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 Green-PT/honey-for-devs --skill honey-ccrgit clone --depth 1 https://github.com/Green-PT/honey-for-devsWrote 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/green-pt/honey-for-devs/honey-ccr)<a href="https://agentmods.dev/skills/green-pt/honey-for-devs/honey-ccr"><img src="https://agentmods.dev/badge/skills/green-pt/honey-for-devs/honey-ccr/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/green-pt/honey-for-devs/honey-ccr"><img src="https://agentmods.dev/badge/skills/green-pt/honey-for-devs/honey-ccr.svg" alt="Reviewed on agentmods" width="80" height="20"></a>- NVIDIA SkillSpector pass
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 | $0.00015 | $0.00600 |
| Opus 5 | $0.00008 | $0.00300 |
| Sonnet 5 | $0.00003 | $0.00120 |
| Haiku 4.5 | $0.00002 | $0.00060 |
Grade A, and why
honey-ccr 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 11d 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 — 60 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Honey CCR
The cheapest token is the one not sent. A 500-row log read for one error line costs 500 rows of context. CCR keeps the rows that carry signal and caches the rest — recoverable by hash, so nothing is lost.
Borrowed from headroom's SmartCrusher. Lossy by design — only for data you
skim, never for payloads where every row matters (e.g. a code-review handoff
where the implementer needs all findings; use eson encode losslessly there).
When to use
- A tool returns a uniform JSON array of many records (≥ ~5 items, redundant).
- You need to read it but mostly skim — logs, search hits, metrics, events.
- The full data stays reachable, so dropping the boring middle is safe.
Don't use
- Small arrays (passes through untouched anyway) or non-array data.
- Any array where every element is load-bearing — encode it, don't crush it.
How
tool-that-spews-json | eson crush # → sampled view + sentinel; originals cached
eson retrieve <hash> # → the full original array, verbatim
crush prints the kept sample plus a sentinel
{"_ccr":"<<ccr:HASH N_rows_offloaded>>"}. When you need a dropped row, read the
hash from the sentinel and run eson retrieve <hash>.
- Cache dir:
.honey-ccr/(override withHONEY_CCR_DIR). - Below the size gate,
crushreturns the array unchanged with no sentinel — safe to pipe anything.
Always-on hook (entry-time GC)
While Honey is active, a registered PostToolUse hook applies this automatically to
Bash output — no opt-in per call. Big uniform JSON arrays (≥20 items, ≥2k chars) are
crushed; repetitive text (≥25 lines) gets consecutive repeats collapsed to ⟨×N⟩.
The original is always stashed first, and the rewritten result ends with the exact
eson retrieve <hash> command that restores it. /honey off disables the hook;
history already in context is immutable — this fires at entry, the only place a
hook can act.
Caveat: Claude Code builds affected by anthropics/claude-code#68951 ignore the
hook's Bash rewrite, making it inert (a session-start warning tells you when this
applies). Manual some-tool | eson crush piping is unaffected — prefer it there.
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.
- 11d ago First seen · 60 lines · 15 tokens per session scan A 6548342e4943
honey-ccr is a skill published in the GitHub repository Green-PT/honey-for-devs (294 stars, last pushed 3d ago), licensed MIT. It adds 15 tokens to every session and 600 once invoked, about $0.0001 per session on Opus 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 2026-08-30.
Other skills, from other repositories
memory-manage
Manage Active updates, atomic session handoffs, and human-gated Long-Term proposals. Use when asked to remember something, preserve a decision, log meaningful work, or review memory.
wrap-up
End-of-session confidence and blind-spot review. Use when closing meaningful work, before committing to action, or when the user asks what is missing or uncertain.
pn-session-handoff
End a session with a structured handoff document and two reflection questions. Captures shipped work, dangling state, resume action, and blind spots. Use via /pn-handoff before closing a session. Pairs with pn-continual-learning for AGENTS.md updates.
claude-session
Find and inspect local Claude Code conversations by session GUID, title, phrase, or recent activity. Use when the user runs /claude-session, asks what Claude Code work they have been doing lately, says a Claude Code conversation died, ran out of context, crashed, or exhausted usage, wants to resume a named…
Send WhatsApp text or media with the local wacli linked-device CLI when the user invokes /whatsapp or explicitly asks to send a WhatsApp message.
pn-context-engineering
Curate what the agent sees and when — rules vs specs vs task files vs errors vs history. Use when starting a session, switching tasks, output quality drops, or the agent ignores project conventions.