Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx skills add ij5a/subrosa --skill checkpointgit clone --depth 1 https://github.com/ij5a/subrosaWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/skills/ij5a/subrosa/checkpoint)<a href="https://agentmods.dev/skills/ij5a/subrosa/checkpoint"><img src="https://agentmods.dev/badge/skills/ij5a/subrosa/checkpoint/github.svg" alt="Measured on agentmods" height="20"></a>Or the 80×15 button, for a site that already has a row of RSS and ATOM ones. Only the verdict fits; the numbers stay here.
<a href="https://agentmods.dev/skills/ij5a/subrosa/checkpoint"><img src="https://agentmods.dev/badge/skills/ij5a/subrosa/checkpoint.svg" alt="Reviewed on agentmods" width="80" height="20"></a>What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5.1 | $0.00075 | $0.02538 |
| Opus 5.5 | $0.00030 | $0.01015 |
| Sonnet 5.5 | $0.00015 | $0.00508 |
| Haiku 4.5 | $0.00007 | $0.00254 |
Grade A, and why
checkpoint scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 23d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 108 lines — stays where its author put it; the contents beside it link to each section on GitHub.
checkpoint: flush memory before a context wipe
The user is about to run /clear or /compact. Save durable facts before either command, then report whether it is safe to wipe.
MEMORY.md is generated from (~/.claude/subrosa/memory.db). Never hand-edit it. Write leaf files and register facts. The generator rebuilds the byte-budgeted index. Use the subrosa CLI (subrosa fact …, subrosa generate, subrosa search …). If it is not on PATH, bootstrap installs it at ~/.claude/subrosa/bin/subrosa.
Procedure
-
Scan the full conversation. Review every turn, not only the last. Classify each candidate as one of these four types:
- user: role, preferences, knowledge, and working context
- feedback: corrections and validated approaches. Include the reason for both.
- project: ongoing work, deadlines, motivations, roles, and reasons
- reference: pointers to external systems, dashboards, and ticket projects
-
Apply the exclusion list strictly. Never save:
- Code patterns, conventions, file paths, or architecture
- Git history, commit hashes, blame info, or PR numbers
- Debugging recipes. The fix is in code, and the reason is in the commit message.
- Anything already covered by
CLAUDE.md - Ephemeral state, such as task details, investigation chains, or bare ticket numbers
- Routine activity logs. Save only what was surprising or non-obvious.
These exclusions apply even when the user says, "save this." Ask what was non-obvious or whether they meant
CLAUDE.md. -
Check existing memory before writing. Run
subrosa fact listfor curated facts. Runsubrosa search "<keyword>"for the full transcript archive. Read related leaf files. For each candidate:- A correct similar fact exists: skip it.
- A stale similar fact exists: update it in place.
- No similar fact exists: write a new one.
-
Write the leaf, then register the fact. Do NOT hand-edit
MEMORY.md.- Create or update the leaf with frontmatter:
name,description,type. - Use Why: and How to apply: for
feedbackandproject. The why helps future readers judge edge cases. - Register it:
subrosa fact upsert --leaf <file.md> --hook "<one-line hook, under ~150 chars>" --origin-session <id>.<id>is the current session id from step 10. Pass it for new and updated facts, becausecheckpoint-markrefuses a session with no fact stamped with its id. Type and title come from frontmatter. Pass--pinfor an always-loaded guardrail, but pinned facts still compete for the byte budget. On update, the stored title stays unless you pass--title. New facts append; updates keep their place. - Link related leaves with
[[slug]], using the other leaf'sname. A not-yet-written leaf is allowed. Runsubrosa fact link <slug>after registration.[dangling]means a typo or missing leaf.
- Create or update the leaf with frontmatter:
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 23d ago Changed 493376a136bf
- 28d ago Changed 4c1dc470844a
- 1mo ago Changed · +2 lines 61630bca6263
- 1mo ago First seen · 106 lines · 75 tokens per session scan A d71daff323ad
checkpoint is a skill published in the GitHub repository ij5a/subrosa (5 stars, last pushed 3d ago), licensed MIT. It adds 75 tokens to every session and 2,538 once invoked, about $0.0003 per session on Opus 5.5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-31.
Other skills, from other repositories
performance-audit
Full-stack performance health check across 12 dimensions. Rendering, bundles, assets, API/network, algorithms, memory, database, caching, Web Vitals, backend runtime, concurrency, and framework-specific pathologies. Evidence-based Impact Models with confidence tiers and a prioritized optimization roadmap. Switches…
memory-manager
Optional local conversation memory using SQLite FTS5 and the existing scoped continuity ledger. Semantic capture, bounded cited retrieval, privacy, session isolation and explicit maintenance.
mushroom
Live graph memory for agents — query and create entities, derive relationships by rule, explain why two things are associated, read the graph at a past commit, ask what a change would do, and answer as one role. Trigger on: memory, remember, recall, relationship, entity, association, why are X and Y related, as of…
memory-status
Report where memory is stored (the SQLite file), the active project scope, and entity/observation counts, or list every project in the store with its usage. Use when the user asks whether memory is working, where it is stored, how much has been remembered, what projects are stored, or how much each project takes.
summer-ratelimit
Summer Framework rate limiting (v0.2.2+ ONLY) — fixed-window, sliding-window, token-bucket strategies with Redis or in-memory storage. acquire() for auto-429, tryAcquire() for manual control. Use when implementing rate limiting in Summer Framework projects — per-user, per-IP, per-tenant, or global rate limits with…
money
Skill "money" from Hmbown/Codewhale, covering money, when to use, setup, workflow and non-goals.