feat: enable Codex CLI as a first-class agent alongside Claude Code (#153)

* feat: enable Codex CLI as a first-class agent alongside Claude Code

Full dual-agent support so Codex CLI users get the same Pilot Shell
experience as Claude Code users across all Console features.

Sessions: CodexSessionReader parses ~/.codex/sessions/ JSONL files,
  merged with Claude sessions in Dashboard and Sessions views with
  agent badge and All/Claude/Codex filter toggle.

Usage: Codex cost tracking via codex-parser with model-specific
  pricing, aggregated into daily/monthly charts with agent filtering.

Memories: Hook pipeline fixed for Codex (hookSpecificOutput passthrough
  instead of systemMessage, context handler enabled, no ugly display),
  all memory lifecycle hooks added (session-init, observation, summarize,
  session-end), agent column migration on sdk_sessions, observations API
  agent filter, Memories UI toggle.

Extensions: Codex extensions discovery and display alongside Claude
  extensions in the Extensions view.

Quality hooks: file_checker (with apply_patch format support),
  tool_token_saver, context_monitor (transcript estimation fallback),
  spec_stop_guard — all ported to codex_hooks.json.

Installer: codex_files step installs hooks, skills, MCP config, and
  AGENTS.md for Codex CLI. Agent runtime detection for both CLIs.

Docs: Codex CLI getting-started guide, updated hooks/extensions/MCP
  docs for dual-agent coverage.

* fix: address code review findings from xhigh + ultrareview

- PaginationHelper: use COALESCE(s.agent, 'claude') so orphan observations
  aren't dropped when agent filter is applied (LEFT JOIN → INNER JOIN bug)
- file_checker: strip trailing whitespace/CR from regex-captured paths
  to handle Windows line endings in apply_patch content
- Memories UI: wrap JSON.parse of facts/concepts in safe helper to prevent
  render crash on malformed data
- hook-command: restore suppressOutput:true in error handler else branch
- codex_hooks.json: remove no-op context hook (handler early-returns for codex)
- installer codex_files: restrict TOML key detection to top-level scope,
  catch _TomlStructureError, remove dead regexes, cleanup

* fix: address PR #153 review findings from CodeRabbit, Cubic, and Claude

- Fix CI: add command_exists early-exit to install_pbt_tools
- Fix fish shell %self not expanding in concatenation (use $fish_pid)
- Fix silent exception swallowing in codex_files skill builder
- Add CODEX_HOME absolute path validation
- Remove unreachable else branch in license_check.py
- Fix HeroSection whitespace-nowrap clipping on mobile
- Fix MCP tool naming (underscores → hyphens) in PRD skill
- Fix test_codex_rules local_mode mock (was truthy, skipping Path.home path)
- Add missing test assertions for dependency orchestration
- Update installation.md step count (7 → 8) with Codex files row
- Update codex-cli.md hooks section to match actual coverage
- Fix hooks.md count (19 → 18) to match table rows
- Add language tags to fenced code blocks in extensions.md

* fix: update Codex docs, enable mem-search MCP and memory context injection

- Installer post-steps: show all 6 Codex skills (was 3), remove Claude Code Only section
- Enable mem-search MCP for Codex: remove _CODEX_SKIP_MCP entirely
- Enable memory context injection for Codex: remove early-return in context handler,
  restore context hook in codex_hooks.json
- codex-cli.md: remove Console Views table (all ✅), update MCP list (all 7 servers),
  remove memory/context from Claude Code-only table, update hooks section
- README: update Codex skill list (all 6), remove persistent memory from limitations
- prerequisites.md: update Codex description
- FAQSection.tsx: remove persistent memory from limitations
- Test updates: add benchmark/setup-rules/create-skill to Codex skill safety tests,
  update invocation regex, fix context handler test to expect injection
- create-skill: simplify allowed-tools examples to pass Codex safety checks
- Codex rules audit: all rules properly adapted via CC-ONLY/CODEX blocks

* fix: Codex context handler returns compact stderr instead of hookSpecificOutput

Codex CLI doesn't support Claude Code's additionalContext protocol — returning
hookSpecificOutput causes 'invalid session start JSON output' error. Instead:
- Return { continue: true } (valid Codex JSON)
- Write compact one-liner to stderr: '[pilot] Memory loaded for [project]'
- mem-search MCP still available for on-demand memory lookups

* fix: ensure all skills and rules work with both Claude Code and Codex

- Add CC-ONLY/CODEX blocks to 5 Skill() calls in spec-verify and
  spec-bugfix-verify that would break the verify→implement loop in Codex
- Add orchestrator-level Codex AskUserQuestion handling to fix, prd,
  benchmark, and setup-rules skills
- Add CODEX blocks to mcp-servers.md (ToolSearch → tool_search) and
  browser-automation.md (Chrome tool detection)
- Fix wrong CodeGraph tool names in spec-implement (trace_call_path →
  codegraph_callers/callees, search_graph → codegraph_search)
- Add CODEX blocks to spec-plan step 00 and setup-rules step 01 for
  AskUserQuestion consistency

* fix: treat Codex CLI as first-class agent alongside Claude Code

- README + Docusaurus restructured: prerequisites, examples, and feature
  pages treat Claude Code and Codex as equals (drop Claude-only framing).
- Console: dashboard / sessions / memories / requirements show consistent
  agent badges; settings page agent styling matches across the app.
- Console Help page renamed to Documentation and moved below Settings;
  vim-style G-sequence shortcuts removed; sidebar ordering: Usage →
  Settings → Documentation.
- Console screenshot assets renamed help.webp → documentation.webp
  (full, _sm, thumbs) and surfaced in marketing site.
- Codex usage cost calculation fixed — parser now reads model from
  JSONL turn_context.payload.model instead of relying on the brittle
  ~/.codex/state_5.sqlite lookup. Previously all Codex sessions resolved
  to model='unknown' and total cost = $0.
- Codex skill migration: skills, hooks, and installer support both
  agents via shared .pilot/hooks layout and Codex agent skills under
  ~/.agents/skills/. uninstall.sh cleans Codex artifacts too.
- Statusline simplified to surface only Pilot Shell updates (Claude/Codex
  CLI updates removed — agents manage their own update prompts).
- Website SEO updated: index.html keywords, HowTo schema, FAQ schema,
  noscript fallback, SEO component defaults, and Docusaurus config
  (blog/RSS descriptions, metadata keywords) all include Codex/GPT-5.
- pilot/scripts and pilot/ui bundles rebuilt against the new console
  source.

* fix: split installer file steps; detect agents instead of installing them

- Add cmd_install agent prerequisite gate (EXIT 1 with README links when
  neither Claude Code nor Codex CLI detected)
- Remove install_claude_code + upstreams.yaml entry: installer no longer
  installs either agent (README contract)
- Three sibling file steps with clear separation:
  * PilotFilesStep — always; ~/.pilot/* runtime + canonical skill source +
    ~/.pilot/rules/ raw sources read by Codex adapter
  * ClaudeFilesStep — gated on is_claude_installed(); ~/.claude/{rules,
    agents,settings.json} + Claude post-install merges
  * CodexFilesStep — gated on is_codex_installed(); ~/.agents/skills/ +
    ~/.codex/*, per-category output mirroring Claude
- Fix is_claude_installed paths to launcher canonical set
  (~/.claude/local/bin/claude + Claude.app bundle)
- Unify agent detection: dependencies._init_rtk and codex_files now use
  platform_utils.is_*_installed
- PilotFilesStep switches inheritance → composition over ClaudeFilesStep
  to avoid LSP violation
- CodexFilesStep output via _CodexReport accumulator + label formatters
- Promote cache keys to named constants; remove dead
  _post_install_processing and dead _get_cached_pilot_files fallback
- Update README, Docusaurus installation guide, installer patterns rule

* fix: address codex hook and review feedback

Author: maxritter

.gitignore

@@ -243,18 +243,21 @@ benchmarks/**
 skill-creator
 spec-annotate
 changes-header
+review-output.json
 apm.yml
 .playwright-*/
 .apm
 .agents
 .codegraph
 .vercel
 .codex
+.review
 .mypy_cache
 .mcp.json
 demo
 skills-lock.json
 sx.json
+context-hud
 .env**
 playwright/
 .claude.backup.*