Claude Daily Retro
LynchzDEV/claude-daily-retro
Retrospectiva noturna contínua: aprenda enquanto dorme
Ver no GitHubSkill que executa uma retrospectiva automática noturna sobre suas sessões do Claude Code, analisando correções, loops de rework e melhorias via um conselho de 3 lentes (historiador, causa raiz, engenheiro de ação). Propõe ou aplica automaticamente novas skills e regras ao CLAUDE.md com deduplicação inteligente.
Instalação
npx convez-ai LynchzDEV/claude-daily-retro1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Melhorar automaticamente a configuração do Claude Code ao longo do tempo
- Revisar padrões de rework e interrupções em múltiplos projetos
- Gerar e atualizar skills baseado em comportamento real de uso
- Manter histórico de melhorias e changelog automático
Veredito

Skill sofisticado e bem-arquitetado para quem quer automação de verdade: coleta sessões, executa análise multi-lens, deuplica e aplica/propõe mudanças com safety gates. Produção-ready com retry logic, agendamento cross-platform e modo sandbox. Exigente em setup mas robusto.
README
A tradução para o português deste README ainda está na fila.
claude-daily-retro
A nightly continuous-improvement retrospective for Claude Code.
Every evening it reviews everything that happened across all of that day's Claude
sessions — every correction, interruption, repeated instruction, rework loop, and
win — then runs a 3-lens "council" over each one (like a scrum retro) and turns the
top findings into new/improved skills and CLAUDE.md rules — with a dedup gate so
it improves what already exists instead of piling up duplicates.
It's the loop that makes your Claude setup get better while you sleep.
What it does (3 steps)
- Gather — reads your local session transcripts (
~/.claude/projects/*/*.jsonl)- built-in memory across all projects for the target day, and writes a flat,
non-judgmental list of everything that happened →
~/.claude/retro/<date>/01-events.md.
- built-in memory across all projects for the target day, and writes a flat,
non-judgmental list of everything that happened →
- Council — 3 parallel sub-agent lenses (Historian / Root-cause / Action-engineer)
analyze every event: what / why / how to respond / how to prevent / what to build →
02-council.md. - Act — ranks the top 3, runs a dedup gate (improve existing, never duplicate),
and either:
- apply mode → auto-edits global
~/.claude/CLAUDE.md(in a demarcated auto-section it never lets itself overwrite your hand-written rules), bumps~/.claude/IMPR-CHANGELOG.md, updates a registry; or - propose mode → writes recommendations to
03-proposals.mdand changes nothing.
- apply mode → auto-edits global
Runs unattended via a scheduler (macOS launchd / Linux cron) at a time you pick, with a 7-day catch-up so a powered-off evening isn't lost.
Reliability: the wrapper is rate-limit aware — if the nightly run dies on a usage/session limit it parses the advertised reset time, waits (bounded), and retries up to 3×. Step outputs carry a completion sentinel, so a retry resumes at the step that died instead of redoing the whole retro. A lock dir prevents overlapping runs (RunAtLoad + schedule + manual).
No
--dangerously-skip-permissions. It runs headless with--permission-mode acceptEditsand an explicit tool allowlist.
Install
Option A — the wizard (skill and scheduler, recommended)
git clone https://github.com/LynchzDEV/claude-daily-retro.git
cd claude-daily-retro
./install.sh
The wizard asks you to choose: apply vs propose mode, run time, catch-up depth,
whether to seed history skip-markers, and the scheduler backend. Re-run it any time to
change settings. Uninstall with ./uninstall.sh.
Option B — Claude Code plugin (skill only, cross-platform)
/plugin marketplace add LynchzDEV/claude-daily-retro
/plugin install daily-retro@lynchzdev
Gives you the /daily-retro skill. Run it manually (/daily-retro 2026-01-31 propose)
or wire your own scheduler. No OS job is installed this way (plugins can't register cron).
Usage
/daily-retro # today, apply mode (or your config default)
/daily-retro 2026-01-31 # a specific day
/daily-retro 2026-01-31 propose # review-only, edits nothing
Outputs land in ~/.claude/retro/<date>/.
Modes
| Mode | What Step 3 does |
|---|---|
propose (safe default in wizard) | Writes 03-proposals.md with ranked, deduped recommendations + proposed diffs. Edits nothing. |
apply | Auto-creates/improves skills, edits the auto-section of ~/.claude/CLAUDE.md, bumps the changelog + registry. |
Start with propose, read a few nights of output, then switch to apply if you trust it.
How it keeps itself honest
- Recurrence escalation — every event is checked against the rules the retro already wrote. If an existing rule should have prevented an event, that's a recurrence: the rule failed as prose. 2+ recurrences escalates it up the enforcement ladder (rule → sharpened rule → hook).
- Deferred revisit — actions parked in the registry's
_deferredlist (blocked tooling, project-specific routing) are re-checked every run and promoted the moment their blocker clears. - Consolidation — the auto-maintained CLAUDE.md section is loaded into every session's context, so growth is capped: past 10 rules the retro merges overlapping rules and retires zero-recurrence stale ones (their text is archived in the registry, never lost).
- Output validation (0.4.0) — a day is only marked done after the scheduler verifies all
three step files end with their completion sentinel (including the new
03-applied.mdapply-proof). An exit-0 run that stalled mid-step is retried, not silently trusted. - Failure visibility (0.4.0) — non-rate-limit failures get one automatic retry, then a
desktop notification (macOS
osascript/ Linuxnotify-send). No more silently missed nights. - Pending-artifact staging (0.4.0) — when a scheduled run can't write into a project repo
(macOS TCC blocks launchd from
~/Desktopetc.), the full artifact is staged under~/.claude/retro/pending/<repo>/and a SessionStart hook lists it at the start of your next interactive session for one-step installation. - Cost hygiene (0.4.0) — days with zero transcripts are skipped without invoking Claude,
and bulky raw capture files are pruned after
RETENTION_DAYS(durable knowledge lives in the step files + registry).
⚠️ Before you trust apply mode
- It edits your global
~/.claude/CLAUDE.md— affecting every project. It only writes inside a## Continuous-improvement rules (auto-maintained by daily-retro)section and never edits your hand-written rules, but review the diffs. - It runs Claude headless with file-write permission on a schedule. Read
install.shandskills/daily-retro/SKILL.mdfirst. - It reads your session transcripts locally. Nothing is sent anywhere beyond your normal Claude Code usage.
- Token cost: quiet days (<10 events) run with no sub-agents; busy days spawn at most 3.
proposeandapplycost the same. You can pin a cheaper model for retro runs (RETRO_MODEL="sonnet"inconfig.env) to spare your premium quota — the run shares your account rate limit.
Linux: systemd timer (alternative to cron)
# ~/.config/systemd/user/daily-retro.service
[Service]
Type=oneshot
ExecStart=%h/.claude/skills/daily-retro/run.sh
# ~/.config/systemd/user/daily-retro.timer
[Timer]
OnCalendar=*-*-* 18:00:00
Persistent=true # back-fills missed runs after boot
[Install]
WantedBy=timers.target
systemctl --user enable --now daily-retro.timer
Files it manages
~/.claude/skills/daily-retro/SKILL.md the 3-step skill
~/.claude/skills/daily-retro/run.sh scheduler wrapper (catch-up)
~/.claude/skills/daily-retro/config.env your wizard choices (mode, schedule, model pin, retry policy)
~/.claude/skills/daily-retro/registry.json dedup manifest
~/.claude/retro/<date>/ per-day outputs
~/.claude/IMPR-CHANGELOG.md versioned action log
~/.claude/retro/pending/<repo>/ artifacts staged for repos the scheduler can't write
~/.claude/hooks/retro-pending.sh SessionStart surfacing hook (registered in settings.json)
~/Library/LaunchAgents/com.claude.daily-retro.plist (macOS)
License
MIT © LynchzDEV