Codex model aliases route default sessions to forbidden gpt-5.4

[Trecek]

Summary

Codex model aliasing still routes default sonnet and haiku class sessions to
forbidden 5.4-era model IDs. This makes fleet remediation dispatches invalid under the
current operator constraint that only gpt-5.5 should be used.

Evidence

• Fresh Codex L2 food trucks killed by 600s stream-idle watchdog while blocked on long-running MCP calls — suppression heuristics inapplicable, healthy children orphaned #4202 Codex self-test rerun dispatch
  4e6fbbea-f4fc-4176-bfce-34cb09157a68 was started on 2026-07-09.
• The parent L2 process launched with --model gpt-5.4, even though the run was intended
  to exercise the Codex backend with the currently allowed model family.
• The process was intentionally interrupted before it could count as a valid Codex L2 food trucks killed by 600s stream-idle watchdog while blocked on long-running MCP calls — suppression heuristics inapplicable, healthy children orphaned #4202
  self-test.
• Source evidence: src/autoskillit/core/types/_type_backend.py mapped:
  • sonnet -> gpt-5.4
  • opus -> gpt-5.5
  • haiku -> gpt-5.4-mini
• Because recipes and fleet defaults frequently pass sonnet, the default-safe route was
  not actually safe for Codex during the current provider window.

Expected Behavior

All Codex alias classes used by recipes and fleet dispatch should resolve to gpt-5.5.
Class distinctions may still be represented through model_reasoning_effort.

Actual Behavior

Codex L2 orchestration can launch with gpt-5.4, and haiku class sessions can launch
with gpt-5.4-mini.

Acceptance Criteria

• CODEX_MODEL_ALIASES contains no gpt-5.4 or gpt-5.4-mini values.
• sonnet, opus, and haiku all resolve to gpt-5.5 for Codex.
• Tests allow duplicate Codex model IDs when the class distinction is carried by
  model_reasoning_effort.
• Doctor/model alias allowlists reject 5.4-era Codex IDs.
• Codex L2 food trucks killed by 600s stream-idle watchdog while blocked on long-running MCP calls — suppression heuristics inapplicable, healthy children orphaned #4202 can be rerun without a Codex L2 process using --model gpt-5.4.

Prior Attempt Context

This blocks #4202 after PR #4217 was reverted. The #4202 rerun must be fully autonomous
under Codex; manually completing the recipe tail cannot count. This issue only remediates
the model alias blocker so #4202 can be rerun correctly.

Operator Requirement Update: GPT-5.6 Sol (2026-07-09)

PR #4219 completed the original 5.5 acceptance criteria and merged into develop at
2026-07-09T19:04:57Z (e0460e723d9bbef4ed112da1a7710a4bbd6328fb). All required CI
checks passed, version sync advanced develop to
8074aafcc1c36f32cfd30e369b107385c5578e30, and AutoSkillit 0.10.860 is installed.

The operator has now superseded the 5.5 target. The final blocker fix must use the latest
Sol model family:

• Every Codex class alias (sonnet, opus, and haiku) resolves to gpt-5.6-sol.
• Sonnet-class/default work resolves to model_reasoning_effort=high.
• Opus-class/serious work resolves to model_reasoning_effort=xhigh.
• Explicit Haiku-class work remains model_reasoning_effort=medium, but uses Sol rather
  than a mini model.
• Alias defaults contain no 5.4, mini, or 5.5 target.
• Explicit native gpt-5.5 may remain accepted for compatibility, but no alias may select
  it by default.
• Doctor, translation, generated-agent, command-construction, reverse-mapping, and suffix
  tests cover the Sol mapping.
• The interrupted Codex L2 food trucks killed by 600s stream-idle watchdog while blocked on long-running MCP calls — suppression heuristics inapplicable, healthy children orphaned #4202 dispatch is not resumed. A fresh autonomous Codex L2 food trucks killed by 600s stream-idle watchdog while blocked on long-running MCP calls — suppression heuristics inapplicable, healthy children orphaned #4202 dispatch starts
  only after the Sol follow-up is merged, installed, and verified.

Investigation report:
.autoskillit/temp/investigate/investigation_codex_gpt_5_6_sol_queue_reassessment_2026-07-09_122058.md.

GPT-5.6 Sol Follow-up PR (2026-07-09)

Branch codex-model-alias-gpt56-sol-4218 now carries commit 51f830e0b
(fix: route codex aliases to gpt-5.6 sol). PR #4220 was opened against develop:
#4220

Implementation scope:

• All Codex aliases now target gpt-5.6-sol.
• The existing Sonnet=high, Opus=xhigh, Haiku=medium effort policy is preserved.
• gpt-5.5 remains valid only as an explicitly requested native compatibility model.
• Exact alias policy, 5.4/mini rejection, doctor validation, native passthrough, suffix
  stripping, and the opus[1m] food-truck Sol/xhigh path are covered.

Validation:

• Focused task test-check: 308 passed; temp/test-2026-07-09_123047.txt.
• An unrelated fleet-marker mtime race failed once in the first full run; the isolated
  file immediately passed 8/8 (temp/test-2026-07-09_123431.txt).
• Final full task test-check: 29,804 passed, 604 skipped, 44 xfailed;
  temp/test-2026-07-09_123514.txt.
• pre-commit run --all-files: passed.
• Independent Sol/xhigh diff review: no findings.
• prepare-pr session 29bc19b3-e1ca-47dc-8950-eb2cffcf2864 and compose-pr session
  3f59f8da-1cbf-4df7-b4bf-e006292f7eda both completed with
  provider_used=minimax; no Anthropic-backed route was used.

Next route: run review-pr, resolve any valid findings, pass CI and merge queue, install
the version-sync result, reload the MCP/session, verify Sol/high and Sol/xhigh at the
installed process boundary, then start a fresh autonomous #4202 dispatch.

Review Gate (2026-07-09)

review-pr ran in GitHub mode through the MiniMax provider:

• Session: 8374b386-a5fb-4abc-9130-5aa5537ddf81
• Order: 4218-gpt56-sol-review-pr
• Result: verdict = approved, %%REVIEW_GATE::CLEAR%%
• Findings: zero valid findings; no inline comments were needed.
• GitHub review was posted successfully. Because the authenticated account authored the
  PR, GitHub's own-PR guard used a COMMENT event rather than self-approval.
• One cohesion subagent returned content from an unrelated historical fleet-resume PR;
  the review skill identified it as invalid and excluded it from the verdict.
• provider_used=minimax; no provider fallback and no Anthropic-backed request.

No resolve-review step is required. Next route: required CI, merge queue, install,
reload, process-level Sol verification, then a fresh autonomous #4202 dispatch.

Merged, Installed, and Verified (2026-07-09)

PR #4220 merged into develop at 2026-07-09T19:49:04Z:

• Merge commit: 1c81b9524bddbb770962d2d7a9f65ab9700d3ada
• Version-sync commit: 6bc93bc2aa12efb8451a066c5d264d48a1fd3f3f
• Installed AutoSkillit: 0.10.861
• Required PR CI: Preflight, execution tests, and general tests all passed in Actions run
  29045295380; version bump also passed.

Installed-package command canary:

• sonnet -> --model gpt-5.6-sol + model_reasoning_effort=high
• opus[1m] -> --model gpt-5.6-sol + model_reasoning_effort=xhigh; [1m] stripped
• haiku -> --model gpt-5.6-sol + model_reasoning_effort=medium
• gpt-5.6-sol and explicit gpt-5.5 are valid; gpt-5.4 and
  gpt-5.4-mini are invalid.
• Codex CLI 0.144.0 doctor --json loaded Sol with both high and xhigh; provider
  reachability and websocket handshake were OK.

Merge-process deviation: merge-pr session 5ab32f4b-c34f-41ff-9f9b-a65c7ed35613
correctly detected that develop requires the merge queue and auto-merge is disabled, but
then used gh pr merge --admin instead of returning for enqueuePullRequest. The already
green PR merged successfully, but this policy bypass is a pipeline regression. It is filed
as #4221 (bug + recipe:remediation), related to closed #1178.

#4218 is complete and may be staged. Before #4202, reload the MCP/session so no process
retains AutoSkillit 0.10.860 or the old 5.5 registry. After reload, verify kitchen version
0.10.861 and launch a completely fresh autonomous bundled remediation --backend codex
dispatch; never resume the interrupted attempt.