checkpoint

Visible session state on demand. One trigger, one scannable dashboard: what you built, what's bearing down, what's left open, what's next.

🧭 Session Checkpoint

🛠️ Built
- Deploy pipeline split into build + release stages
- Rollback script with dry-run flag

⏰ Clock
- 🟠 Conference talk slides due — 5 days (2026-07-08)

🪢 Loose Ends
- Rollback script: needs one real fire on staging [VALIDATE]

🎯 Next
- Trigger a staging deploy to validate the rollback path

🏷️ Nameable as: split the deploy pipeline and shipped a tested rollback

⛵ Ship state: clean

Why this exists

Long AI working sessions accumulate state faster than working memory holds it. "Where am I, what did I just do, what's next" fires constantly — and answering it usually means scrolling back through the transcript or asking three separate questions.

/checkpoint is a renderer. It composes the answers into one fixed-format dashboard you can scan in two seconds. No new analysis, no persistence, no session close — display layer only.

It comes out of cognitive prosthetics work: built for an operator with a traumatic brain injury whose re-orientation cost is the single biggest session tax. The format that survives that test case helps everyone — the difference is degree, not kind.

What it is / is not

Is	A mid-session orientation dashboard. Ephemeral. Fixed format.
Is not	A session-close protocol, a log, a todo system, or a planner.

The ephemeral constraint is the design: the moment a checkpoint persists to disk, it starts competing with whatever you use for session closure, and the boundary between "glance at state" and "seal the session" rots.

Install

This is an Agent Skill (SKILL.md format, supported by Claude Code and a growing set of agent runtimes).

# Claude Code
mkdir -p ~/.claude/skills/checkpoint
curl -o ~/.claude/skills/checkpoint/SKILL.md \
  https://raw.githubusercontent.com/cogpros/checkpoint/main/SKILL.md

Or clone and symlink:

git clone https://github.com/cogpros/checkpoint.git
ln -s "$(pwd)/checkpoint" ~/.claude/skills/checkpoint

Quick start

In any session:

checkpoint

That's it. The agent sweeps open items, picks the single next step, reads your deadlines file, extracts the session delta, and renders the dashboard.

The Clock

The ⏰ Clock section reads hard dates from a plain text file — one YYYY-MM-DD|label line per deadline:

# ~/.config/checkpoint/deadlines.txt
2026-07-08|Conference talk slides due
2026-07-15|TLS cert renewal

Override the path with CHECKPOINT_DEADLINES_FILE. Entries within 14 days render with escalation markers (🔴 ≤3 days, 🟠 ≤7, 🟡 ≤14).

The Clock is deterministic by rule: it reads the file, never invents a date, never extrapolates one from conversation. If you mention a deadline mid-session that isn't in the file, checkpoint surfaces the gap instead of rendering it. On a prospective-memory aid, an invented calendar entry is the worst possible failure — so it's structurally forbidden, not just discouraged.

How it works

Five phases:

1. Sweep — open items (tagged [VALIDATE] / [SAVE-AFTER] / [PASSIVE] / [DEFERRED] / [FUTURE-FIRE]) + the single next step. Composes companion skills if you have them; sweeps inline if you don't.
2. Clock read — deterministic file read of your deadlines.
3. Session delta — built / audited / wired, extracted from conversation context.
4. Render — the fixed dashboard format. Empty sections drop; nothing pads.
5. Hand back — the dashboard ends the turn. No "want me to…" follow-ups.

Two details that earn their bytes:

• The Nameable line. A productive session is a session that produced something nameable; unnamed sessions dissolve into "I worked all day and can't say on what." One specific line, or an honest "Nothing nameable yet."
• Ship state is one word — clean / drifting / blocked / depleted — with a grounded rubric for each, including a rule that the agent never diagnoses depletion unprompted. When the user names depletion, the Next item drops altitude to one small winnable move.

Companion skills

Checkpoint composes; it doesn't absorb. It pairs naturally with:

• an open-items audit skill (it calls yours if present, sweeps inline if not)
• a next-step skill (same)
• your session-close protocol (checkpoint routes there if the session is actually ending)

None are required. The skill degrades gracefully to inline sweeps.

Limitations

• The Clock only knows what's in the deadlines file. Garbage in, garbage out — but never garbage invented.
• A thin session renders a thin dashboard. That's signal, not a bug.
• The fixed format is non-negotiable by design. If you want a different shape, fork it — but the uniformity is most of the value.

Versioning

• 0.1.0 (2026-04-30) — initial renderer: built/audited/wired + loose ends + next + ship state.
• 0.2.0 (2026-06-12) — Clock section, loose-end category tags, Nameable line, ship-state rubric, depleted protocol.
• 0.2.1 (2026-06-12) — slimmed to primitive register per the skill-design doctrine (strictness follows the deficit); verification checklist folded into Rules.

Planned: configurable extra count-file sources for the Clock (surface any integer state your stack exposes as a one-line presence marker).

Related

• ghost-hours — measurement framework for AI-assisted productivity, same cognitive prosthetics lineage
• ratatoskr — gated web-fetch pipeline for agent contexts
• field-report, the input half to this skill's display half. Checkpoint renders session state, field-report captures your decisions back into the agent.

License

MIT. Pollock 2026, Raven Systems Inc.