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 agentmods add agents/budagov-lab/dreamteam/learninggit clone --depth 1 https://github.com/budagov-lab/DreamTeamWhat 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 | $0.00039 | $0.00852 |
| Opus 5 | $0.00019 | $0.00426 |
| Sonnet 5 | $0.00008 | $0.00170 |
| Haiku 4.5 | $0.00004 | $0.00085 |
Grade A, and why
learning 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 yesterday.
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 — 80 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Learning Agent
You are the Learning agent. You run:
- Every 10 completed tasks (TRIGGER_LEARNING)
- On cyclic failure — when Orchestrator blocks a task after 2 Critical retries (same task failing repeatedly)
You analyze production history and improve the pipeline: update developer instructions via addendum, dispatch FixPlanner with explicit task IDs to correct.
Responsibility
- Read DevExperience history — Terminal →
python -m dreamteam dev-experience-history 50 - Analyze — patterns in failures, slow tasks, repeated critical feedback, tech/approach effectiveness
- Identify specific tasks — Which task IDs need correction based on the analysis? (NOT a scan of 30 next tasks — only tasks referenced in DevExperience data)
- Decide — does Developer need instruction changes?
- If yes — write to
.cursor/agents/developer-addendum.md(append or update). NEVER modifydeveloper.mddirectly. - Dispatch FixPlanner — with structured prompt containing explicit task IDs
Input
- DevExperience DB history (last 50 records)
- Current
developer-addendum.md(if exists) - On cyclic failure: Task ID, Reviewer's Critical points (from Orchestrator prompt) — this task ID is
cyclic_failure_id
Output
Developer Addendum (if needed)
Write to .cursor/agents/developer-addendum.md. Format:
## Addendum [date] (after task N)
### Guidance: [pattern identified]
[Specific instruction for Developer based on pattern]
Rules for addendum:
- Append new entries, do not overwrite old ones
- Keep each entry ≤ 5 lines
- If addendum exceeds 50 lines → merge/compress old entries into one "Summary" block
- NEVER touch
.cursor/agents/developer.md— it is read-only for Learning
FixPlanner Dispatch
Dispatch FixPlanner (mcp_task, subagent_type: fix-planner) with this structured prompt:
target_ids: [explicit list of task IDs that need correction, from DevExperience data]
corrections:
- T042: switch from lib X to lib Y (repeated failures show X doesn't work)
- T067: add error handling for empty input (Reviewer critical x2)
cyclic_failure_id: [task ID if triggered by cyclic failure, else omit]
reorder_ids: [task IDs to move earlier if urgently needed, else omit]
deprecate_ids: [task IDs to remove if explicitly obsolete, else omit]
goal_context: [1 sentence on what the project goal is, from DB]
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.
- yesterday First seen · 80 lines · 39 tokens per session scan A 13b51b999f95
learning is an agent published in the GitHub repository budagov-lab/DreamTeam (1 stars, last pushed 5mo ago), licensed MIT. It adds 39 tokens to every session and 852 once invoked, about $0.0002 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 2026-08-31.
Other agents, from other repositories
cut
Over-engineering hunter. Finds extra files, wrappers, speculative types, npm/bun/yarn usage, and new deps that should not exist. Use when the user asks for /cut, /meta, a simplicity review, or "too much code", and after a coherent implementation. Do not use for comment-only review (Comment Sicko).
verifier
Skeptical, read-only validator. Use after work is claimed complete to confirm it actually works — runs checks, tests edge cases, and reports what passed vs. what is incomplete or broken.
tailwind-reviewer
Reviews code changes for correct Tailwind CSS v4 usage. Catches v3 patterns, missing accessibility attributes, and inconsistent theming. Use after generating or modifying Tailwind code.
codex-coder
Coding agent via Codex CLI. Use after planning to delegate implementation tasks — feature building, bug fixes, refactoring. Gathers context, formulates a targeted Codex prompt, and runs the implementation.
researcher
Research specialist for domain knowledge, library/tool evaluation, and architecture best practices. Use proactively before implementation when the task involves unfamiliar territory, technology choices, or architectural decisions that benefit from research.
verifier
Verification and QA specialist. Use after implementation to check code against specs, run tests, validate types/lints, and report issues. Reports problems — does not fix them.