Skip to content

Commit 505f6f0

Browse files
authored
Merge pull request #60 from alanshurafa/claude/swebench-poc-status
Record the SWE-bench POC run as executed
2 parents 4f68804 + b321359 commit 505f6f0

1 file changed

Lines changed: 27 additions & 3 deletions

File tree

.planning/notes/2026-09-02-swebench-poc-plan.md

Lines changed: 27 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -1,8 +1,32 @@
11
# SWE-bench proof-of-concept run: one task, every arm, published (2026-09-02)
22

3-
Status: APPROVED (Alan, 2026-09-02). Execute from this file in a fresh Opus
4-
session with Sonnet workers. Orientation: read `benchmarks/COMPLETE-SUITE-PLAN.md`
5-
(on branch `codex/code-benchmark-battery`) and `benchmarks/code/README.md` first.
3+
Status: EXECUTED 2026-09-02. All four phases complete. Page live at
4+
<https://alanshurafa.github.io/co-evolution/>. Nine arms on `pallets__flask-5014`,
5+
9/9 resolved by the official evaluator, zero infrastructure failures. Landed as
6+
PRs #57 (battery to master), #58 (arms H and I), #59 (results page + Pages
7+
workflow). Fable spend $21.96 against a $15 estimate.
8+
9+
Four corrections to this plan, found in execution:
10+
11+
- `evaluate-swebench.sh` and `validate-predictions.sh` take a predictions FILE,
12+
not `--run-id`. Evaluation is nine invocations, one per condition, which is
13+
also how the evaluator names its reports.
14+
- `run-canary.sh` had no `--task` flag and routed every condition to the agentic
15+
driver, so the plan's own P2 command would have killed arms F and G. Both
16+
fixed in #58.
17+
- A one-task run against the 5-task canary renders as "1 / 5". Added a
18+
`swebench-verified-poc` suite so the denominator matches what ran.
19+
- Codex 0.144.5 on Windows accepts `--sandbox workspace-write` and enforces
20+
read-only. Arms B and E had to be discarded and rerun under
21+
`CODE_BENCH_CODEX_SANDBOX=danger-full-access`.
22+
23+
Open follow-ups: `run-canary.sh` aborts the whole batch when one arm produces no
24+
patch instead of scoring it zero; Codex 0.144.5 -> 0.152.1 upgrade and a retest
25+
of `workspace-write` on Windows; charts and filters on the page once a suite has
26+
enough tasks to make them mean anything.
27+
28+
Original orientation: read `benchmarks/COMPLETE-SUITE-PLAN.md` and
29+
`benchmarks/code/README.md`.
630

731
## Goal (Alan's words)
832

0 commit comments

Comments
 (0)