codex-feature-adoption

name	codex-feature-adoption
description	Review what shipped in the OpenAI Codex CLI since the last adoption analysis and file an umbrella issue (☂️) + one subissue per adoptable feature, each grounded in how Sybra actually drives Codex headless. Use when asked to "check what's new in Codex", "adopt new Codex features", or to refresh the Codex CLI adoption tracker. For Claude Code, use claude-feature-adoption instead.
argument-hint	[--from <version>] [--dry-run]
user-invocable	true
allowed-tools	["Agent","Read","Grep","Glob","Bash","WebFetch"]

Codex CLI Feature Adoption

Find what shipped in the OpenAI Codex CLI since Sybra last did an adoption pass, decide what's worth adopting (grounded in how Sybra invokes the codex CLI as a provider), and file an umbrella issue with one subissue per item.

This is the Codex variant. For Anthropic's Claude Code, use claude-feature-adoption — same workflow shape, different changelog source and grounding points.

Repo: Automaat/sybra (public personal project). Everything this skill writes is about Codex CLI public features — safe to file. Never paste work-repo content (see the Work-Data Confidentiality rule in the root CLAUDE.md); only reference Sybra's own code.

Phase 1 — Establish the baseline (what's already covered)

1. Find the last umbrella. List prior Codex adoption trackers, newest first:

   gh issue list --repo Automaat/sybra --state all --search '"Adopt Codex" in:title' \
     --json number,title,state,createdAt --limit 20
   

   The most recent ☂️ Adopt Codex CLI vX-Y features gives the floor = its upper version Y. If none exists (first run), ask the user for a starting version, or default to ~1 month of Codex releases back.

2. Read the last umbrella + its subissues so you don't re-file already-adopted items:

   gh issue view <last-umbrella> --repo Automaat/sybra --json title,body
   gh api graphql -f query='{ repository(owner:"Automaat",name:"sybra"){ issue(number:<last-umbrella>){ subIssues(first:50){ nodes{ number title state } } } } }' \
     -H "GraphQL-Features: sub_issues"
   

   Closed subissues = already adopted (skip). Open subissues = carryovers to list in the new umbrella.

3. Current installed version = the ceiling:

   codex --version    # e.g. "codex-cli 0.142.2"
   

   New range = (version after the floor) → installed — exclude the last covered release so it isn't re-processed. Override the floor with --from <version> if passed. Codex versions move fast (0.x.y) — expect a wide numeric range per month.

Phase 2 — Read the changelog for the new range only

Codex has no local changelog cache (unlike Claude Code). Fetch it from the source repo:

• Primary: WebFetch https://raw.githubusercontent.com/openai/codex/main/CHANGELOG.md. Ask for the entries between the floor and installed versions.
• Cross-check releases if the changelog lags HEAD:

  gh release list --repo openai/codex --limit 40
  gh api repos/openai/codex/releases --jq '.[].tag_name' | head -40
  

  Read individual release notes with gh release view <tag> --repo openai/codex for anything not yet in CHANGELOG.md.

Focus only on the codex exec / CLI / sandbox / config surface — ignore IDE-extension, TUI, and unrelated changes.

Phase 3 — Ground the analysis in Sybra's code

A changelog entry only matters if it touches how Sybra invokes or manages codex. Re-verify against the current tree every run (the map drifts) — spawn Explore agents in parallel, or read directly. Known integration points (verify, don't trust):

Concern	Where to look
Headless codex exec flags	internal/agent/runner_headless_invocation.go (the a.Provider == "codex" branch: exec --json --skip-git-repo-check --ignore-user-config --ignore-rules, -C <cwd>, --model)
Sandbox / approvals	codexSandboxArgs(...) in provider_codex.go; --dangerously-bypass-approvals-and-sandbox is always used for real dispatch (headless-only now) — --sandbox workspace-write only survives as a dead branch in the display-command builder
Stream parsing	internal/agent/stream.go (ParseCodexLine → CodexEvent) parses the NDJSON; internal/agent/runner_headless_stream.go (codexEventToStreamEvent) maps Codex events (thread.started, turn.started, turn.completed, item.started/item.completed, error) into Sybra's StreamEvent shape
Sessions	Codex has no --resume; sessions persist as ~/.codex/sessions/rollout-<sessionID>.jsonl (resolved by resolveCodexSessionFile in discovery.go). Watch for any new resume/continue flag.
Skills	discoverCodexSkills() + rewriteSkillInvocations() (Codex has no native skills; Sybra rewrites /skill → $ prompt injection)
Model list	internal/sybra/svc_info.go runs codex debug models (dynamic); normalizeModel maps aliases (sonnet/opus → gpt-5.5, haiku → gpt-5.4-mini)
Config	--ignore-user-config means Sybra-passed flags/env are the only config surface — note any new codex exec flag that replaces a config-file-only setting
Server / unattended	root CLAUDE.md "Server Deployment" — Codex config + hooks live in /data/sybra/codex (codex_hooks = true)

For each candidate feature, state exactly where it plugs in (file:line) and the effort. Drop anything UI-only, TUI, IDE-extension, or already adopted.

Phase 4 — Bucket the findings

Mirror the prior umbrellas (#685, #1019):

• High value — clean adopts improving headless resilience, reliability, or cost (e.g. a new resilience flag, structured-output/JSON mode, sandbox/approval granularity that lets Sybra drop blanket bypass).
• Medium — worthwhile but more effort or a behavior change (e.g. session resume/continue support, new sandbox modes, model-list/effort handling).
• Smaller — confidentiality/security/cleanup (e.g. credential-scoping, attribution, config hardening).
• Parallel groups — Group A = changes that all touch the codex branch of runner_headless_invocation.go (serialize); Group B = independent.
• Carryovers — still-open subissues from prior Codex umbrellas / the parity tracker (e.g. #256 "Codex provider parity").
• Strategic watch — large/convergent items, not drop-ins.
• Skipped — features evaluated and rejected, with the one-line reason.

Present this to the user before filing. If invoked --dry-run, stop here.

Phase 5 — File the umbrella + subissues

1. Umbrella (title prefixed with ☂️, per the global git rules):

   gh issue create --repo Automaat/sybra \
     --title "☂️ Adopt Codex CLI vX-Y features" \
     --body-file <umbrella.md>
   

   Body = the buckets above + parallel groups + carryovers (with #refs) + a ## Source line naming the changelog/release range. Write bodies to files in the scratchpad; never use -F/heredoc-to---body for long text.

2. One subissue per adoptable item. Each body: the bucket + parallel group, the Sybra grounding (file refs), concrete steps, and a Source: Codex CHANGELOG vX.Y.Z (or release tag) line. Create, then link:

   gh-add-subissue <umbrella#> <child#> Automaat/sybra
   

   (gh-add-subissue is in ~/.local/bin.)

3. Verify the links landed — gh-add-subissue occasionally drops one:

   gh api graphql -f query='{ repository(owner:"Automaat",name:"sybra"){ issue(number:<umbrella#>){ subIssues(first:50){ totalCount } } } }' \
     -H "GraphQL-Features: sub_issues"
   

   If totalCount < number created, re-run gh-add-subissue for the missing child.

4. Report the umbrella URL and the subissue list, and recommend the highest-leverage / lowest-effort starting pair.

Conventions

• Issue titles: conventional-commit style where applicable; umbrella titles lead with ☂️.
• No work-repo identifiers anywhere. No AI attribution. No #-ref noise that title hooks reject.
• Keep subissue bodies terse and grounded — a generic body ("adopt feature X") is a failure; name the file it lands in.
• Codex moves fast and breaks flags between minor versions — when a changelog entry changes existing codex exec flags Sybra already passes, that's a regression-risk item, not just an adoption: flag it High.