Skip to content

Native: version the CLI-reported model label for Claude Code 2.1.325 (C10 refusal) - #6296

Merged
realmarcin merged 4 commits into
mainfrom
codex/native-cli-model-label-20261119
Nov 20, 2026
Merged

realmarcin merged 4 commits into
mainfrom
codex/native-cli-model-label-20261119

Conversation

@realmarcin

Copy link
Copy Markdown
Collaborator

Why

Cell 1's C10 scripted observation probe (2026-11-19T21:05Z, pinned Claude Code 2.1.325, --model claude-opus-5[1m]) completed successfully but was refused by native_u4b_probe_sequence._observed. 2.1.325 labels the model claude-opus-5 in the init frame and as the terminal modelUsage key, where the addendum-2 / D112 alias contract (observed with 2.1.269) requires claude-opus-5[1m]. The reported contextWindow is 1,000,000, so the 1M model is in effect; only the label changed.

Owner decision 2026-11-19: amend the code rather than change the pinned binary.

What

  • New explicitly declared alias-basis versions, final basis 4 and provisional basis 2. Each declares requested claude-opus-5[1m], returned claude-opus-5 and CLI-reported claude-opus-5.
  • Under the new versions, a reported label of claude-opus-5 is accepted only when:
    • the reported contextWindow is exactly 1,000,000;
    • the requested alias matches the saved invocation argv.
  • Old final versions 2/3 and provisional version 1 keep their semantics: [1m] is still required. Synthetic profiles keep their old versions.
  • The consumers updated are:
    • probe sequence;
    • model aliases;
    • u4b runtime;
    • capture words;
    • gate engine;
    • shared observations and permissions;
    • u4b policy;
    • the probe-packet drafter.
  • Tests use a sanitized real 2.1.325 transcript and request fixture (dummy local key only). They cover:
    • acceptance;
    • refusal of a 200k window, a [1m] label under the new version, mixed labels, and an old version with the new label;
    • the old-version boundaries.

Verification so far (local)

  • Focused tests: 249 passed (439.8 s).
  • D128: 28 passed; no pin change.
  • Whole-suite collection: 27,705 tests.
  • Ordinary native cohort: still running.
  • CI dedicated lanes: covered by this PR's CI.

Review

A Codex adversarial review follows. Findings will be filed as issues and the in-scope ones addressed on this branch before merge.

🤖 Generated with Claude Code

realmarcin and others added 4 commits November 19, 2026 13:25
Declare final basis 4 and provisional basis 2 for the unsuffixed CLI report while retaining the requested 1M alias, exact million-token proof and saved argv binding. Preserve historical and synthetic declarations; cover all active readers and explicit draft selection with sanitized offline regressions.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Reject synthetic use and cross-version observation joins. Preserve the CI dedicated-case partition in the offline verification runner so long fixtures do not contend for freshness windows.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…6298)

The fast-path, post-probe chain and envelope-comparison wrappers pin the
frozen e772 source, which lacks native_model_aliases.reported_model, and
they need companion scripts this repository does not carry. They belong to
the operator tooling, which will select the merged head. The draft-producer
tests stay; the tests that loaded the wrappers are removed.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
#6299)

The CLI-reporting label is consulted only when the runtime basis declares
final 4 or provisional 2; every other runtime compares against
runtime['model'] exactly as at e7721b6.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@realmarcin
realmarcin merged commit c959da6 into main Nov 20, 2026
91 checks passed
@realmarcin
realmarcin deleted the codex/native-cli-model-label-20261119 branch November 20, 2026 00:30
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