# Codex Sol + Luna Worker — Execution Setup Contract Contract version: `v4.1.0-rc5`. Release status: - RC5 is a source candidate with Source CI `PASS` on Windows, Ubuntu, and macOS; - RC5 real Global upgrade is `NOT RUN`; - RC5 O1–O10 runtime acceptance is `NOT RUN`; - RC5 is not published and is not Stable; - Stable remains `v4.0.0`; - Published Preview remains `v4.1.0-rc4`. The RC5 Source Commit A repository suite recorded `152/152` tests passing. These tests and Source CI are source / repository validation only. They do not establish a real Global upgrade, fresh runtime acceptance, or RC5 publication. > [!IMPORTANT] > This is an independent community project. It is not affiliated with, sponsored by, or endorsed by OpenAI or ModelDial. The user must review this contract before execution. RC5 remains an unpublished source candidate; the public README installation entry continues to point to the published RC4 Preview. Never silently expand permissions, installation scope, or network access. This document is an execution contract for Codex, not a replacement installer. Orchestrate `scripts/install.py`; do not reproduce its merge, ownership, backup, migration, rollback, or uninstall logic with ad hoc shell commands. Approved RC5 runtime source commit: ```text 5ae88ff9190b31174c55a6136c0c8c8611d0b34c ``` This setup contract installs the immutable RC5 runtime payload from Source Commit A shown above. The repository README may later pin this contract document from a documentation commit. That documentation commit is not the runtime payload source and is never passed to the installer. The RC5 acceptance-boundary redesign keeps the product runtime payload frozen: `src/selector.py`, `scripts/install.py`, `templates/AGENTS.global.md`, and `.codex/agents/*` do not change. The acceptance contract consists of this setup contract, `RUNTIME_TESTS.md`, `ARCHITECTURE.md`, and `SECURITY.md`; those documents may change when the acceptance design changes. Such contract-only updates are not a product-runtime change. `PRODUCT_RUNTIME_CHANGED = NO`; `ACCEPTANCE_CONTRACT_CHANGED = YES`. `SETUP_CONTRACT_SELF_REFERENCE_REQUIRED = NO`. ## 0. Document Role You are the installation agent. Your job is to inspect, plan, back up through the installer, install, validate, and report. Preserve an evidence-based boundary between repository checks and real runtime checks. You must: 1. obtain the exact approved RC5 runtime source commit shown above; 2. discover the current OS, Python, Codex, and target `CODEX_HOME` without guessing; 3. inspect existing state without exposing unrelated private content; 4. run the existing installer dry-run before any mutation; 5. classify the installation mode and stop on uncertainty; 6. use the existing installer transaction for apply, migration, rollback, and uninstall; 7. validate static payloads and then require a fresh Codex task for runtime smoke testing; 8. report what was changed, preserved, backed up, or blocked. You must not guess paths, overwrite unknown configuration, install system dependencies, change authentication, update Codex, move Git tags, execute mutable `master` content, bypass ownership checks, or silently widen permissions. If a required condition is missing, stop with `BLOCKED` and the exact reason. ## 0A. Installation Preflight Before cloning or downloading source, creating a temporary source directory, writing capability-probe state, or invoking any installer mode, run a read-only prerequisite preflight. Do not install system dependencies, modify `PATH`, or search internal Codex Desktop application directories for an executable. This copy-and-paste installation path supports Codex Desktop on Windows, Ubuntu/Linux, or macOS, with WSL treated as a separate Linux environment. It requires all of the following before source acquisition: - the current Codex Desktop task; - a `codex` command already resolvable by the task environment and able to complete `codex --version`; - Python 3.11 or newer with standard-library `tomllib`; - Git, because this contract has no archive or downloader fallback and requires an immutable exact-commit checkout; - HTTPS access to the public GitHub repository for a read-only `git ls-remote` check. On Windows, use `Get-Command` for command discovery. On macOS, Linux, or WSL, use `command -v`. Then run the equivalent of: ```text codex --version --version -c "import sys, tomllib; assert sys.version_info >= (3, 11); print('PYTHON_CAPABILITY_PASS')" git --version git ls-remote https://github.com/SuperDaddyV/codex-sol-luna-worker.git HEAD ``` Do not treat a Codex Desktop internal application path, versioned cache directory, or content-hash directory as a supported CLI discovery mechanism. A Desktop-bundled executable satisfies this contract only when the current task environment already exposes it as a resolvable `codex` command and `codex --version` succeeds. Display the result before any filesystem write: ```text Sol/Luna Installation Preflight Codex Desktop: CURRENT_SESSION / NOT_CONFIRMED Codex CLI: PASS / MISSING_OR_UNUSABLE Python: PASS / MISSING_OR_UNSUPPORTED Git: PASS / MISSING GitHub HTTPS: PASS / BLOCKED Ready: YES / NO ``` `Ready: YES` requires every hard prerequisite above to pass. If any item fails, set `Ready: NO`, do not create or modify files, and stop with exactly one concise remediation: - `BLOCKED: CODEX_CLI_REQUIRED` — install or expose a supported Codex CLI yourself, reopen Codex Desktop, and retry only after `codex --version` succeeds; - `BLOCKED: PYTHON_3_11_WITH_TOMLLIB_REQUIRED` — install a supported Python yourself, reopen the task environment, and retry; - `BLOCKED: GIT_REQUIRED_FOR_IMMUTABLE_SOURCE` — install Git yourself, reopen the task environment, and retry; - `BLOCKED: GITHUB_HTTPS_REQUIRED` — restore read-only HTTPS access to the public repository and retry without changing proxy or system settings automatically. PowerShell is the native Windows shell used to run these checks, but it is not a separate cross-platform product dependency. `curl` and other download tools are not required by this contract. ## 1. Target State The global target contains five native custom agents: | Agent name | Model | Reasoning effort | Leaf setting | | --- | --- | --- | --- | | `luna_low` | `gpt-5.6-luna` | `low` | `[agents] enabled = false` | | `luna_medium` | `gpt-5.6-luna` | `medium` | `[agents] enabled = false` | | `luna_high` | `gpt-5.6-luna` | `high` | `[agents] enabled = false` | | `luna_xhigh` | `gpt-5.6-luna` | `xhigh` | `[agents] enabled = false` | | `luna_max` | `gpt-5.6-luna` | `max` | `[agents] enabled = false` | The managed global `config.toml` block enables multi-agent work and sets `max_concurrent_threads_per_session = 3`. The managed global `AGENTS.md` block makes Sol the planner, orchestrator, ambiguity resolver, and final acceptance owner; it requires the Daily Selector, bounded Luna delegation, a Task Contract, a Context Firewall, Sol acceptance, and one concise Delegation Receipt after receipt-eligible non-trivial work. The Receipt summarizes facts already observed after the delegation decision. `Luna unavailable` is evidence-gated and requires a current-task, parent-visible selector or native-agent availability failure that arose naturally on the normal delegation path. The Receipt cannot force delegation, lower the threshold, invoke a selector, probe, tool, child, or network solely to create evidence, write state, telemetry, or repository content, or expose private reasoning. The selector is installed at `/sol-luna-v4/selector.py`. Its state root is `/sol-luna-v4/state`. The v4 ownership manifest is `/sol-luna-v4/install-manifest.json`. ## 2. Explicit Non-goals Do not install or introduce any of the following: - Hook Router; - `PreToolUse` enforcement; - managed-child registry; - daemon or background scheduler; - database or dashboard; - plugin framework; - custom orchestration engine. Do not lower the delegation threshold, force Luna use, add a Receipt-specific state or configuration flag, treat no delegation as Luna unavailability, use `Luna unavailable` as a default fallback, or treat Receipt text as runtime attestation. RC5 adds observability metadata and policy UX without adding a second selection or state authority. Do not modify the selector algorithm, agent payloads, concurrency limit, ModelDial policy, manifest schema, migration contract, Git tag, or GitHub Release as part of setup. ## 3. Environment Discovery Identify the runtime as native Windows, macOS, Linux, or WSL. Treat WSL as Linux with its own home and `CODEX_HOME`; do not mix WSL paths with native Windows paths. Resolve the target in this order: 1. an explicit target approved by the user; 2. the current `CODEX_HOME` environment variable when it is explicitly set; 3. the standard user Codex home, `~/.codex`. On native Windows, the standard expansion is the current user's `.codex` directory under the profile. On macOS, Linux, and WSL, it is `$HOME/.codex`. Resolve the absolute path locally but do not hard-code or publish a username. Confirm the path with the user if discovery produces multiple plausible targets. Record the OS, architecture, Codex version, Python version, source commit, target path category, and whether the target already exists. Do not print credentials or full user configuration. ## 4. Python Capability This repository requires Python 3.11 or newer and standard-library TOML support. Select `python` or `python3` based on the current environment and assign it conceptually as `` for the commands below. Run the equivalent of: ```text --version -c "import sys, tomllib; assert sys.version_info >= (3, 11); print('PYTHON_CAPABILITY_PASS')" ``` Do not install or upgrade Python. If either check fails, stop with `BLOCKED: PYTHON_3_11_WITH_TOMLLIB_REQUIRED`. ## 5. Source Acquisition The installer source must be exactly the approved RC5 Runtime Source Commit A: ```text 5ae88ff9190b31174c55a6136c0c8c8611d0b34c ``` Do not substitute `master`, `main`, `origin/master`, `latest`, a floating branch, mutable `target_commitish`, or any other commit. The commit containing this setup contract is a separate documentation anchor and is not an installer source. Git is required. After the Installation Preflight reports `Ready: YES`, create a temporary workspace outside `CODEX_HOME` and run the equivalent of: ```text git clone --no-checkout https://github.com/SuperDaddyV/codex-sol-luna-worker.git git -C checkout --detach 5ae88ff9190b31174c55a6136c0c8c8611d0b34c git -C rev-parse HEAD ``` Require `git rev-parse HEAD` to equal `5ae88ff9190b31174c55a6136c0c8c8611d0b34c` exactly. Inspect `git status --short` and require a clean checkout before installation. Confirm `scripts/install.py` declares installer `VERSION = "v4.1.0-rc5"` before running it. Do not execute installer code from a mutable branch. This contract defines no archive fallback. If the exact commit cannot be verified after the Git preflight passed, stop with `BLOCKED: IMMUTABLE_SOURCE_UNVERIFIED` rather than adding a downloader. ## 6. Codex Capability The Installation Preflight has already confirmed that the `codex` command resolves and `codex --version` succeeds. The installation also requires native custom agents, multi-agent/subagent support, a primary task using GPT-5.6 Sol, and account access to GPT-5.6 Luna at `low`, `medium`, `high`, `xhigh`, and `max`. From the verified immutable checkout, first display the capability probe plan: ```text scripts/probe_capabilities.py ``` With the user's installation authorization, run the real probe into the temporary source workspace: ```text scripts/probe_capabilities.py --execute --state /.var/setup-capabilities.json ``` The probe uses ephemeral Codex calls, ignores global user configuration, and does not edit global Codex settings. Require all five client calls to be supported. Exact echo is diagnostic; the current probe treats successful client execution as model availability. If required multi-agent behavior is absent or a required Luna effort is unavailable, stop before inspecting or writing `CODEX_HOME`; do not upgrade Codex or change authentication. ## 7. Existing State Inspection Before writing, inspect these paths under the resolved target without printing sensitive contents: - `config.toml`; - `AGENTS.md` and `AGENTS.override.md`; - `agents/`, especially the five stable Luna filenames; - `sol-luna-v4/install-manifest.json`; - `sol-luna-router/install-manifest.json`; - legacy Hook definition files and scripts; - existing v4 backup roots and any same-name files the installer would own. Classify exactly one mode: - **clean install**: no v4 manifest, no supported legacy manifest, and no unowned destination conflicts; - **existing equivalent v4**: a valid v4 manifest owns equivalent payloads; - **existing v4 upgrade**: a valid v4 manifest owns safely replaceable older payloads; - **supported legacy migration**: no conflicting v4 installation and the legacy manifest schema is exactly `3.2`; - **conflict**: non-empty `AGENTS.override.md`, same-name unowned agents, modified owned files, unsafe marker state, or conflicting user-owned agent settings; - **unsupported state**: malformed or unknown manifest, ambiguous ownership, fuzzy legacy version, or an arbitrary older installation. The presence of historical Hook files or audit bundles alone does not authorize deletion. Exact legacy `3.2` migration may remove only manifest-owned legacy Hook content. Unowned audit bundles and evidence remain preserved for review. If classification is ambiguous, stop. Do not force an unsupported state into clean install or migration. ## 8. Dry Run From ``, display the current help and then run the real dry-run against the explicit target: ```text scripts/install.py --help scripts/install.py --dry-run --codex-home --source-commit 5ae88ff9190b31174c55a6136c0c8c8611d0b34c ``` Dry-run is non-mutating. Report each stable agent action (`create`, `identical`, or `conflict`), the future artifact inventory, target platform, and conflicts. Supplement it with the read-only inspection from Section 7 to summarize expected shared-file modifications, legacy removals, preserved unowned content, and the chosen mode. Do not claim that the current dry-run output is a byte-for-byte preview of every transactional merge; the apply path performs the definitive ownership validation. Stop before apply if the dry-run returns a nonzero status, reports an unsupported platform, or reveals a conflict. ## 9. Backup Do not create a parallel backup system with `Copy-Item`, `cp`, or hand-written archive commands. The installer's apply path chooses `/backups/sol-luna-v4/`, snapshots every effective target and commit-marker path, records whether each path existed, hashes copied bytes, and verifies the inventory before writing. Capture the `backup` value returned by the installer. Treat it as the only rollback path for that transaction. If backup creation or verification fails, the installer must stop before partial installation. ## 10. Clean Install For a classified clean install, run from the immutable checkout: ```text scripts/install.py --apply --codex-home --source-commit 5ae88ff9190b31174c55a6136c0c8c8611d0b34c ``` The explicit target is mandatory. Do not pass `--validation-sandbox` for a real user global installation; that option only permits repository-local test targets below `.tmp/installer-validation/`. Accept `INSTALLED` as the expected clean-install result. Record created and modified paths and the returned backup. A nonzero exit, `FAIL`, or any reason code is a blocker; do not bypass it. ## 11. Legacy v3.2 Migration Use migration only when the read-only inspection proves that `sol-luna-router/install-manifest.json` has the exact supported schema version `3.2` and no conflicting v4 state. Run: ```text scripts/install.py --apply --migrate-v3 --codex-home --source-commit 5ae88ff9190b31174c55a6136c0c8c8611d0b34c ``` Do not accept similar, malformed, missing, or guessed versions. The migration does not contact ModelDial, convert a legacy Daily Profile or LKG, or delete unowned audit bundles. It uses the same transaction backup, validates v4 payloads and shared merges, writes the v4 manifest last as the commit marker, and then attempts independently retryable cleanup of the old manifest. Accept `INSTALLED` with completed cleanup as the normal result. `LEGACY_MANIFEST_CLEANUP_PENDING` means valid v4 content committed but old-manifest cleanup is incomplete; report it precisely and do not declare full installation complete until an idempotent authorized rerun completes that cleanup. Any pre-commit failure must roll back exactly. ## 12. Existing v4 For a valid installer-owned v4 installation, run the same apply command without `--migrate-v3`: ```text scripts/install.py --apply --codex-home --source-commit 5ae88ff9190b31174c55a6136c0c8c8611d0b34c ``` Interpret the real result: - `IDEMPOTENT_PASS`: payload and manifest are already equivalent; zero effective changes and no new backup; - `UPGRADED`: installer-owned payload changed safely; retain the returned transaction backup; - `LEGACY_CLEANUP_COMPLETED`: an already committed supported migration finished pending old-manifest cleanup; - `LEGACY_MANIFEST_CLEANUP_PENDING`: cleanup remains incomplete and must be reported; - `FAIL` with an ownership, marker, config, manifest, target, or backup reason code: stop without manual repair. Do not use `--migrate-v3` merely because historical files remain beside a valid v4 installation. An existing valid `v4.1.0-rc4` installation is an existing v4 upgrade for RC5. The only expected effective installed changes are: 1. the installer-managed Global `AGENTS.md` block; 2. `sol-luna-v4/selector.py`; 3. `sol-luna-v4/install-manifest.json`. The five Luna agents, `config.toml`, Daily Profile, LKG, and unrelated user configuration must remain byte-preserved. If the real dry-run reports any other effective change, stop before apply. The installer must create and verify its normal transaction backup before applying the three expected changes. If a valid same-day RC4 Profile already exists, RC5 must reuse it without refresh, backfill, rewrite, state migration, or ref-cost network access. Optional RC5 metadata may first appear through normal selection on the next natural Beijing calendar day. `OLD_SAME_DAY_PROFILE_FORCE_REFRESH = NO`; `STATE_MIGRATION_REQUIRED = NO`. ## 13. ModelDial / Daily Selector The installer copies the selector but does not need live ModelDial access and does not create selector state. First actual selection occurs when the installed global policy invokes: ```text /sol-luna-v4/selector.py --state-dir /sol-luna-v4/state --ensure-daily --print-role ``` The selector requires the canonical five efforts, chooses the highest score, and breaks a tie toward the lower effort. If invoked with a restricted `--supported` set, it can choose the best locally supported alternative and record capability degradation. The normal installation capability gate requires all five efforts, so the installed policy uses the complete default set. Selection occurs once per Beijing calendar day. A valid same-day Profile, including an older RC4 Profile without RC5 optional metadata, is reused byte-for-byte without a new live fetch. On a new day, the selector tries the anonymous official ModelDial API v1 endpoint, then the official full snapshot JSON, then a valid last-known-good record. A valid API response stops acquisition; sources are never merged, and v4.1 has no Radar HTML runtime fallback. First use with neither valid live data nor valid fallback fails closed with `NO_LUNA_PROFILE_AVAILABLE`. `ultra` is never allowed. Do not write a particular day's score into configuration or documentation as a permanent fact. RC5 preserves the existing `--print-role` contract and adds structured selection metadata: ```text /sol-luna-v4/selector.py --state-dir /sol-luna-v4/state --ensure-daily --print-selection ``` The saved selection is the single source for delegated Receipt suffixes. `Luna ref-cost ↓X.X%` is an optional ModelDial configuration reference-cost comparison from the same comparable batch, pricing snapshot, provider, route, and selected effort for Sol and Luna. It does not measure actual token savings, actual billing savings, subscription quota savings, or whole-task savings. If reference-cost metadata is missing, invalid, or non-comparable, omit only that suffix; selection and delegation continue normally. Append `LKG` only when the saved selection used fallback. Append `capability →` only when local capability forced a lower selected effort. Receipt generation performs no additional selector, network, state, capability-probe, or child work. ## 14. Read-only Status and Diagnostic The natural-language request `检查 Sol/Luna 状态` runs the installed selector once with `--status-json`. It is strictly read-only: ```text STATUS_NETWORK = 0 STATUS_SELECTOR_LOCK = 0 STATUS_STATE_WRITES = 0 STATUS_LUNA_SPAWN = 0 ``` Health classification must preserve these distinctions: - Profile missing → `Healthy / TODAY_SELECTION_NOT_INITIALIZED`; - Profile content invalid → `Unavailable / DAILY_PROFILE_INVALID`; - Profile read failure → `Misconfigured / DAILY_PROFILE_READ_FAILED`. A local read failure is never reported as Luna unavailable. Overall health precedence remains `Misconfigured > Unavailable > Degraded > Healthy`. The natural-language request `生成 Sol/Luna 诊断报告` uses the same single read-only `--status-json` result. Its fixed whitelist is limited to schema and generation metadata; OS, architecture, Codex, Python, installed version, and verified source SHA; manifest, Global policy, selector, agent, native-leaf, configuration, and parallel-limit status; Beijing-day selection, selected role and effort, source, fallback, capability, snapshot identifiers, and optional ref-cost status; project override status; health and reason codes; and symbolic sanitized locations. It must not expose full configuration, full `AGENTS.md`, environment variables, credentials, logs, child reasoning, real private paths, private remote URLs, exception text, or arbitrary file content. ## 15. Upgrade to the Latest Published Version The natural-language request `升级 Sol/Luna 到最新版本` means the latest published project Release, including Stable and prerelease versions. Release discovery is outside the installer. Consider only published, non-draft Releases with strict project SemVer; validate the prerelease flag, reject malformed or duplicate normalized versions, and select by SemVer precedence rather than publication time. Before any prerelease apply, show a clear prerelease / Public Beta risk notice. Resolve the immutable Release tag to an exact commit SHA, detect tag movement, check out that commit detached, and verify the installer version and setup/source alignment. Pass the verified immutable SHA to the transactional installer through `--source-commit`; never execute from a branch or mutable `target_commitish`. Preserve ownership validation, transaction backup, apply verification, rollback, uninstall, no-downgrade behavior, and zero-write idempotency. This installed-policy feature does not authorize the current setup run to replace RC5 Runtime Source Commit A. Every RC5 command in this contract remains pinned to `5ae88ff9190b31174c55a6136c0c8c8611d0b34c`. ## 16. Validation Before apply, validate the immutable source checkout: ```text -m unittest discover -s tests -v ``` After apply, perform read-only inspection and require all of the following: 1. the five expected files exist under `/agents/`; 2. every agent TOML parses, has the expected name, `gpt-5.6-luna`, matching effort, and `[agents] enabled = false`; 3. `config.toml` parses and the managed v4 block enables agents with maximum direct concurrency 3; 4. the managed v4 block exists exactly once in `AGENTS.md`, contains the installed absolute selector and state command, and includes the non-trivial-task Delegation Receipt policy without forcing delegation or extra Receipt work; 5. the selector file compiles and its `--help` command succeeds; 6. `sol-luna-v4/install-manifest.json` parses, reports `v4.1.0-rc5`, records source commit `5ae88ff9190b31174c55a6136c0c8c8611d0b34c`, and records the expected owned files and blocks; 7. installer-reported created, modified, removed, and preserved content matches the selected mode; 8. no active configuration or Hook definition still invokes the legacy Sol/Luna Router. Historical directories, trusted metadata, backups, or audit evidence may remain without being runtime dependencies. Do not delete them during validation. Do not expose unrelated configuration values in the report. The Source Commit A result of `152/152` tests and its Windows / Ubuntu / macOS CI `PASS` are repository validation only. They cannot replace the real Global upgrade or the following fresh-runtime acceptance scope, which remains unexecuted for RC5: - O1 Natural-language healthy status — `NOT RUN`; - O2 No-profile healthy status — `NOT RUN`; - O3 Degraded LKG Receipt and status — `NOT RUN`; - O4 Capability-degraded selected-effort Receipt and status — `NOT RUN`; - O5 Unavailable evidence status — `NOT RUN`; - O6 Misconfigured precedence — `NOT RUN`; - O7 Safe diagnostic report — `NOT RUN`; - O8 Latest prerelease immutable discovery and notice — `NOT RUN`; - O9 Fail-soft observability selection, status, and Receipt omission — `NOT RUN`; - O10 Fresh-session delegation and Receipt suffixes — `NOT RUN`. RC5 real Global upgrade, idempotency, and rollback readiness are a separate acceptance operation, not O9, and remain `NOT RUN`. ## 17. Fresh Session Requirement Codex loads global and project instructions when a run or task starts. Updating `AGENTS.md`, agent TOML, `config.toml`, or the selector does not make an already-open task reliable evidence of the new effective configuration. After install, upgrade, rollback, or uninstall: 1. finish the installer report; 2. fully reload Codex Desktop/App Server when appropriate for the client; 3. start a new Codex task in the intended scope; 4. perform the runtime smoke test there. Do not require an operating-system reboot unless an independently verified system issue requires it. This project itself does not require a Windows reboot. ## 18. Runtime Smoke Test In the fresh task, use a concise user-facing smoke test rather than reproducing the development G1-G7 protocol: 1. confirm the five global Luna role names are discoverable in the effective context; 2. run the installed selector command and require exactly one allowed role; 3. explicitly delegate a bounded exact-echo task through that selected native role and inspect parent-visible agent metadata; 4. give Sol a second clearly bounded task without naming a role and verify policy-based delegation when Sol judges it worthwhile; 5. confirm the Luna child has no multi-agent/delegation tools because it is a native leaf; 6. require Sol to review the child evidence and own the final conclusion. The recorded RC4 release acceptance used four fresh-task cases after a separately authorized real upgrade. It does not establish RC5 runtime acceptance. Apply the same evidence requirements when later validating an authorized RC5 installation: Runtime Case A is a non-trivial architecture or reasoning task; require zero direct children and `Sol/Luna: Sol-only · reasoning/architecture task`. Runtime Case B contains two or three independent bounded read-only checks; require actual Luna direct children, an actual role and count matching `Sol/Luna: delegated · ×`, and `parallel` only when parent-visible evidence proves execution overlap. Runtime Case C is non-trivial, non-architecture, sequential or tightly coupled work with no clean independent bounded child task; require no availability failure evidence, zero children, selector invocation only if normal execution required it, and `Sol/Luna: Sol-only · no independent bounded work`; `Luna unavailable` is forbidden. Runtime Case D exercises a real-unavailable classification only in a controlled fixture, fake `CODEX_HOME`, test harness, or non-production simulation where the normal path already exposes the failure. Do not damage or reconfigure the real selector, state, account, or Global environment to manufacture evidence. In all four cases, the Receipt is only a user-facing summary; verify child absence, presence, and availability facts with parent-visible runtime metadata. Receipt generation itself must not invoke a selector, capability probe, tool, child, network, state write, telemetry, or repository write. Do not accept Receipt text by itself as proof. The recorded RC4 release acceptance passed all four cases; each new installation or upgrade must still complete its own fresh-task smoke test before returning `INSTALL_RUNTIME_PASS`. Return `INSTALL_RUNTIME_PASS` only when all applicable checks pass. Otherwise return the exact failed step and evidence boundary. A sentinel string by itself is not proof of model, role, or leaf behavior. Static CI and repository tests are not substitutes for this fresh-task runtime check. ## 19. Rollback Rollback is available only with the exact backup path returned by the relevant successful install or upgrade. Require the path to remain under `/backups/sol-luna-v4/`, then run from the same immutable source: ```text scripts/install.py --rollback --codex-home ``` Accept only `ROLLBACK_EXACT_PASS`. The installer verifies the snapshot hashes, restores files that existed, removes transaction-created files, restores the prior target existence state when applicable, and removes the consumed backup after success. Do not hand-edit TOML or copy selected files as a substitute. Reload Codex and start a new task after rollback. ## 20. Uninstall The current public CLI supports manifest-owned uninstall: ```text scripts/install.py --uninstall --codex-home ``` Uninstall requires a valid v4 manifest. It verifies hashes for owned files and marker blocks, removes only owned content, restores surrounding user files by removing the managed blocks, and preserves unrelated agents and runtime state. If owned content changed or the manifest is absent or invalid, it fails closed. Accept only `UNINSTALLED`, then reload Codex and start a new task. Uninstall is not the same as transaction rollback: it removes the installed v4 ownership set rather than restoring a specific pre-install snapshot. ## 21. Final Report Return a concise report in this exact structure without secrets or raw user configuration: ```text 【Environment】 OS： Codex： Python： source commit： CODEX_HOME：confirmed / blocked 【Mode】 clean install / existing equivalent v4 / existing v4 upgrade / legacy 3.2 migration / blocked 【Dry Run】 status： create： modify： remove： preserve： conflicts： 【Backup】 created：YES / NO rollback path：recorded securely / none 【Installed】 status： owned payloads： 【Preserved】 user configuration： legacy evidence： 【Daily Selector】 static validation： fresh-day result： 【Observability】 status：read-only / exact blocker diagnostic：sanitized / exact blocker ref-cost：comparable suffix / omitted 【Runtime Smoke】 fresh task： result：INSTALL_RUNTIME_PASS / exact blocker 【Rollback Available】 YES / NO 【Final Status】 INSTALL_RUNTIME_PASS / BLOCKED ``` Do not claim runtime PASS before the fresh-task smoke test. If setup stops before mutation, say so explicitly. If mutation committed but runtime validation is pending, report installation and runtime as separate states.