[recipes] Clean up Life Engine: add state table, harden permissions, fix skill divergence

[air-captain]

Summary

• Add life_engine_state key-value table for runtime state (cron job ID, sleep schedule) — replaces the cron_state briefing type workaround with a proper system state table
• Simplify Dynamic Loop Timing from 6 tiers to 4 (15m / 30m / 60m / one-shot) for cleaner reschedule logic
• Remove duplicate embedded skill from README — Step 5 now points to life-engine-skill.md as the single source of truth, eliminating a 130-line divergent copy
• Harden permissions approach — restructure Step 6 to lead with settings.json allowlist (scoped, persistent) as the recommended default; demote --dangerously-skip-permissions to testing-only; replace broad mcp__supabase__* wildcards with specific tool names
• Fix skill-to-schema consistency — add user_responded update logic, timezone note, remove phantom suggestion_type and cron_state references
• Fix platform references — consistently say "Telegram or Discord" where both apply

Test plan

• Cross-reference all table/column/type references in life-engine-skill.md against schema.sql — all 9 QA checks pass
• Verify schema.sql creates 6 tables with no destructive SQL
• Verify README contains no embedded skill code (only pointer to life-engine-skill.md)
• Verify settings.json allowlist includes all tools Life Engine needs (reply, react, edit_message, gcal_list_events, gcal_get_event, search_thoughts, list_thoughts, thought_stats, execute_sql, CronCreate, CronDelete)
• Run /life-engine skill to confirm it executes a full cycle with the updated instructions

🤖 Generated with Claude Code

[matthallett1]

@claude

[matthallett1]

@claude summarize this PR

[matthallett1]

@claude review this PR

[matthallett1]

@claude review this PR

[github-actions]

Claude finished @matthallett1's task in 2m 12s —— View job

---

PR Review

• Gather context (CLAUDE.md, CONTRIBUTING.md)
• Read changed files
• Review for alignment, clarity, completeness, safety, value, and design patterns
• Post final review

---

Overall: ✅ Approve with minor notes

This is a solid cleanup PR that makes the recipe significantly more maintainable. The single biggest win is removing the 130-line duplicate embedded skill — that alone justifies the PR. The permissions restructure and life_engine_state table are both genuine improvements. A few things worth noting:

---

What Works Well

Removing the embedded skill duplicate is the right call. A divergent copy is a maintenance liability, and pointing Step 5 at life-engine-skill.md as the single source of truth is clean and clear.

Permissions restructure directly addresses the #1 failure mode ("Life Engine stopped responding"). Leading with the settings.json allowlist and demoting --dangerously-skip-permissions to testing-only is exactly the right hierarchy. The risk table is a nice addition.

life_engine_state table is a proper fix for the cron_state briefing type workaround. Key-value for runtime state is the right pattern here — simple, flexible, and doesn't pollute the user-facing briefing log.

Schema hygiene is good: RLS enabled, GRANTs present, indexes added, auto-update trigger for life_engine_state, verification query updated to expect 6 tables.

user_responded update logic in Rule 7 of the skill closes a gap — the self-improvement protocol was querying this field without anything writing it.

---

Issues to Address

1. life_engine_state has no user_id column — all other Life Engine tables have user_id TEXT NOT NULL, but life_engine_state is keyed only on key. This works fine for a single-user setup, but silently assumes one Life Engine instance per Supabase project. A brief comment in the schema explaining this single-user assumption (and the implication: if you want multi-user, prefix keys with your user ID) would prevent confusion. schema.sql:115

2. Skill shows Telegram-specific channel event format only — life-engine-skill.md:17 shows <channel source="telegram" chat_id="..." ...>. The PR description says "Fix platform references — consistently say 'Telegram or Discord' where both apply," but the skill still only documents the Telegram channel event format. Discord events presumably use a different source attribute. A user running Discord would be confused. life-engine-skill.md:17

Fix this →

3. delivered_via column has no CHECK constraint — life_engine_briefings.delivered_via TEXT DEFAULT 'telegram' accepts any string. With Discord now supported as a first-class option, a CHECK (delivered_via IN ('telegram', 'discord')) would prevent typos and make the column self-documenting. Minor, but easy to add. schema.sql:82

Fix this →

---

Nitpicks (no action required)

• The hardcoded coordinates in the weather curl example (latitude=45.52&longitude=-122.68) look like they need to be manually changed, but the skill correctly says to read from life_engine_state first. A comment like # replace with your actual lat/lon, or store in life_engine_state would clarify this is illustrative. life-engine-skill.md:159

• The one-shot cron behavior (recurring: false in the sleep window) is load-bearing but unexplained. A brief inline comment in the reschedule logic block would help future readers understand why the sleep window creates a one-shot vs. a recurring job.

• metadata.json version is still 1.0.0 despite this being at least the 3rd round of changes to the recipe. Consider bumping to 1.1.0 per semver.

---

Issues 1 and 2 are worth fixing before merge. Issue 3 is a nice-to-have. The rest are optional polish.
|

[matthallett1]

Good cleanup overall — the duplicate skill removal and permissions restructure are solid wins. However, Codex review (gpt-5.4) caught three issues, two of which are functional regressions:

P1: Bash missing from the unattended allowlist. The skill now uses curl for weather fetches, but the recommended settings.json allowlist doesn't include Bash. The first morning weather check will block on a permission prompt, breaking the unattended loop for users who follow the docs. Fix: add Bash(curl:*) to the allowlist example.

P1: Discord reply events ignored. The skill rule at line ~240 only fires for <channel source="telegram">. On a Discord setup, check-in replies and YES/NO approvals never match, so user_responded stays false and the self-improvement logic breaks. The README advertises Discord as first-class, but the skill doesn't support it.

P2: No upgrade migration. Changing user_id from UUID to TEXT inside CREATE TABLE IF NOT EXISTS doesn't update existing installs. Users who already have the tables need explicit ALTER TABLE ... TYPE text migration steps, or they'll get runtime errors when storing Telegram/Discord identifiers.

Happy to re-review once these are addressed!

[alanshurafa]

Review feedback from @matthallett1 has been addressed across commits 6edb477 and 26910eb:

1. P1 — Bash permissions: Broadened to Bash(*) since scoped patterns were fragile. Documented in settings.json allowlist example.
2. P1 — Discord support: Added Discord reply event matching alongside Telegram.
3. P2 — Migration: Added upgrade migration steps for existing installs (UUID → TEXT column change).

Ready for re-review.

[justfinethanku]

Thanks for the outstanding contribution. We took a final pass on the packaging, README clarity, and overall fit with OB1, and this is in good shape to land.

I’m merging this now so people can start using it. Really appreciate the work you put into this.