Porting OMP productivity mechanisms to Grok — AFK, suggester, retro

OMP extensions get ExtensionAPI: sendUserMessage({ deliverAs: "steer" }), status bar labels, setTimeout timers inside the session process. Grok gets lifecycle hooks, skills, and /loop — no steer API, passive hook stdout often ignored. Four mechanisms were ported from OMP/Claude Code to Grok on the personal Mac (2026-06-21). Timestamps failed; three others shipped with degraded but usable parity.

What OMP has that Grok lacks

Implication: ports are hook + skill + rules architectures, not line-for-line extension translations.

Timestamps — the port that didn’t work

Goal: OMP chat-timestamps inline prepend ([Jun 21, 05:45:53 PM] your message) plus lifecycle banners from claude-code-timestamps.

What we tried:

• timestamps.py on all CC lifecycle events, wired via ~/.grok/hooks/timestamps.json + Claude-compat settings.json
• UserPromptSubmit → hookSpecificOutput.updatedPrompt with prepended timestamp
• PostToolUse / Stop → systemMessage banners

What actually works on Grok:

• [ui] show_timestamps = true in ~/.grok/config.toml — scrollback times on messages (native, visible)
• Hook scripts run (annotations in hook log) but stdout is ignored on passive events per 10-hooks.md
• updatedPrompt on UserPromptSubmit does not alter scrollback display in Grok Build/Cursor sessions (verified in grok-ts-test)

Verdict: Use native /timestamps for scrollback. Abandon inline prepend and lifecycle banners on Grok until the harness exposes input-transform or MessageDisplay. Keep timestamps.py for Claude Code and OMP-compat wiring only.

AFK — works (skills + state + PreToolUse deny)

OMP: agent-afk-omp — /afk sets status bar AFK 🔴, /back clears; in-memory afkActive flag.

Grok port:

hooks/afk.py                    # state: ~/.claude/afk-state/<session_id>
~/.grok/hooks/afk.json          # PreToolUse deny git push / gh pr when active
skills/afk, skills/back         # engage/disengage via python CLI
rules/afk.md                    # policy when state file exists

Why it works: AFK only needs (1) durable session state and (2) blocking PreToolUse — Grok supports {"decision":"deny","reason":...} the same way as Claude Code. No steer injection required.

Trade-off: No status-bar indicator; python3 ~/.claude/hooks/afk.py status or /back exit code is the signal.

Suggester — partial parity (PreToolUse deny for visible nudges)

OMP: agent-suggester-omp — setTimeout + sendUserMessage({ deliverAs: "steer" }), editor draft save/restore, /suggester-{pause,resume,now}.

Grok port:

hooks/suggester.py
~/.grok/hooks/suggester.json    # SessionStart, PostToolUse, PreToolUse, SessionEnd
~/.grok/agent/agent-suggester.json
skills/suggester-{pause,resume,now,arm,tick}
rules/suggester.md

Timer: PreToolUse checks elapsed intervalMs since last_fire — fires on the next tool call after the interval, not wall-clock mid-turn like OMP setTimeout.

Injection (v2): PreToolUse deny with 🔄 Suggestion text — same visibility pattern as retro and AFK. Replaced an earlier Stop + additionalContext path that logged fires but was invisible in Grok Build/Cursor.

skipIfIdle: PostToolUse increments tool_call_count; zero since last fire → skip (same semantics as OMP).

Optional Grok TUI layer: /suggester-arm → scheduler_create or /loop 10m /suggester-tick for wall-clock fires during idle gaps.

Evidence (2026-06-21):

• /suggester-now fired स्थितप्रज्ञ item with exact wrapWith; log shows session_start — armed timer.
• Timer fire at 20:32:50 via Stop path — logged in agent-suggester.log but not visible in the session UI.
• Fix in 52ade8e: moved due nudges to PreToolUse check block; verified decision: deny + suggestion text blocks the next tool call.

Retro — shared script with Claude Code, weaker Grok delivery

OMP: agent-retro-omp — clock-aligned setTimeout, activity summary, rotate picker across 7 retro lenses.

Claude Code port: retro-tracker.py — PreToolUse deny at boundary (blocks next tool until retro runs); Stop + asyncRewake + exit 2 for session-end retro. See clock-aligned retro in Claude Code.

Grok port: same retro-tracker.py, ~/.grok/hooks/retro.json:

SessionStart → start
PostToolUse  → tick (async)
PreToolUse   → check block   # deny when boundary passed + activity
Stop         → check         # additionalContext fallback (no asyncRewake)

Grok vs CC on Stop: CC wakes a new turn via asyncRewake. Grok detects GROK_HOOK_EVENT=stop and emits additionalContext instead — passive, may be low-visibility.

Enforcement: rules/retro.md — mandatory 6-question response when Segment: HH:MM → arrives. Knowledge gem Q5 requires a live web_search; list every search URL in the retro reply.

Manual fallback: /loop 10m /retro-core

Evidence (2026-06-21, Grok Build/Cursor):

• Armed 20:22:33; boundary 20:30:00; 10 tool calls in segment.
• Fired 20:31:36 on first PreToolUse after boundary (not at :30 exactly — event-driven).
• Log: fired: 10 calls in segment | mode=check block
• Deny text: Segment: 08:22:33 PM → 08:31:36 PM (9min), 10 calls (1 Glob, 1 Grep, 4 Read, 3 Shell, 1 Write) + Core Retro prompt.
• Hook annotation: global/settings:pre_tool_use[1].hooks[0] blocked tool Shell
• Next boundary reset to 20:40:00.

Parity summary

Portable config: claude-config/grok/ — not a separate repo

Question: Do we need a standalone grok-config repo like claude-config?

Answer: No. Hook scripts, skills, rules, and JSON configs (agent-suggester.json, retro.config.json) are harness-agnostic and already live in claude-config. What Grok adds is thin and machine-specific:

A second repo would duplicate the shared scripts. Instead, claude-config/grok/ holds templates + install.sh:

git clone git@github.com:ankitg12/claude-config.git ~/ghq/github.com/ankitg12/claude-config
cd ~/ghq/github.com/ankitg12/claude-config
./grok/install.sh          # renders ~/.grok/hooks/*.json, symlinks hooks/skills/config

install.sh detects python3/python, substitutes and into templates, symlinks ~/.claude/hooks → repo hooks/, and links Grok skills. Override with env vars on Windows:

PYTHON=python HOOKS_DIR="$HOME/.claude/hooks" ./grok/install.sh

Reload: /hooks then r. Verify: grok inspect → Hooks > 0.

Windows (BLRANGAUR): merge claude-config/settings.hooks.json into settings.json; PYTHON=python; ghq path may be ~/repos/... not ~/ghq/....

Design rule for future Grok ports

1. If the feature needs to block an action → PreToolUse deny (AFK, retro, suggester) — ports cleanly.
2. If the feature needs to inject a user turn mid-stream → no clean Grok equivalent; use /loop, scheduler_create, or accept turn-boundary Stop injection.
3. If the feature needs to transform display text → wait for harness API; hooks won’t do it today.

Repo: github.com/ankitg12/claude-config (hooks, skills, rules, grok/ templates, retro.config.json).

Related posts

• AFK toggle for AI coding agents — OMP original
• agent-suggester-omp — OMP original
• Clock-aligned retro in Claude Code — CC hooks version
• claude-code-timestamps — works on CC, not Grok banners