Skip to content

task #1338: crash-persist harvests per-run logs before oversized-dir skip#1093

Open
superkaiba wants to merge 4 commits into
mainfrom
issue-1338
Open

task #1338: crash-persist harvests per-run logs before oversized-dir skip#1093
superkaiba wants to merge 4 commits into
mainfrom
issue-1338

Conversation

@superkaiba

Copy link
Copy Markdown
Owner

Closes task #1338 (workflow-fix, kind: infra). Crash-persist: new _up_data_logs() pass harvests per-run *.log tracebacks from swept data dirs BEFORE the EPS_PERSIST_DIR_CAP_BYTES oversized-dir skip; worker-log sweep excludes committed logs/daily|weekly .md notes. Plan v2 + code-review PASS + test-verdict PASS on events.jsonl.

…versized-dir skip

EPS_PERSIST_PY heredoc: hoist the swept-dirs tuple to PARTIAL_DIRS; new
_up_data_logs() (any logs/*_logs path component or .attempt filename,
newest-first, tail-capped + count-bounded by the #885 knobs, staged under
EPS_PERSIST_DATA_LOG_STAGE_DIR, ONE upload_folder commit to
{dest}/data_logs/<dirname>/<relpath>) sequenced after _up_logs() and
before the partial-dirs sweep; _up_logs() excludes committed *.md notes
with one aggregate SKIP line. Existing repo upload paths + output lines
byte-identical. Print-cap recount ~104 < 120 (comment updated).
…rate snapshot regen

3 new behavioral heredoc tests (harvest-before-oversized-skip incl. the
worker_logs/gpu0.log arm + the under-cap dual-path cell; count-bound
newest-first via os.utime; worker-log .md notes exclusion), a render
string-pin sibling, harness EPS_PERSIST_DATA_LOG_STAGE_DIR isolation,
no-dirs SKIP extension, PARTIAL_DIRS loop anchor update (newline-prefixed
to pin the top-level sweep loop). Snapshot fixture DELIBERATELY
regenerated (+89/-5 lines vs #1205) with #1338 provenance;
test_render_startup_script_is_valid_bash ran green BEFORE the regen.
…md parenthetical

compute-backend-failover.md Part A gains item 4 (data_logs harvest,
items 4-8 renumbered 5-9 with cross-refs updated) + the .md-notes
exclusion sentence in item 3 + the Sweep-scope paragraph update;
CLAUDE.md crash-diagnostics parenthetical gains data_logs/.
…eate-argv test

test_gcp_create_argv_forwards_real_xet_kill_switch's hermetic delenv list
predated #1205's GITHUB_TOKEN addition to STARTUP_SECRET_ENV_KEYS: a
batch-mate's load_dotenv() puts the repo-root .env's GITHUB_TOKEN into
os.environ mid-pytest-session and the #1205 secrets guard then raises
(reproduced with the BASE ffea76b gcp.py — NOT introduced by this
task's diff; it would have failed the Step 9c gate's 47-file selection).
Fix derives the delenv set from gcp.STARTUP_SECRET_ENV_KEYS minus the two
threaded keys, so a future key addition cannot re-break it.
@superkaiba
superkaiba marked this pull request as ready for review July 15, 2026 16:12
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant