...`, and `-A output_format=...` are forwarded to Cawdex CLI flags when present.

package/resources/hal/cawdex_agent/__init__.py

@@ -1 +1 @@
-"""HAL adapter package for Cawdex."""
+"""HAL adapter package for Cawdex."""