Repository navigation
fix(scripts): make the Windows full run startable and resumable - #20
Merged
Merged
Conversation
What the first full run on Windows/WSL found, and the fixes: - run_full.sh resume exited silently before resuming: with `set -euo pipefail`, `ls jobs/full-<ts>/config.json ... | head -n 1` failed on the missing top-level path (harbor writes config.json one level down). Now a plain loop. Resume also filters AuthenticationError, since a run started without a key leaves 401 results behind. - env.sh: a failed `op read` exported an empty key (the pipeline's status was tr's); now it fails loudly, and an empty value is an error too. A variable already in the environment is kept and op is not called, so the key can be passed in from outside. - scripts/run_full_windows.ps1: the 1Password app never authorizes op.exe when it is called from inside WSL (authorization timeout every time), so the key is read on the Windows side and handed to the tmux session through WSLENV. Modes: start, resume, task. - scripts/full_progress.sh: done/pass counts, whether harbor is alive, log age. A tmux session outlives harbor, so `tmux ls` says nothing. - docs/setup-windows.md: swap=8GB and autoMemoryReclaim in .wslconfig (harbor itself passed 11 GB RSS at trial 56 and was OOM-killed), the 1Password reality, resume semantics, score.py's inner job directory, and a troubleshooting list from the night. - tests/test_env_sh.py: env.sh with a fake op on PATH. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
`run_full.sh subset <job dir> <task>...` runs the named tasks only, N_TRIALS attempts each (default 1), into the given job dir — harbor's -i per task. The Windows launcher gets the same as `run_full_windows.ps1 subset -Tasks a,b -Trials 3 -Job jobs/<name>` (tmux session `subset`, log `jobs/<name>.log`), and uses an LLM_API_KEY already in its environment so a wrapper that probes the gateway first costs one Allow click, not two. full_progress.sh drops the /89 denominator and reads the subset's log for non-full jobs. This is how the 20-task candidate run and the dev-slice runs are started. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Two things the first subset run from a fresh worktree hit: - `tee -a jobs/<name>.log` ran before harbor had created jobs/, so on a clone with no jobs/ yet the log couldn't be opened, tee exited, harbor died on the broken pipe and the tmux session vanished with no trace. `mkdir -p jobs` first. - `powershell -File ... -Tasks a,b,c` hands the parameter one string, not an array (that split only happens when called from inside PowerShell), so run_full.sh was asked for a task named "a,b,c". Split on commas in the script so both forms work. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
rzhang569
added a commit
that referenced
this pull request
Oct 14, 2026
ruoshi/feat-sampling at 6867ff0 (main: medium effort, 16k), 13 tasks, 3 trials each, concurrency 2. Baseline: no extra settings. Card sampling: LLM_TEMPERATURE=1.0, LLM_TOP_P=0.95, LLM_TOP_K=20, LLM_MIN_P=0. Never give up: AGENT_MAX_TURNS_WITHOUT_ACTION=1000. Exported with scripts/export_trials.py.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Fixes found during the first full 89-task run on Windows (2026-10-07, 11 h on a 16 GB laptop).
op.execalled from inside WSL, sorun_full_windows.ps1reads the key on the Windows side and passes it in viaWSLENV;env.shkeeps a key already in the environmentenv.sh: a failed or empty read is now an error (the first run exported an empty key and got 401s)run_full.sh resumesilently exited underpipefail; now finds harbor's nested config.json and also retries auth failuresfull_progress.sh: one-line progress, including whether harbor is still aliveautoMemoryReclaim, troubleshootingrun_full.sh subset <job dir> <task>...(N_TRIALS each) andrun_full_windows.ps1 subset -Tasks a,b -Trials 3 -Job jobs/<name>, for dev-slice runsenv.sh