[P1.59] Move the Sonnet-tier defaults to claude-sonnet-5-5 before claude-sonnet-4-5 reaches end of life (2026-11-30)

[frankbria]

Spun off from PR #1347 (issue #1267) during post-merge disposition.

Context

claude-sonnet-4-5 reaches end of life on 2026-11-30 (anthropic 1.11 DEPRECATED_MODELS). It is still the default for planning, execution, correction and supervision (adapters/llm/base.py:59-63). #1267 made every current model usable, which unblocks the move. The dated plan is in docs/AGENT_SYSTEM_REFERENCE.md under "Plan: moving the defaults off Sonnet 4.5".

Why it was not done in that PR

#1267 asked for the plan. The move changes cost and behaviour (Sonnet 5.5 runs adaptive thinking by default), so it needs its own comparison and review.

Definition of done (by 2026-10-31)

• DEFAULT_PLANNING/EXECUTION/CORRECTION/SUPERVISION_MODEL are claude-sonnet-5-5, and tests/adapters/test_model_defaults_guard_1112.py is updated.
• scripts/lifecycle --mode cli runs on the new default, and its cost and gate pass rate are compared with Sonnet 4.5 in the PR.
• A release carrying the new defaults ships at least two weeks before 2026-11-30, and scripts/quickstart-cleanroom/run.sh passes against the published package (the [P0.31] Published codeframe-ai 0.9.1 pins retired Anthropic model IDs — every LLM command 404s on a fresh install #1112 lesson).
• claude-haiku-4-5 (generation default) stays until it has an announced retirement date.

[frankbria]

Parked (needs-operator): the definition of done needs scripts/lifecycle --mode cli with a real key for the cost comparison, and the agent environment has none. Decided 2026-10-02: this ships as 0.9.5, after 0.9.4 (#1277), and still before the 2026-10-31 deadline.

[frankbria]

Operator steps: tactical, in order (2026-10-06)

Priority 3 of 7. Blocked by #1277 (0.9.4 must ship first). Deadline: tag v0.9.5 by 2026-10-31, two weeks before claude-sonnet-4-5 retires on 2026-11-30. About 1.5 hours: two real-key lifecycle runs (about $1–3 each) and a release.

⚠️ scripts/lifecycle defaults to haiku and overrides the compiled-in defaults (tests/lifecycle/conftest.py). A defaults comparison must pass --model sonnet, or both runs measure haiku.

1. Start the agent's half. Say "go on #1348". The agent opens a PR that:

• sets DEFAULT_{PLANNING,EXECUTION,CORRECTION,SUPERVISION}_MODEL = "claude-sonnet-5-5";
• updates tests/adapters/test_model_defaults_guard_1112.py;
• repoints the lifecycle sonnet alias to 5.5, so step 4 measures the new model;
• bumps the version to 0.9.5 with a CHANGELOG entry.

Note the PR number:

PR=<the PR number>

2. Load the key.

read -rs ANTHROPIC_API_KEY && export ANTHROPIC_API_KEY

3. Baseline: Sonnet 4.5 on main.

cd ~/projects/codeframe && git checkout main && git pull --ff-only
scripts/lifecycle --mode cli --model sonnet 2>&1 | tee /tmp/lifecycle-sonnet45.log

4. Candidate: Sonnet 5.5 on the PR branch.

gh pr checkout "$PR" --repo frankbria/codeframe
scripts/lifecycle --mode cli --model sonnet 2>&1 | tee /tmp/lifecycle-sonnet55.log

5. Hand both logs back. Tell the agent: "lifecycle logs are at /tmp/lifecycle-sonnet45.log and /tmp/lifecycle-sonnet55.log". It posts the cost, token and gate pass-rate comparison on the PR. The bar: the pass rate is not lower than 4.5's, and the cost is acceptable to you.

6. Approve the merge. Reply on the PR, or tell the agent, "merge #1348". It merges after CI and bot triage.

7. Release 0.9.5. Follow #1277's steps 2–12 with 0.9.5 in place of 0.9.4:

• cleanroom --source .;
• set the heading date;
• git tag -a v0.9.5 … && git push origin v0.9.5;
• watch release.yml;
• check the PyPI version;
• cleanroom against PyPI.

The release's model guard does a live check of the new default, using the staging key.

8. Close.

gh issue close 1348 --repo frankbria/codeframe --comment "0.9.5 ships claude-sonnet-5-5 defaults; lifecycle comparison on PR #$PR; cleanroom green on PyPI."

[frankbria]

Update (2026-10-06): the cost comparison runs in CI, not on your machine

lifecycle.yml takes mode/model inputs on workflow_dispatch and runs in the staging environment with the working Anthropic key. It has passed every Sunday. So steps 2–5 above (the two scripts/lifecycle --model sonnet runs) are the agent's:

gh workflow run lifecycle.yml --repo frankbria/codeframe --ref main -f mode=cli -f model=sonnet    # baseline, Sonnet 4.5
gh workflow run lifecycle.yml --repo frankbria/codeframe --ref <PR branch> -f mode=cli -f model=sonnet   # candidate, Sonnet 5.5

These are still billed to your Anthropic account (about $1–3 each). Your part is only to approve that spend and the merge. Order unchanged: #1443 (the discovery cap) lands first, then this, then the 0.9.5 release.