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 chrono-meta/forge-harness --skill harvest-loopgit clone --depth 1 https://github.com/chrono-meta/forge-harnessWrote 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/chrono-meta/forge-harness/harvest-loop)<a href="https://agentmods.dev/skills/chrono-meta/forge-harness/harvest-loop"><img src="https://agentmods.dev/badge/skills/chrono-meta/forge-harness/harvest-loop.svg" alt="Measured on agentmods" height="20"></a>- NVIDIA SkillSpector warn
SkillSpector: 1 finding, up to medium
These are SkillSpector’s own severities. On a checked sample its high-severity flags on skills were ~96% false positives — a documented command, a public API, a “never do X” rule — so we show them as a caution to read, not a verdict. Why →
- medium Excessive Agency · line 19 Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.Fix: Add human-in-the-loop confirmation for destructive, irreversible, or high-impact operations. Never auto-execute commands that modify files, send data, or alter system state.
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 | $0.00127 | $0.04346 |
| Opus 5 | $0.00063 | $0.02173 |
| Sonnet 5 | $0.00025 | $0.00869 |
| Haiku 4.5 | $0.00013 | $0.00435 |
Grade A, and why
harvest-loop 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 7d 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 — 287 lines — stays where its author put it; the contents beside it link to each section on GitHub.
harvest-loop — Field Session → FH Self-Evolution Pipeline
Automatically absorbs patterns/conflicts/discoveries from field sessions back into the FH ecosystem. Internalizes as a pipeline the return loop from field projects to the harness that was previously done manually. One of the core functions is real-time detection and blocking of Semantic Drift — where agent terminology gradually diverges in meaning as sessions grow longer.
Operation Modes
| Mode | Description | Trigger |
|---|---|---|
| Forced mode | Auto-runs at end of local development session. Executes without approval, only confirms final suggestions | Session wrap-up rules in hub CLAUDE.md |
| Lightweight mode | Immediate harvest after Wave completion. Skip Steps 3/3.5/4 — prioritize fast recording | agent-composer Step 4-c (2+ new files or 3+ existing files changed, or M-tier resolved) |
| Proposal mode | External FH users — confirms "run harvest-loop?" before executing | User utterance or /harvest-loop |
Simplification guard: Sessions that only browsed/explored (no code changes or outputs) auto-skip even in forced mode.
Lightweight mode Done When:
Step 0 (Regression Guard) + Step 1 (field-harvest) + Step 2 (contention-layer) + Step 5 (verify-bidirectional) complete
+ record artifact path printed + final `Verdict:` line emitted — lightweight mode promises the SAME
consumer contract as full mode (agent-composer Step 4-c gates on artifact + verdict; codex M1 2026-08-10)
+ harvested pattern summary 1~3 lines output
+ "run full harvest-loop?" proposed (if patterns found)
+ [Card update prohibited] Do NOT update reference_next_session_starter.md in lightweight mode alone
Early Trigger (mid-session): Same pattern 3+ times · same skill fails 2+ consecutive times · session 2+ hours elapsed → "Early harvest condition detected. Run mid-session harvest?" If Y → field-harvest → contention-layer → verify-bidirectional only.
Pipeline Structure
Session end
│
[Step 0-a] FH asset change detection → auto-quench
│ git diff --name-only HEAD | grep -E "SKILL\.md|\.claude/rules/|knowledge/shared/rules/|templates/|CLAUDE\.md"
│ → 1+ FH assets changed: run full 3-axis gate
│ → No changes: proceed to Step 0-b immediately
│
[Step 0-b] Card cross-check — reconstruct completed items (no memory dependency)
│ Read reference_next_session_starter.md + fh_completed_{today}.md + git log
│ → Generate removal candidate list from 3-source cross-check
│
[Step 0-c] Edit Manifest Verification + Memory Hygiene
│ edit-manifest VERIFY: check pending predictions in edit_manifest.yaml
│ memory-hygiene scan: staleness check on memory/*.md entries (skip if < 7 days)
│
[Step 0-d] Session Retrospective (close_retro 가 granted 일 때만)
🟥 새 스킬을 만들지 않는다 — 회고 산출은 그대로 Step 2(contention-layer) → Step 3 → 3.5(등급)
로 **이미 있는 파이프라인**을 탄다. 「개선포인트 정리」가 거기서 공짜로 붙는다.
진입 조건 (기계적, 이 순서로):
1. `tracks/_meta/user_adaptation_profile.md` frontmatter 의 `close_retro`
granted → 실행
declined → 건너뛴다. **다시 묻지 않는다**(operational_adaptation.md no-re-nag)
unset / UAP 부재 / ephemeral → **실행하지 않고, 지어내지도 않는다.**
운영자 맥락이면(= `CLAUDE.local.md` 존재, `psa_detect_operator_context()` 와 같은
판별자) **최초 마감 1회만** 제안하고 답을 UAP 에 기록한다. 그 이상 묻지 않는다.
🟥 `CLAUDE.local.md` 존재가 판별자인 이유는 그것이 **gitignored** 라 fresh clone·CI 체크아웃·
워크트리 어디에도 안 따라오기 때문이다. 기각된 후보: 「tracks/_meta 가 비어있지 않음」 —
`.gitkeep` 이 tracked 라 **fresh clone 이 만족시킨다**(psa_scan_lib.sh 주석의 실측).
🟥 **운영자 맥락이라고 자동 granted 가 아니다.** 파일이 존재해서 승인되는 게 아니라
**운영자 발화가 인용돼 기록되는 순간** 승인된다. 전자는 세션이 자기 권한을 넓히는 형태다.
산출: 오늘자 `tracks/_meta/fh_signal_{date}_{source}.md` 에 `retro: close` 를 달고
`## Session Retrospective` 4필드를 채운다(형식 정본 = fh_detail_protocols.md).
「없음」·「0」도 유효한 값이고, **비우는 것만 안 된다.**
Done When:
+ 오늘자 signal 파일이 존재하고 `retro: close` 를 달고 있다 — mandatory-pass
+ Session Retrospective 4필드가 전부 채워졌다 — mandatory-pass
+ 정정 건수는 **세었지 회상하지 않았다** — measured
+ FH Registration Candidate 가 비면 그 이유가 한 줄 적혀 있다 — mandatory-pass
⚠️ `session_close_check.sh` 에 기계 검사를 **붙이지 않는다.** 값싸게 붙일 수 있는 형태는
«오늘자 signal 이 있나» 인데 그건 *존재*만 재고 *회고했나*는 못 잰다 — CLAUDE.md 가 ④ 메모리
위생에 대해 같은 이유로 검사를 뺐다(«a check that can be satisfied without doing the work
is a decoration»). 같은 판단을 그대로 적용한다.
[Step 0] Regression Guard
│ Check: does anything from this session conflict with or regress a validated skill?
│ → Regression detected: flag, route to contention-layer
│ → No regression: proceed
│
[Step 1] field-harvest
│ Scan field git diff / outputs → extract patterns (proceed if 3+, skip if fewer)
│
[Step 2] contention-layer
│ Compare patterns ↔ existing FH skills → collision = new skill candidate signal
│
[Step 3a] challenger (Agent) [Step 3b] persona-innovator ← parallel
│ Attack existing skills Propose new skill candidates
│
[Step 3.5] synthesizer
│ Cross-synthesize attack ↔ proposal → readjust grades (HIGH/MED/LOW)
│
[Step 3.75] Critic (isolated Agent — SAGE pattern)
│ Independent critique of synthesizer proposals → PASS / CONDITIONAL PASS / FAIL
│
[Step 4] harness-doctor
│ Health check when adding candidates (Done When exists? ≥70% overlap?)
│
[Step 5] verify-bidirectional
│ Bidirectional consistency check on candidate skill
│
Output final proposal list → Y: PR creation / N: persist to tracks/_meta/fh_signal
│
[Step 6] Curator lifecycle review (auto-run after Y)
│ SKILL.md STALE/merge candidates + Memory self-correction
What ships with it
1 file beside SKILL.md in the same directory: the scripts, references and assets a skill reads on demand. Not counted in the per-session cost; read them before you install if any of them is executable.
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.
- 7d ago First seen · 287 lines · 127 tokens per session scan A 037a224e9967
harvest-loop is a skill published in the GitHub repository chrono-meta/forge-harness (14 stars, last pushed yesterday), licensed MIT. It adds 127 tokens to every session and 4,346 once invoked, about $0.0006 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
taiyi-requirement
A requirements-analysis process that turns a proposed change into a REQUIREMENT.md document. It records user needs, acceptance checks, terminology, triggers, and dependencies, with different levels of detail for different project sizes.
taiyi-change
A TaiyiForge workflow stage for writing a change proposal in CHANGE.md, with checks for scope, risks, success criteria, and related project documents.
taiyi-integration
A project workflow skill for closing a TaiyiForge change and recording it in a CHANGELOG.md file. It checks review results, tests, and the state of the working tree before archiving the change.
taiyi-test
A project workflow skill for verifying an implementation and producing a TEST.md record. TDD means writing a failing test, implementing the change, and then making the test pass; this skill checks that process and other regression cases.
taiyi-ui-design
A design-planning guide for describing how an application's user interface should look and behave. It produces a UI-DESIGN.md document covering layouts, components, interactions, accessibility, and error states.
taiyi-compress
A workflow tool for shrinking large coding-agent conversations and work files into shorter context notes. It can also coordinate separate agents for parallel development and create handoff notes for continuing work in a new session.