issue-622: contrastive-negatives dose-to-failure sweep#460
Draft
superkaiba wants to merge 305 commits into
Draft
issue-622: contrastive-negatives dose-to-failure sweep#460superkaiba wants to merge 305 commits into
superkaiba wants to merge 305 commits into
Conversation
|
The latest updates on your projects. Learn more about Vercel for GitHub.
|
…r review) Generated with [Claude Code](https://claude.ai/code) via [Happy](https://happy.engineering) Co-Authored-By: Claude <noreply@anthropic.com> Co-Authored-By: Happy <yesreply@happy.engineering>
…onomous plan-gate (tas [task.py]
Stale totals drift from the scripts they describe ("13-check" /
"19 checks" vs the actual 20-entry CHECKS list; "c0-c11" rots on the
next verify_plan.py check). Point at the script docstring catalogs
instead: workflow-improver.md (candidate target), analyzer.md,
promote-clean-result/SKILL.md (x2).
Generated with [Claude Code](https://claude.ai/code)
via [Happy](https://happy.engineering)
Co-Authored-By: Claude <noreply@anthropic.com>
Co-Authored-By: Happy <yesreply@happy.engineering>
…1 subagent=critic [task.py]
…rkflow-improver.md, analyzer.md, promote-clean-result SKILL.md)
…emoved (workflow-improver.md, [task.py]
…tail": "gcloud create ret [task.py]
…t) also quota-ref [task.py]
…to /home/thom [task.py]
…onomous plan-gate (tas [task.py]
…round=1 subagent= [task.py]
…read (watcher orp [task.py]
… effort=high write=True pol [task.py]
… effort=high write=True pol [task.py]
…onpolicy-testbed-v2) [task.py]
…itic-ensemble [task.py]
[task.py]
…onomous plan-gate (tas [task.py]
… effort=high write=True pol [task.py]
[task.py]
…round=26 subagent [task.py]
…jy phase=done after 60s. [task.py]
…:0", "job_id": "16001862" [task.py]
…ubagent=experimen [task.py]
…ubagent=experimen [task.py]
…central1 A100-80 [task.py]
…oj phase=done after 423s. [task.py]
… v2 --> ## Codex Interpretatio [task.py]
…ubagent=analyzer [task.py]
…tick-time drift-check candid [task.py]
…", "backend": "gcp", "ins [task.py]
…tail": "gcp provision ret [task.py]
…963315672) with r [task.py]
…a scripts/dispatch_issu [task.py]
[task.py]
… round 3 **Status:** READY-FOR-REV [task.py]
## Upload Verification — [task.py]
… subagent=code-re [task.py]
[task.py]
…mplant down ~5 [task.py]
…00) terminated post upl [task.py]
…hema Two surgical fixes for the round-3 crash (epm:failure code). (1) vLLM 0.11.0 _run_engine ZeroDivisionError. The throughput calc divides total_in_toks by tqdm's `elapsed` field; on the LoRA-swap fast path the engine finishes its first batch before tqdm advances and ZeroDivisionError aborts the call before any outputs return. Pass use_tqdm=False on every llm.generate(...) under experiments/contrastive_neg_geometry_472/: - eval_trajectory.py:280 (the crash site, _generate_on_policy_R) - r_generate.py:153 (_generate_batch) - eval_one_cell.py:271 (eval payload) vLLM keyword is documented + supported in 0.11.0 (LLM.generate signature shows `use_tqdm: Union[bool, Callable] = True`). Per-phase log lines already cover progress; the tqdm progress bar adds nothing. (2) FAILED-sentinel schema. scripts/dispatch_neg_setpoint_601.py was writing the FAILED sentinel as a raw 4-key dict, bypassing _write_sentinel which the same file defines for poll_pipeline conformance. The poller's drain step warned "missing required keys ['sentinel_schema_version', 'kind', 'version']; skipping" and read status=done off the supervisor's [phase=done] tail, turning a FAILED run into a silent success. Route the write through _write_sentinel(kind="epm:failure", task_id=sentinel_task_id) so the poll-pipeline contract is satisfied; the threaded task_id (not the default 601) keeps #622 dispatches correctly tagged. Smoke evidence (## Smoke run in epm:experiment-implementation v4): - signature smoke: vllm 0.11.0 LLM.generate accepts use_tqdm kwarg - sentinel-schema smoke: patched FAILED-sentinel parses via poll_pipeline._parse_sentinel with task_id=622 (not the default 601) - dispatcher --dry-run: wiring validated end-to-end without GPU work Refs: - Crash: vllm/entrypoints/llm.py:1610 in_spd = total_in_toks / pbar.format_dict["elapsed"] - Sentinel: poll_pipeline.py:1387 _SENTINEL_REQUIRED_KEYS check
…5 trajectory globs + phase0 gate/panel + completion sentinel, 26 MB)
superkaiba
added a commit
that referenced
this pull request
Jul 8, 2026
…n pinned-tree footer rows (#904) * task #1090: epm:merged — artifact_confirmed: true | full_rebase_deferred: true | surg [task.py] * task #1144: create — Port issue-1090 stranded shared-src fixes to main (sft.py CV [task.py] * task #1132: epm:merged — Step 10d auto-merge: issue-1132 -> main via PR #900 (gh pr m [task.py] * task #1132: epm:step-completed — <!-- epm:step-completed v1 --> ## Step Completed step: 10 a [task.py] * task #662: epm:progress — [autonomous_session_watch:orphan-respawn] active task (statu [task.py] * task #663: epm:progress — [autonomous_session_watch:orphan-respawn] active task (statu [task.py] * task #867: epm:progress — [autonomous_session_watch:orphan-respawn] active task (statu [task.py] * task #1139: epm:consistency — Verdict: PASS. No parent_id; no artifact/code reuse; single [task.py] * task #1137: epm:results — <!-- epm:results --> ## Completion Report **Task:** #1137 — [task.py] * task #1092: epm:progress — Launch 7: G2 PASSED + ~4 cells swept clean (265 shards; miti [task.py] * task #1142: plan v1 [task.py] * task #1142: epm:plan-verify — verify_plan.py: PASS, n_fail=0, n_warn=0, plan v1 (kind:infr [task.py] * task #1136: epm:merged — ## Step 10d merge — LANDED - **Merge SHA (main):** df2caa4 [task.py] * task #1136: epm:step-completed — <!-- epm:step-completed v1 --> ## Step Completed step: 10 a [task.py] * task #1090: epm:follow-ups — <!-- epm:follow-ups v1 --> ## Proposed Follow-Up Experiments [task.py] * task #1090: epm:followup-value-critique — Single-pass redundancy screen (Claude-only; Codex twin no-sh [task.py] * task #1142: plan v2 [task.py] * task #1139: epm:codex-task-failed — codex-quota-exhausted: org usage limit exhausted until 2026- [task.py] * task #1142: epm:consistency — N/A — kind:infra docs-only workflow-fix; no parent experimen [task.py] * task #1090: epm:step-completed — <!-- epm:step-completed v1 --> ## Step Completed step: 9a-b [task.py] * task #1139: epm:progress — Codex ensemble no-show all 3 lenses this round: codex-quota- [task.py] * task #1139: epm:plan — Plan v1 written to tasks/planning/1139/plans/plan.md (gpu_ho [task.py] * task #1139: planning → approved [task.py] * task #1139: epm:plan-approved — Auto-approved by the code-enforced autonomous plan-gate (tas [task.py] * task #1143: plan v1 [task.py] * task #662: epm:progress — [autonomous_session_watch:orphan-respawn] active task (statu [task.py] * task #663: epm:progress — [autonomous_session_watch:orphan-respawn] active task (statu [task.py] * task #867: epm:progress — [autonomous_session_watch:orphan-respawn] active task (statu [task.py] * task #1143: epm:plan-verify — verify_plan.py mechanical pre-pass on plans/v1.md: PASS — 0 [task.py] * task #1142: epm:progress — Phase 2 critic ensemble dispatched: 3 Claude lens critics + [task.py] * task #1139: approved → running [task.py] * task #1139: epm:step-completed — <!-- epm:step-completed v1 --> ## Step Completed step: 4b a [task.py] * task #1137: epm:code-review — <!-- epm:code-review v1 --> ## Code-Reviewer Verdict — PASS [task.py] * task #1139: epm:results — <!-- epm:results --> ## Completion Report **Task:** #1139 — [task.py] * task #1137: running → reviewing [task.py] * task #1138: epm:results — ## Completion Report **Task:** #1138 — add ONE canonical "B [task.py] * task #662: epm:progress — [autonomous_session_watch:orphan-respawn] active task (statu [task.py] * task #663: epm:progress — [autonomous_session_watch:orphan-respawn] active task (statu [task.py] * task #867: epm:progress — [autonomous_session_watch:orphan-respawn] active task (statu [task.py] * task #1112: epm:experiment-implementation — <!-- epm:experiment-implementation v10 --> ## Implementation [task.py] * task #1142: epm:codex-task-failed — codex-quota-exhausted: org usage limit exhausted until 2026- [task.py] * task #1142: plan v3 [task.py] * task #1142: epm:plan — Plan v3 written to /home/thomasjiralerspong/explore-persona- [task.py] * task #1142: planning → approved [task.py] * task #1142: epm:plan-approved — Auto-approved by the code-enforced autonomous plan-gate (tas [task.py] * task #1143: plan v2 [task.py] * task #1139: epm:code-review — <!-- epm:code-review v1 --> ## Code-Reviewer Verdict — PASS [task.py] * task #1142: approved → running [task.py] * task #1139: running → reviewing [task.py] * task #662: epm:progress — [autonomous_session_watch:orphan-respawn] active task (statu [task.py] * task #663: epm:progress — [autonomous_session_watch:orphan-respawn] active task (statu [task.py] * task #867: epm:progress — [autonomous_session_watch:orphan-respawn] active task (statu [task.py] * task #1143: epm:progress — stage-dispatch stage=planning round=1 subagent=critic-ensemb [task.py] * task #825: epm:results — {"eval_numbers": {"mode": "production", "gate_fail": false, [task.py] * task #825: epm:progress — stage-dispatch stage=followup-verifying round=1 subagent=upl [task.py] * task #825: epm:progress — stage-dispatch stage=followup-interpreting round=1 subagent= [task.py] * task #1143: epm:consistency — Verdict: PASS — infra analogue check (no parent experiment; [task.py] * task #1142: epm:results — <!-- epm:results --> ## Completion Report **Task:** #1142 — [task.py] * task #1137: epm:test-verdict — Verdict: PASS. Scope: touched (33 files from select_step9c_t [task.py] * task #1142: epm:progress — Step 5 code review dispatched (Claude-only; codex-code-revie [task.py] * task #1137: epm:completion-audit — ## Completion Audit — PASS Audited against original body.md [task.py] * task #1137: reviewing → completed [task.py] * task #1137: epm:done — Outcome: stalled-alert noise fixed in the watcher, two mecha [task.py] * task #1143: epm:codex-task-failed — codex-quota-exhausted: org usage limit exhausted until 2026- [task.py] * task #825: epm:upload-verification — <!-- epm:upload-verification v1 --> ## Upload Verification — [task.py] * task #1112: epm:code-review — <!-- epm:code-review v10 --> ## Code-Reviewer Verdict — PASS [task.py] * task #1143: plan v3 [task.py] * task #1143: epm:plan — Plan v3 written to /home/thomasjiralerspong/explore-persona- [task.py] * task #1143: planning → approved [task.py] * task #1143: epm:plan-approved — Auto-approved by the code-enforced autonomous plan-gate (tas [task.py] * task #1143: approved → running [task.py] * task #1143: epm:progress — stage-dispatch stage=implementing round=1 subagent=implement [task.py] * task #1138: epm:code-review — <!-- epm:code-review v1 --> ## Code-Reviewer Verdict — PASS [task.py] * task #1112: epm:progress — stage-dispatch stage=followup-running round=1 subagent=dispa [task.py] * task #1112: epm:backend-selected — {"attempts": [], "chosen_kind": "gcp", "cluster": null, "ela [task.py] * task #825: epm:upload-verification — <!-- epm:upload-verification --> ## Upload Verification — ro [task.py] * task #1142: epm:code-review — <!-- epm:code-review v1 --> ## Code-Reviewer Verdict — PASS [task.py] * task #1138: running → reviewing [task.py] * task #1112: epm:cluster-launched — {"attempt_id": "att-20260708-111901", "backend": "gcp", "ins [task.py] * task #1112: epm:backend-selected — {"attempts": [{"cluster": null, "detail": "gcp rung flexstar [task.py] * #1143: check 30 Pattern D — backtick sub-path + parenthetical count in pinned-tree footer rows Adds a fourth claim position to verify_task_body.py check 30: a backtick `dir/` token immediately followed by a parenthetical OPENING with the count-noun (the #1112 footer shape — `raw_completions/` (7,165 files: …)), bound via the new _nearest_preceding_pinned_tree_link binder to the nearest preceding hex-pinned /tree markdown link on the same line (bracket-free, newline-free, <=400-char gap) and scoped to <link-prefix>/<sub-path> through the existing whole-prefix verification leg (two-sided files compare + folder-inflation diagnostic; WARN-only, shared probe memo/cap unchanged). The count-noun lookahead is widened to decline ANY distributive "per <word>" qualifier (the #460 per-adapter shape); A/B keep their narrow #833 per-namespace lookahead by design. 6 new offline tests incl. the FULL verbatim #1112 footer fixture (>=2 pinned links before the claim pins nearest-of-N binding). Co-Authored-By: Happy <yesreply@happy.engineering> Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * issue-1143: reset tasks/ to origin/main after recovery merge (stale pre-rebase task-state duplicates; union-merge line dupes) --------- Co-authored-by: Happy <yesreply@happy.engineering> Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
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
1 participant
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.
Closes task #622. Plan: tasks/running/622/plans/plan.md (v3).