36's sshd reset new connections during key exchange for a few minutes; three jobs in
a row gave up after one try per host and paused the batch.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Every attempt kept result.tar, its unpacked tree and the copies in the attempt, so a
third of the hot disk went to a tree nothing reads again. result.tar and its manifest
still record exactly what the job returned.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A downloader task's agent served test files from 127.0.0.1 inside its own job; the rule
the user set is about the outside network. Look-alike hosts such as 127.0.0.1.evil.io
and localhost@evil.io still count.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
rollout.salty_run keeps the gateway, grading and evidence on 36 and moves only the
agent step into a Salty job. The gateway listens on 127.0.0.1 with a per-attempt
bearer token; the job reaches it over port 22 with a restricted one-time key that is
removed after each attempt. rollout.batch gains --backend salty and repeatable
--credential-file, and writes pass/fail SFT exports through rollout.sft at the end.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
rollout.dashboard serves catalog counts, attempts and complete trajectories, and the public
prompt and template of each task. catalog.controllers()/complete() are split out so the
dashboard can reuse receipts it has already read.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
100 tasks from the HWE arm-linux code-gen pool (outside the lk/man/pth
range given to batch 22), each with at least one independent
contract-edge defect and plain public docs. Claude reviewed 100/100
USABLE against current digests; no ID or source overlap with main or
batch 22. Difficulty is unmeasured.
progress.md records the wrap-up: all Codex sessions and Qwen schedulers
stopped for the user's rollout, Qwen results so far (57 HARD of 373),
and what is left on each batch branch.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
From the 101 lk/man/pth code-gen sources: 63 tasks (C11 36, C++17 27,
11 canonical families), 38 skipped for scope or duplication. Each task
has at least one independent contract-edge defect, boundary types
rotate, and public docs state trap clauses as plain contract prose.
Claude reviewed 63/63 USABLE against current digests. Difficulty is
unmeasured.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Only author-side spec.md files change, so no task digest moves. The specs of
concur-release-capsule and resource-temp-link follow batch 20's family and
version. The 53 rejected Bos environments wait: JuiceFS returns EIO for the
whole source tree.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
210 batch-15 tasks used 119 ad-hoc family names. Each now carries one of the
canonical root-cause families (11 of the 19 are used), with version + 1 and a
spec.md note, and was recalibrated through both gates. Claude checked that only
family, version and spec.md changed, that every family is canonical, and the
new trace6 Docker receipts against current digests (210/210 USABLE).
Squashed from the 36-local branch batch20-family.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
One user-space repair task per eligible HWE debug-diag / compile-deploy base
ID: 320 admitted (C 218, C++ 97, build scripts 5), 1 skipped by the content
filter. Each ships a public spec with a check-to-basis table, at least two
defects with single-fix fixtures, a symptom-masking fixture and an alternative
correct implementation. Claude re-checked all 320: structure, the HWE quality
items and trace6 Docker calibration receipts against current digests
(320/320 USABLE). Difficulty is unmeasured. Known follow-up: 212 tasks use 120
ad-hoc family names; the next batch maps them to the canonical list.
The batch's 1,256-line progress log stays in its factory report and the
36-local branch batch15-hwe.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
After the importer accepted fixture scripts that rewrite a target (dd69720),
54 of the 55 environments batch 16 could not import loaded and were imported
with graders narrowed to the declared files and READMEs stating the delivery
boundary. Claude re-checked all 54 packages and their trace6 Docker
calibration receipts against current digests (54/54 USABLE). Difficulty is
unmeasured. Squashed from the 36-local branch batch19-rglob2.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The 191 Bos environments whose graders rglob the workspace were imported under
the new rule: graders read only the declared candidate files and the READMEs
state that only those files are delivered. After the import-scan fix, 135 were
approved; 55 are still refused because a fixture script writes the same file
twice, and 1 has a target missing from its template. Claude re-checked all 135
packages, the delivery-boundary wording and their trace6 Docker calibration
receipts against current digests (135/135 USABLE). Difficulty is unmeasured.
Squashed from the 36-local branch batch16-rglob.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The 72 tasks that had no difficulty conclusion were measured with
qwen3.8-27b-nvfp4 on trace6: 68 ROUTINE, 3 MEDIUM, 1 held (S500_n3_100,
insufficient basis), none HARD. 172 raw attempts, 148 valid. Root review also
flagged 14 tasks whose README and grader disagree or whose grader misses a
README requirement; batch 17 fixes and re-measures them. No task changes;
records only. Squashed from the 36-local branch batch13-screen.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
192 unused Bos n1/a environments inventoried in the multi-agent mode: 159 are
rglob-type and 6 ROS2-type (not importable under the current contract), 13 were
rejected by the importer or missing targets, 14 imported. Claude re-checked all
14 packages and their trace6 Docker calibration receipts against current
digests (14/14 USABLE). Difficulty is unmeasured. Squashed from the 36-local
branch batch14-n1a.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
119 Bos n3 environments inventoried in the multi-agent mode: 69 imported with
graders rewritten to run candidates in subprocesses, 18 rejected because the
source gold fails its own contract, 32 rglob-type skipped. S500_n2_094 and
S500_n2_115 READMEs now state the oracle's integer, nesting and exception rules,
which lifts their holds. Claude re-checked all 71 packages: structure, isolation
and trace6 Docker calibration receipts against current digests (71/71 USABLE).
Difficulty is unmeasured. Squashed from the 36-local branch batch11-n3.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
120 tasks screened with qwen3.8-27b-nvfp4 on trace6 in the multi-agent mode:
29 HARD, 13 MEDIUM, 75 ROUTINE, 3 held for ambiguous prompts (053/094/115).
364 raw attempts, 285 valid measurements. No task changes; records only.
Squashed from the 36-local branch batch10-screen; see docs/progress.md.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Clarify four ambiguous prompts (S500_n2_053, S500_n2_096, S500_n2_103,
mem-mapping-lease), drop the unused tests.py from eleven rewritten n3 tasks,
add cheat_read_grader to two concurrency tasks and normalise their source
fields. All 18 recalibrated on trace6 (820/820 as expected).
Squashed from the 36-local branch batch12-cleanup; see docs/progress.md.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Batches now commit on their own branch and land on main with git merge
--squash, so main gains one commit per batch. Record the four-step push.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
With batch 8's rewritten verifiers and its trace6 calibrations on cold
storage, all 161 tasks are USABLE: the 33 former BLOCKED tasks, the two
ambiguous HOLD tasks and the three replacement concurrency tasks have new
digests and are UNMEASURED. Batch 6 session 2/3 trajectories move seven
tasks to ROUTINE and three to PARTIAL.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
36 main (ca48b33) carries batch 6 session 2/3 measurements and batch 8
(20261019-C-usable-8), which rewrote the 33 isolation-blocked tasks, the two
ambiguous ones and replaced the three held-out concurrency tasks. Mac main adds
the trace6 catalog, the self-heal batch runner and its trial on 36.
docs/progress.md keeps both sides by time: 36's session 2 log, then Mac's
trace6 catalog and self-heal sections, then batch 8 and session 3.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>