[HTExplicit]

Problem

#229-#231 removed the Codex ticket subsystem but left pieces that only existed for it. An audit of main found 78:

• some survived under new names, such as the response model guard and the "logical turn" staging;
• some were tolerance code for its stored data, such as retired-key dropping and retired-extra filters;
• one was the old module's shell: the "native Codex runtime host", which takes a database lease on every OAuth request to serve one boolean.

The owner's instruction: retired features are deleted, and stored data is fixed at the root instead of being tolerated in code.

Change

Deleted

• Response model-declaration guard, on every path including the account test. Behaviour is upstream's: the reply is delivered and the usage log records upstream_model_mismatch.
• Codex quality diagnostics: the /codex-quality-runs routes, the grant/trial headers, every hook in forwarding, and the service and repository code.
• 「Codex 出站指纹」: the panel, its two endpoints, profile selection and the per-request wire observation. The account test still prints its identity snapshot and the headers it sent; it now builds them itself.
• Plugin-owner / account-view job metadata decoding (account_job_legacy.go).
• Tolerance added by refactor(codex): remove ticket and route-qualification harvesting #229/chore(codex): purge dormant ticket data (migration 263) #230: retired config keys, retired account-extra keys, kind_unsupported and its text, placeholder parameters, and tests that only pinned tolerance.

Replaced

• Runtime host. The lease, epoch, generation, receipt headers, encrypted config and module dispatch give way to one plain settings row, codex_runtime_config = {"request_zstd":bool}. It is decoded strictly, and a missing row means gateway.openai_codex_request_zstd. A package-level compression gate and direct typed calls into codexruntime/profile and codexruntime/recovery replace the dispatch. GET/PUT /api/v1/admin/settings/codex-runtime keep their path and body.

Restored

• x-codex-turn-state to the fork's own contract from before Add GPT-6 Sol/Luna, verified Cookie routing and account-scoped prompt rules #192 (hunk-level reverts; the logical-turn file is deleted).
• Responses-Lite default instructions to upstream v0.2.11.
• Create-account dialog: the fingerprint-mode pseudo-option default is gone; the dialog writes device explicitly, as Edit and Bulk already do.

Where a differing upstream model is shown (replies that declare another model now reach the UI)

• Single account test: prints the line the batch test already shows, 「↳ 上游响应: （模型不一致）」.
• Quality check: such a run is judged 疑似降智 whatever its answer, with the reason next to the verdict (owner decision from the preview).

Migration 264

• Compression value. Carried over by comparing the recorded normalized-config hash with the two possible bodies. An unknown hash raises instead of guessing.
• Dropped. sub2api_plugin_state (it raises if another plugin has rows there) and sub2api_plugin_leases.
• Retired installation. Removes the stored configuration and package of the retired codexrip.codex-runtime installation, and deletes the two legacy settings rows.
• Fingerprint mode. Written explicitly for OAuth accounts that had none.

The migration is forward-only. A read-only production pre-check confirmed its three preconditions: the receipt lists the plugin, the hash is the known "on" value, and the state table holds only this plugin's 42 rows.

Not changed here, by agreement with the parallel reasoning-recovery rework: recovery semantics, and a few always-nil error returns in plugin_codex_recovery.go that the rework removes next.

Verification

• go build ./...; go vet under no tag, unit and integration; golangci-lint v2.13 on the touched packages;
• targeted tests for identity, compression, turn-state (the TurnStateCommits suite restored to its pre-Add GPT-6 Sol/Luna, verified Cookie routing and account-scoped prompt rules #192 form), the settings load and PUT, the account test output, and account jobs;
• migration 264 on a throwaway PostgreSQL 18.1: a fresh database, seeded pre-migration rows of every kind plus rows that must survive, each raise case, and a second apply;
• frontend vue-tsc --noEmit, check:i18n, eslint and the specs of touched components;
• a six-lens review (hot path, turn-state, completeness against the audit, migration, frontend, kept features) with every finding verified: 25 confirmed, 22 fixed, 3 deferred to the parallel rework;
• a local mock-data before/after preview of the changed screens.

Not run: race suites locally (no cgo), Redis-dependent integration tests, real model calls.

🤖 Generated with Claude Code