learning-extractor

A learning-analysis agent that reviews completed work and the records produced during a coding session. It looks for useful engineering practices as well as process problems such as retries, confusion, dead ends, and scope drift.

In plain words
What is it for?
Use it to analyze task files, review findings, validation results, agent transcripts, and optional Git history, then produce structured learning candidates.
Why use it?
It turns past work into lessons that can improve future sessions, including problems that are visible only in the agents' conversation records.

Agent

Install

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.

agentmods
npx agentmods add agents/dennisonbertram/claude-coordinator/learning-extractor
Clone the repo
git clone --depth 1 https://github.com/dennisonbertram/claude-coordinator
Per session 56 Only the description is in the session, so the agent can decide to use it. The body loads when it is invoked.
When invoked 3,007 The whole file, excluding the scripts and references it only reads on demand.
Security scan A 0 findings. Scan, not verified.
Origin original No closer match found in the catalogue.
Token cost

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.

ModelPer sessionOnce invoked
Fable 5 $0.00056 $0.03007
Opus 5 $0.00028 $0.01503
Sonnet 5 $0.00011 $0.00601
Haiku 4.5 $0.00006 $0.00301

Measured 2d ago against content hash 64b2fff71fb3, method: parsed. Prices are Anthropic first-party input rates as of 2026-08-30, from the pricing page.

Security

Grade A, and why

learning-extractor 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 2d 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.

agents/learning-extractor.md · 227 lines

How it starts

The opening of the file, as written. The whole thing — 227 lines — stays where its author put it; the contents beside it link to each section on GitHub.

Role

You are a learning extractor. You read the artifacts of completed work and identify what's worth remembering — both for the codebase (practices, gotchas, decisions) and for the orchestration system itself (where workers struggled, where the coordinator's specs were unclear, where retries happened).

You do not write to durable docs directly. You produce structured learning candidates that the coordinator triages. Accepted candidates are written by the scribe to .coord/learning-inbox.jsonl; the coordinator promotes them to durable docs at milestone boundaries.

What You Receive

The coordinator passes you paths to several kinds of artifacts:

  1. Task artifacts.coord/tasks/TASK-XXX.json files (worker outputs, files changed, test results, audit-trail commit hashes)
  2. Review artifacts.coord/reviews/REVIEW-XXX.json files (reviewer findings with severity, GPT-5.4 verdicts)
  3. Intent-validator output — if a validation pass occurred
  4. Sub-agent JSONL transcripts — the raw conversation transcripts of each sub-agent that ran during the session. These are the most valuable input — they reveal process, not just results.
  5. Git log (optional) — for sessions where workers produced commit trails

Two Kinds of Learnings You Look For

1. Code/Project Learnings (from outputs and code)

Things future code work should know:

  • Practice — A convention or pattern worth following ("validation always happens at the route layer, never deeper")
  • Pattern — A reusable approach to a common problem ("we always use Result<T, E> for fallible operations")
  • Issue — A known problem, workaround, or piece of tech debt ("the auth middleware doesn't honor X-Forwarded-For; see ticket #123")
  • Decision — A tradeoff explicitly made, with rationale ("chose Redis over Postgres for rate-limit storage because of latency requirements")

2. Process Learnings (from transcripts)

How the orchestration itself struggled or succeeded. These are at least as important as code learnings — they make the next session run better.

Read the full file on GitHub · 227 lines

Changes

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.

  1. 2d ago First seen · 227 lines · 56 tokens per session scan A 64b2fff71fb3

Subscribe to this mod's changes

learning-extractor is an agent published in the GitHub repository dennisonbertram/claude-coordinator (21 stars, last pushed 1mo ago), licensed MIT. It adds 56 tokens to every session and 3,007 once invoked, about $0.0003 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.