Back to directory

agent-crop

In a curated listMaintenance: Active

oriliz/agent-crop

Records where an agent's time goes, nudges on red lines, and files retro reports back into AGENTS.md/skills from a right-sidebar panel.

View on GitHub
$ dsh plugin --profile <your profile> add agent-crop-dsh

171

stars

0

forks

Python

Language

MIT

License

2026-11-14

Created

2026-11-20

Last push

MIT-licensed local-first agent observability plugin for DeepSeek Harness (also Claude Code): measures session time, nudges drift, and pushes each retro back into instructions and skills — JSONL on disk, no upload, no scoring.

DSH integration

Native runtime

Author-claimed

Safety audit

Unaudited

Last verified

2026-11-20

License

MIT

01What can it help you accomplish?

  • Find out where an agent session actually spent its time

    A local timeline: tool name and success per call, a `stop` mark per round, `prompt` marks with `user_wait` / `idle_gap` between them, plus rework signals — shown in the session Crop tab and the cross-session Crop view

    Developers running long agent sessions who cannot tell a slow agent from their own thinking time

  • Turn each session into a small, checkable change to the agent instructions, skills or tools

    A `crop_update` self-report (phase, task, `cause` / `category`, next step) plus a work summary, with the ledger keeping what changed, why and whether it helped

    Individuals and teams doing harness engineering — treating agent improvement as a measured loop, not a vibe

  • Check whether the recorder itself has gone quiet

    `crop health --days 7` (human summary) or `crop health --json` (full `health_summary`), also exposed as the `crop_health` tool with `days` / `host` arguments

    Anyone who suspects the numbers are lying — the README says to call it when you doubt the data

02How to install into DeepSeek Harness

Installation steps

  1. 01

    Install the plugin: `dsh plugin --profile <your profile> add agent-crop-dsh`

    $ dsh plugin --profile <your profile> add agent-crop-dsh

  2. 02

    Restart dsh — the plugin takes effect in new sessions

  3. 03

    Optional: make the `crop` CLI available with `npm i -g agent-crop-dsh`, or call it ad hoc via `npx --yes --package=agent-crop-dsh -- crop status`

    $ npm i -g agent-crop-dsh

  4. 04

    To hack on the source instead: `git clone https://github.com/oriliz/agent-crop.git ~/agent-crop`, then `dsh plugin --profile <your profile> add ~/agent-crop`

    $ git clone https://github.com/oriliz/agent-crop.git ~/agent-crop

Verify the integration

  • `crop status` runs and its last line prints the real data directory
  • `tail -1 ~/.agentcrop/data/events.jsonl` shows an event with `"host":"dsh"`
  • `crop health --days 7` (or the `crop_health` tool inside a session) returns a health summary

Rollback

  • Remove the plugin: `dsh plugin --profile <your profile> remove agent-crop-dsh`
  • Uninstalling does not touch the data; to wipe it too, run `rm -rf ~/.agentcrop`

03DSH integration and capability boundaries

DSH integrationNative runtime

Installed as the `agent-crop-dsh` DSH plugin: registers the recording protocol into the system prompt and exposes `crop_update` / `crop_health` plus the Crop session tab and sidebar view

  • Passive timeline capture

    Every tool return, round end, and your next message in a session→`tool` events with name and success, `stop` marks with the round’s active minutes, `prompt` marks with `user_wait` / `idle_gap`

    one injected line of 100–300 characters per roundJSONL append under `~/.agentcrop/data`, mode 0700
  • Nudge plus AI self-report

    A session that goes quiet or crosses the yellow / red line→A nudging prompt, then a `crop_update` report with `phase`, `cause`, `category` and next step

    `category` is a closed enum of nine values — a wrong value fails the callthe self-report costs roughly 50–100 tokens when blocked
  • Two-panel view

    Recorded events of one session or of many→Session Crop tab: session clock (active / waiting), per-work lanes, nudge stickers with counts, reflections and improvements. Global Crop: five friction metrics, daily active/waiting bars, improvements, remaining pitfalls, work list

  • Health check for the collector itself

    `crop health --days 7`, `crop health --json`, or the `crop_health` tool with `days` / `host`→Valid sessions, closure ratio, must-capture rate, abnormal tools, coverage bars per event type, and a diagnosis list

  • Measured self-improvement loop

    Events from previous sessions→Measure → attribute → change instruction / skill / tool → measure again, with each retro landing as one verifiable edit in the ledger

04Who is it for? When not to use it?

Good for

  • Developers running long agent sessions who cannot tell a slow agent from their own thinking time
  • Individuals and teams doing harness engineering — treating agent improvement as a measured loop, not a vibe
  • Anyone who suspects the numbers are lying — the README says to call it when you doubt the data

Not for

  • The author calls it work in progress, and the README lists its own gaps: data coverage is low (`type` / `difficulty` often empty, so period-over-period comparisons often read “no data”), and the health thresholds (coverage ≥ 80%, field completeness ≥ 95%) are hand-set and not yet calibrated; under three sessions it only reports “too few samples”.
  • The AI-written fields are judgements, not facts. The README says self-reports are never a conclusion on their own and have to be read against the automatically captured times and failure counts.

05Compatibility, maintenance and safety notes

  • The author calls it work in progress, and the README lists its own gaps: data coverage is low (`type` / `difficulty` often empty, so period-over-period comparisons often read “no data”), and the health thresholds (coverage ≥ 80%, field completeness ≥ 95%) are hand-set and not yet calibrated; under three sessions it only reports “too few samples”.
  • The AI-written fields are judgements, not facts. The README says self-reports are never a conclusion on their own and have to be read against the automatically captured times and failure counts.
  • Small but not free: one injected line of 100–300 characters per round, and about 50–100 tokens when a stuck agent reports. The README estimates roughly 2k tokens for a 20-round piece of work.
  • This page covers the DeepSeek Harness plugin path only. The README points Claude Code users to `docs/INSTALL.md`, where the integration runs through MCP (`crop_health` is exposed next to `crop_update`); the health view separates hosts with a dsh / claude-code dropdown, and the `crop_health` tool takes a `host` argument.
2026-11-142026-11-17v0.8.4

MIT · actively maintained (latest release v0.8.4, 2026-11-18)

06Frequently asked questions

How do I install it in DeepSeek Harness?

Run `dsh plugin --profile <your profile> add agent-crop-dsh`, then restart dsh; it takes effect in new sessions. What comes from npm is a pre-built artifact, so no build script runs on your machine and you do not have to allow anything via pnpm `allowBuilds`.

Do I have to edit AGENTS.md or any config?

No. The README says the protocol is registered into the system prompt through the plugin and removed automatically on uninstall, so there is nothing to configure and nothing to add to `AGENTS.md`.

How do I check that it is actually recording?

Run `crop status` — the last line prints the real data directory — and `tail -1 ~/.agentcrop/data/events.jsonl`, which should contain `"host":"dsh"`. The `crop` command ships with the npm package but `dsh plugin add` does not put it on `PATH`: install it globally with `npm i -g agent-crop-dsh` or run it through `npx --yes --package=agent-crop-dsh -- crop status`.

Where does the data live?

Only on this machine: `~/.agentcrop/data` with mode 0700, overridable with `KIT_DATA_DIR`. Conversation text is not stored, and the README lists “no upload” as a hard boundary. The older data directory is `~/.ai-coding-efficiency`; `crop status` prints the real path if `~/.agentcrop` does not exist yet.

Can I change how hard it pushes?

Yes — open the Agent Crop row on the sidebar plugin page, hit Configure and pick a nudging style; saving takes effect without a restart. Only the default style is a real whip; the kind mentor, experienced colleague and laissez-faire parent variants are reminders. Custom reminder lines can be dropped into `~/.agentcrop/templates/voice/*.md`, and that override only applies to the default style.

08Data and sources

  • Author-claimedgithub.com57330c355cfc…

    dsh plugin --profile <你的 profile> add agent-crop-dsh

  • Author-claimedgithub.com57330c355cfc…

    被抽完,AI 要**反思与上报**——调 `crop_update` 写下 `cause` / `verdict` / `category`,这是整条链上唯一由 AI 主动补的动作。

  • Author-claimedgithub.com57330c355cfc…

    DSH 插件:工具 `crop_health`(参数 `days` / `host`),怀疑数据不可信时直接调。

This page is generated from the project’s public documentation, repository metadata and a structured parse of DSH Plugins; last verified on 2026-11-20. Found an error? Submit a correction.

🏆

Best DeepSeek Harness Plugins

Twelve plugins worth installing first — picked from the whole catalog, across every category.

DSH Plugins is an independent community directory of DeepSeek Harness plugins. Not affiliated with or endorsed by DeepSeek. Third-party plugins are not security-audited — review the source before installing.

New DeepSeek Harness plugins, weekly. No spam.