[d6veteran]

R1 and R3 of the lean review, re-measured on main at 0.1.20. Tokens are words times 1.33.

Saving

Read	Before (words)	After (words)	Saved
Session-start listing	2,549	531	2,018 words, about 2,700 tokens, per session
Workflows (10)	1,386	172
Contract and helper skills (15)	634	235
Agents (21)	453	90
Memory-gate rule	76	34
One memory-gate run	3,014	603	2,411 words, about 3,200 tokens, per memory
The same run, choosing "Add to CLAUDE.md"	3,014	808	2,206 words

The listing figures leave out two SessionStart lines this release does not change: the old-tool warning and the team line.

Can a plugin hide a workflow from the listing?

No, so there is no further saving. Claude Code 2.1.310 has disableModelInvocation on a workflow's command, but its plugin loader passes only name, description, whenToUse, phases and the script, and the docs name no such meta field. The 172 words stay.

What changed

• Descriptions: each workflow's meta.description is one sentence of 15 to 20 words, and the old text moved word for word to a comment under meta. Contract and helper skills are about 15 words. Agents are name and role only, set in scripts/generate.sh and regenerated.
• Gate rule: 34 words. The sentence about Claude Code's approval prompt went, because the hook's own prompt and the skill's step 4 cover it.
• Sort: moved to skills/memory-gate/sort.md, read by the gate and by step 4 of /ck:claude-md.
• CLAUDE.md steps: moved to skills/memory-gate/claude-md-change.md, read only when the User picks that choice.
• Links to /ck:claude-md: the gate used to point at its step 8 for the Trash and the merge. It now says both itself, so it has no reason to open that skill.

Reviewed by Fable

Verdict: nothing breaks a workflow or the gate. Three fixes were taken:

• draft now says its argument defaults to prd. A launch with no argument writes the PRD, so a session wanting the architecture document needs to know.
• plan-draft now says no text means the release plan.
• /ck:claude-md step 4 now says to read sort.md. It used to say "by" the file, which a session could skip.

One change was kept: in /ck:claude-md, Retire now also covers a memory that repeats another memory. Step 7 still asks the User.

Tested

• bash tests/run.sh: 495 passed, 0 failed.
• A real claude -p session on a throwaway repository, on Sonnet 5.5: the gate ran, read sort.md and not the claude-md skill, and ended with "Reply 1, 2 or 3". After "2" it read claude-md-change.md.
• Not exercised: the trial repository has no remote, so the worktree, push and draft pull request steps did not run. The Save to memory write was not run either. Those step texts moved unchanged, apart from the Trash and merge wording above.

🤖 Generated with Claude Code

[d6veteran]

ck 0.1.21: short listings and the memory gate cost summary

Lane	Model	Input	Output	Cache write	Cache read	Total
main	opus-5-5	46 · $0.00	30k · $0.60	141k · $1.13	3.0M · $0.59	3.1M · $2.32
claude-code-guide	haiku-5-5	6 · $0.00	3.0k · $0.01	117k · $0.03	161k · $0.00	281k · $0.04
general-purpose	fable-5-1	430 · $0.00	15k · $0.75	260k · $3.25	1.2M · $0.30	1.5M · $4.30
All		482 · $0.00	48k · $1.35	518k · $4.41	4.3M · $0.90	4.9M · $6.66

Prices from https://platform.claude.com/docs/en/about-claude/pricing.md, as of 2026-10-29.
Claude Code wrote no cost record for this session, so background calls are not counted.