Codex SDLC Wizard Roadmap

Current State

• codex-sdlc-wizard@0.7.34 and v0.7.34 are the current published release; 0.7.35 is the Desktop-readiness release candidate
• npm trusted publishing is configured and the GitHub release workflow is now proven for real OIDC publish
• the repo now ships a skills-only Codex plugin (.codex-plugin/plugin.json pointing to skills/), one installer skill at skills/codex-sdlc-wizard/SKILL.md, and the installer/setup adapter (install.sh, setup.sh)
• the plugin declares no MCP servers, apps, or plugin-level hooks; repo-scoped enforcement remains the responsibility of the adaptive installer
• the npm CLI now defaults to adaptive interactive setup instead of requiring an explicit setup subcommand for the main human path
• setup now layers deterministic scan plus live Codex gpt-5.6-sol / high refinement when available
• setup now keeps detected values automatically, asks inferred values conversationally, and asks only missing core repo facts directly
• the repo-scoped Codex discovery bridge for $sdlc is now part of the shipping path
• consumer-path hardening for auth-heavy boundaries, capability detectors, and docs-strong scaffold repos is shipped
• honest Codex architecture guidance, confidence/reporting guidance, direct-issue capture, and repo-focus rules are now part of the shipped path
• the model-profile toggle is now shipped as a user choice:
  • mixed: gpt-5.6-terra medium main pass + gpt-5.6-sol review with an explicit high effort override
  • maximum: gpt-5.6-sol / high throughout
• researched GPT-5.6 reasoning policy now uses high as the consumer agentic-coding default, adapts xhigh escalation scopes to detected repo risks, preserves this repo's measured Sol xhigh maintainer exception, and keeps Max/Ultra as explicit escalations instead of defaults
• setup/install now offer issue-ready feedback for obvious wizard-level failures instead of only failing vaguely
• install/setup/update now use maximum as the standing Sol high default; mixed is an experimental explicit opt-in that update preserves only when already selected
• setup/update guidance now treats verification as diagnostic for product failures and stops before editing application code or application tests without explicit user consent
• setup/update guidance now tells users to exit and reopen Codex after hook/skill repairs, without rerunning setup/update just for that restart
• install/setup/update now write and repair repo-local .codex/config.toml model keys for the selected profile, while preserving unrelated MCP, sandbox, approval, and custom config
• first-run live setup now defaults to plain codex after bootstrap and requires an explicit full-trust choice to start that setup handoff with codex --dangerously-bypass-approvals-and-sandbox
• first-run handoff now uses a clearer prompt, recommends model-explicit codex resume -m ... -c ... for interrupted handoffs, and avoids the deprecated Windows shell:true plus args launcher path
• first-run live Codex handoff now runs as a managed child process with opt-in timeout cleanup, POSIX signal forwarding, process-group termination, repeated-interrupt handling, and explicit retry/resume guidance
• setup/install output now prints Codex's canonical full-trust flag (--dangerously-bypass-approvals-and-sandbox) for users who normally say yolo-style sessions, while keeping full-trust distinct from historical full-auto wording
• update guidance now frontloads the npm version boundary: $update-wizard repairs repo artifacts, while npx codex-sdlc-wizard@latest update consumes the newest package
• setup guidance now includes Codex Desktop handoff notes for auth-heavy browser/computer-use setup flows
• generated setup docs and shipped skills now include a task-routing gate that identifies CLI, Desktop/computer-use, browser automation, or human-only lanes before giving execution steps
• generated setup docs now include a demo runtime claim gate so demo-ready claims must prove the real human-facing runtime, action runner, proof status, live artifact, mutation gates, and not-claimed boundary
• setup guidance now includes Microsoft 365 auth-lane proof rules for tenant-bound Graph PowerShell and fallback OAuth evidence
• sponsor metadata is now shipped for GitHub Sponsors and npm funding surfaces
• the package now treats $sdlc as the single canonical public workflow entrypoint, keeps the Codex display name lowercase, and blocks legacy $codex-sdlc or imperative /sdlc wording from returning
• setup/install now keep $sdlc repo-scoped, install no extra repo-scoped lifecycle skills by default, and install only global helper skills, avoiding same-name global/repo skill collisions
• repo-root skill installs now keep bundled helper definitions under non-discoverable SKILL.template.md names and materialize SKILL.md only in intended destinations, preventing recursive duplicate skill discovery
• setup now detects Playwright MCP browser tooling/profile policy and documents explicit opt-in isolation versus shared persistent auth-heavy flows without rewriting .mcp.json
• setup/update now repair stale platform-specific hook wiring and install universal Node hook entrypoints so a checked-in .codex/hooks.json does not flip between macOS Bash and Windows PowerShell commands
• setup/update now write [features].hooks = true, migrate deprecated [features].codex_hooks config, and remind users to review pending repo hooks through /hooks
• generated Node hooks now use .cjs entrypoints so consumer repos with "type": "module" do not break on CommonJS require
• Codex CLI 0.144.0+ is required for GPT-5.6 profiles; its hook surface is recognized, and the wizard intentionally installs SessionStart, PreToolUse, PreCompact, and PostCompact while leaving PermissionRequest, PostToolUse, UserPromptSubmit, and Stop unwired until a proven SDLC need exists
• compact lifecycle hooks now preserve SDLC carry-forward context around Codex compaction without blocking normal compaction
• update now repairs legacy .js hook commands and stale .js hook manifest entries, including old matching files
• the git guard is now proof-aware: fresh reviewed SDLC proof allows commit/push, while missing, stale, cross-repo, or mismatched-workdir proof still blocks
• public install/README/skill copy now keeps unreleased future workflow labels out of handoff text
• the repo now ships a consumer bug-report template for install/setup/runtime failures
• the public README now leads with the real @latest adaptive setup path and keeps the top section consumer-focused
• the public README now has consumer-parity sections that explain why to use the wizard without exposing later ecosystem branding
• the official Codex plugin distribution boundary is documented in README and ROADMAP: ChatGPT Work, the Codex desktop app, and Codex CLI share plugin distribution, while ordinary Chat does not support plugins
• maintainers can run node scripts/run-proof-suite.cjs for bounded parallel release proof without dropping any checks, with --serial available for debugging
• benchmark and pilot-rollout ledgers now exist so model/default-use decisions can be measured, not guessed
• release, packaging, npm, skill, setup, adapter, update, and E2E tests are green when the parity merge is complete
• bare npx codex-sdlc-wizard@latest now auto-runs the update/check-repair path in already-initialized clones, so cross-machine checkouts sync without remembering separate check/update commands
• setup now supports optional --goals generation for a manifest-tracked GOALS.md active-scope contract, while ROADMAP.md remains backlog/history
• README and generated GOALS.md now document manual Codex /goal usage as SDLC-backed active work anchored to $sdlc, confidence/verification gates, and clean-break commits; programmatic /goal automation remains unassumed
• setup/check now reject unknown arguments before mutating or inspecting the current directory, so mistyped flags do not silently operate on the wrong repo
• upstream sync has been reviewed through agentic-ai-sdlc-wizard / claude-sdlc-wizard v1.73.0; Codex-relevant workflow hardening was ported, while Claude-only precompact hooks and research churn remain intentionally out of scope unless they prove reusable here

Next Release Cycle

0.7.35

Purpose: harden plugin-driven installation before final supported-surface validation and public directory submission.

Scope:

• preserve unrelated host hooks while replacing only wizard-owned hook entries during install and repair
• keep successful artifact installation successful when the optional nested Codex handoff cannot start or exits with an ordinary failure
• refresh manifest hashes only for installer-touched artifacts and preserve untouched customizations
• distinguish plugin exploration, install/repair, and repo-local lifecycle intents in the picker
• validate the exact release candidate through Claude-driven black-box Codex Desktop and ChatGPT Work E2E before publishing

Tracker Cleanup

The 0.7.35 stabilization set is intentionally limited to #55, #56, and the picker-intent fix from #69/PR #70. Issue #66 owns the final supported-surface E2E and public directory submission; remaining docs/research issues stay outside this release lane.

• open a new issue only when pilot consumption exposes another proven reusable wizard bug
• avoid speculative backlog churn while 0.7.35 is being validated on supported Codex surfaces

Remaining Backlog

After 0.7.35, the main backlog is:

• complete supported-surface plugin validation and submit the public directory listing
• any new reusable wizard fixes discovered during the pilot set
• model-profile measurement data collection for mixed vs maximum
• optional Skill Creator maintenance research after the active backlog stays under control

Official Codex Plugin Distribution Plan

Official Codex docs make plugins the installable distribution unit for reusable skills, apps, MCP servers, and presentation assets. The implementation outcome for this repo is a skills-only plugin: .codex-plugin/plugin.json points to the conventional skills/ directory and its single installer skill, while npx remains the repo-mutation engine.

• Package the manifest in codex-sdlc-wizard@0.7.35 so the same artifact carries the plugin skill and the existing CLI/installer.
• Validate discovery in ChatGPT Work, the Codex desktop app, and Codex CLI. Ordinary Chat remains outside the supported plugin surfaces.
• Keep MCP servers, apps, and plugin-level hooks out until a proven product need exists; the installer continues to write repo-local hooks and config.
• The official submission portal is live; the public listing is pending supported-surface validation and submission.
• Do not imply official OpenAI endorsement unless the plugin is actually accepted into the official Plugin Directory.

Working Order

1. Validate the skills-only plugin in ChatGPT Work, the Codex desktop app, and Codex CLI
2. Publish the version-aligned npm package and verify clean installation
3. Submit the public listing through the live submission portal after supported-surface validation
4. Keep pilot rollout and stabilization patches tied to real consumption bugs

Default-Use Gate

Before calling this the default Codex SDLC path, prove it on real pilot repos instead of just repo-self-tests.

• run released builds on 3-5 pilot repos before broadening the default-use claim
• require pilot success >= 95% before default use
• allow no more than 1 reusable wizard bug across the pilot set
• track the pilot set in benchmarks/pilot-rollout.csv
• summarize the gate with bash scripts/summarize-pilot-rollout.sh

Later Research

The creator-tool investigation is complete for the distribution decision: Plugin Creator established the minimum valid package shape and led to the implemented skills-only plugin. Skill Creator remains a possible later maintenance aid rather than a release blocker.

• investigate programmatic /goal automation only if Codex exposes a stable CLI/API path; keep manual /goal guidance anchored to $sdlc
• evaluate Skill Creator later as a possible aid for skill-structure maintenance
• keep the implemented Plugin Creator manifest contract covered by packaging tests
• run an experimental explicit opt-in measurement of gpt-5.6-terra medium for the main working pass with a gpt-5.6-sol review explicitly overridden to high, and compare that against the Sol high maximum profile
• keep an easy toggle between the two explicit profiles:
  • mixed: gpt-5.6-terra medium main pass plus gpt-5.6-sol review with an explicit high effort override
  • maximum: gpt-5.6-sol / high for the whole slice
• require a sample of 20 slices before considering whether mixed should stop being experimental and become a normal-work recommendation
• numeric target for recommending the mixed mode: at least 95% end-to-end success, follow-up rate <= 10%, and at least a 15% improvement in cycle time versus maximum
• separately measure Sol high vs Sol xhigh before lowering this repo's maintainer exception; preserve this repo's current xhigh baseline until Sol high meets the same sample, success, follow-up, and improvement gate
• keep Max as a single-task reasoning escalation and Ultra as a subagent-backed parallel-work escalation; most tasks do not need Max or Ultra, and neither becomes a default wizard profile without separate evidence
• escalate abstract, complex, security-sensitive, or high-blast-radius slices to xhigh when high leaves unresolved risk; do not make every consumer task pay that cost
• keep this behind the active workload so it does not compete with the active pilot-rollout and stabilization backlog