Raise agent/verifier timeouts to 3600s for compile-compcert, gpt2-codegolf,
path-tracing, and make-doom-for-mips — all repeatedly timing out at their
previous limits (900–2400s).
Wrap all apt-get calls in test.sh with `timeout 120` across 82 tasks to
prevent verifier hangs caused by slow/unresponsive apt mirrors inside
containers (root cause of recurring VerifierTimeoutError on caffe-cifar-10).
Increase timeouts for caffe-cifar-10 and kv-store-grpc
caffe-cifar-10: verifier timeout 1200→1800s (verifier installs uv+pytest
from scratch, eating into the budget — timed out in 3/5 jobs)
kv-store-grpc: agent timeout 900→1800s (timed out in 3/6 jobs)
element-web (53 parsers + 3 configs):
- Fix Jest parser to handle ✕/✖ (U+2715/U+2716) as FAILED markers
- Previously ✕ lines were treated as describe block names, corrupting
all subsequent test name matching
- Remove flaky InviteDialog wasm test from pass_to_pass in 3 instances
tutanota (2 parsers + 2 configs):
- Fix parser embedding non-deterministic assertion counts in test names
(e.g. "3029 assertions" changes to "3261" after fix is applied)
- Use stable names: "test/api/Suite.ts | api tests" without counts
- Skip workspace package tests, only parse main api+client suites
teleport (1 task.toml):
- Bump cpus 4→8 for TestConcurrentFanout under go test -race
Fix SWE-bench Pro: apply test-file patches at test-time
Move git checkout of test files from Dockerfile to test.sh so that
agents running git restore/reset on startup cannot undo them.
All 731 tasks affected.
Fix swebenchpro: resolve 37 oracle failures without reducing difficulty
The direct-completion agent extracts the output file path from the
instruction using a regex that requires /workspace/ or /app/ prefix.
Without it, code gets written to /workspace/solution.py instead of
/workspace/HumanEval__N.py, causing all tests to fail (ImportError).
Add Completion prefix section to all humanevalplus instruction.md files
The direct-completion agent (completion mode) needs this section to extract
the function stub as the prompt prefix for base model continuation.
Without it, the fallback wraps the entire instruction in a docstring,
producing garbage completions.
Fix humanevalplus-32: remove 7 ill-conditioned test inputs
These polynomial inputs cause catastrophic cancellation in
the naive evaluation (sum of coeff*pow(x,i)) on Python 3.12,
making atol=0.0001 impossible to satisfy at any float64 x.
Fix humanevalplus-32 oracle: use canonical Newton's method
