Skip to content

Native arm: integration candidate on main 7ad8c9e72 - #5968

Merged
realmarcin merged 842 commits into
mainfrom
codex/native-main-integration-20261112
Nov 19, 2026
Merged

realmarcin merged 842 commits into
mainfrom
codex/native-main-integration-20261112

Conversation

@realmarcin

Copy link
Copy Markdown
Collaborator

What this is

The native agentic arm's integration candidate: the stack of #5839, #5857 and #5871 merged with main 7ad8c9e7295f1575ae9108ea62c12c83b1dd4db1, plus the repairs made on 2026-11-12. It is opened as a draft to get a complete CI run on the integrated source. It is not proposed for merge, and it does not replace the review of the three stack PRs.

No registration, no run and no provider, model or native-binary call is part of this work.

What it adds to #5871 (1a137b588)

  • The merge of main, with the four protected files kept equal to main's bytes.
  • The v2 freshness age at 1200 s (owner decision D136). Guard v1 is unchanged.
  • Guard admission bound to the latest claimed capture.
  • The scanner authority pin repaired after the merge.
  • The CI collection fault of Native stack fails CI at collection: a native test imports a script that requires the d4d-executions directory layout #5875 fixed; three test files made independent of the machine's directory layout; the import-origin failures isolated in tests.
  • The observer's model-label check repaired, with two test repairs and one added test.
  • The reviewed launch sequence script, review bundle version 3 and owner launch terms version 2.

Failures to expect on this head

The previous complete signal was run 42976790438 on 1a137b588: 152 distinct failing ids in 21 native test files, with two shards cancelled at the time limit. This run shows how many of those are cleared.

Not decided here

The pre-reservation margin, the scanner comparison's disposition, the protected post-images and the amendments N1 and NP1 are with the owner.

🤖 Generated with Claude Code

realmarcin and others added 21 commits November 13, 2026 22:13
…ication

Copy complete installed RECORD payloads without downloads or primary mutations; verify one distribution before the remaining closure. Keep broad-environment partial evidence and report final cohorts separately. Align remaining U1 envelope mutations with the current v3 wire calls without weakening refusal.
…imeout-20261114, 5b1c769)

Registers BASH_DEFAULT_TIMEOUT_MS and BASH_MAX_TIMEOUT_MS at 600000 in the
v2 native runtime environment, validated exactly, and makes the synthetic
peer take its Bash timeout from the child environment (synthetic peer pin
rotated, authorized by the lead). Also carries the helper CPU measurement
and read-only profiling scripts.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…meouts (#6235)

native_shared_runtime_values.py a0cffc39 -> 11e2254b: two required v2
environment keys and their exact-value validator. native_runtime_profiles.py
44d32d35 -> 58a71a05: the synthetic peer pin and one validator call. Both
computed from the merged bytes; no loader logic changed.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
realmarcin and others added 3 commits November 17, 2026 03:59
compare_d128_head_20261116 at bf551f4:
- all 132 difference classes identical to the accepted run;
- the nine violations identical;
- hits 5292 -> 5308 (exact 2750, relocations 2530, added 22, removed 6,
  changed 6);
- test-dependency relocations 22.

The fixture rebinds evidence 3c205c5b and reference candidate
bf551f4. All 28 D128 tests pass.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
CI dispatch run 44079015886 (shard 5) failed test_exit_join_has_two_second_bound:
D159 doubled native_pre_dispatch_timing._finish_writer's join timeout from 2 to 4 s.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
All 132 difference classes are identical to the accepted run; the nine
historical violations are unchanged. Fixture 2ba49e02, evidence 6d1aec95.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
realmarcin and others added 13 commits November 17, 2026 12:42
…on (#6290)

The shared saved readback classifies Bash permission denials with the
frozen run_native_canary.classify_denials, reached through the execution
capability projection. 10527af exposed denial_problems but not
classify_denials, so any Bash denial failed capture with
"AttributeError: classify_denials" and a run with a forbidden-command
denial failed its runtime gates. Add the pure helper to the roster and a
fast test that classifies a Bash denial through the projection.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
All 132 difference classes are identical to the accepted run and to the
283d959 refresh; hits 5292 -> 5308 unchanged. Fixture a73539d7,
evidence 05f00368, reference candidate 970eced (#6290 fix).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ion (#6290)

The shared saved readback classifies an ordinary (non-helper, non-lookup)
Bash command with native_command_policy.classify_program_command through
the execution capability projection, which did not list it. A run whose
history holds such a command, e.g. a denied forbidden command, failed
capture with "AttributeError: classify_program_command" (the second gap
found by test_denial_live_and_saved[Bash-not_prescribed] after 970eced).
A static audit of the shared readback path finds no other unlisted member.
Add it to the roster, with a fast test against the ordinary classifier.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
All 132 difference classes are identical to the accepted run; hits
5292 -> 5308 unchanged. Fixture 202bdc16, evidence 5ae5a34c,
reference candidate c175b88 (#6290 second fix).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
All 132 difference classes are identical to the a1d7 baseline; hits
5292 -> 5308 as at the previous refresh. Follows the issue 6289 test
and shard-weight change.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
PR run 44235627132 at 7b9497f cancelled shards 7, 10 and 12 at 60 minutes
with single native lifecycle cases (Write and Bash not_prescribed, Bash
prescribed, stalled shared admission) each still running after 3,438-3,524 s
on four-worker Linux. Whole-file balancing cannot give those cases a margin
inside 60 minutes. Launch admission requires every ordinary shard to succeed,
so the job budget, not the test, has to move. Check names, shard counts and
the required check roster are unchanged.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
All 132 difference classes are identical to the a1d7 baseline; hits
5292 -> 5308 as at the previous refresh. Follows the issue 6289 scheduling
estimates, denial helper-environment fix and shard job limits.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…o-end checks (6289)

Bash, Read and Write not_prescribed each run a complete native lifecycle
(2-3.5 h locally) and have never finished inside an ordinary CI shard: PR
runs 44235627132 and 44390330876 cancelled them at 60 and at 90 minutes
while every other shard passed. Owner decision 2026-11-18: skip them when
CI=true, with a visible reason, rather than block the merge; each passed
end to end locally (Bash at 69d86e0, Read at 283d959, Write at
a8eba65). A dedicated long CI lane for them is the follow-up.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
All 132 difference classes are identical to the a1d7 baseline; hits
5292 -> 5308 as at the previous refresh.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
realmarcin and others added 4 commits November 18, 2026 17:23
The fixture registered a fixed fourteen-day admission window ending
2026-11-19T00:00:00Z and a launch word authorized on 2026-11-11, so every
launch-sequence and launch-entry test refused admission from that instant
(PR run 44428665157: 59 failures on shard 11, 2 on shard 5; the same tests
passed on 2026-11-18). The window now spans today minus seven to today plus
seven days, the word is authorized yesterday, and the delegated-start case
runs its written offsets on today's date. Production windows come from the
real registration and are unaffected.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
All 132 difference classes are identical to the a1d7 baseline; hits
5292 -> 5308 as at the previous refresh.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
PR run 44437340827 failed seven tests in test_native_sequence_start.py
("capture word or current time is outside registration validity", "capture
journal clock moved backwards"): its 31 times were written on 2026-11-12 and
fell outside the now date-relative launch-sequence fixture window. They now
run on yesterday's date, inside the window and after the fixture's launch
word; offsets between times are unchanged. All ten test files that use the
launch-sequence, capture-word or sequence-start fixtures pass (546 tests).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
All 132 difference classes are identical to the a1d7 baseline; hits
5292 -> 5308 as at the previous refresh.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@realmarcin
realmarcin marked this pull request as ready for review November 19, 2026 04:54
@realmarcin realmarcin changed the title Native arm: integration candidate on main 7ad8c9e72 (draft, for a complete CI run) Native arm: integration candidate on main 7ad8c9e72 Nov 19, 2026
@realmarcin
realmarcin merged commit 4b6f8a9 into main Nov 19, 2026
94 checks passed
@realmarcin
realmarcin deleted the codex/native-main-integration-20261112 branch November 19, 2026 04:54
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