retro-writer

A final-stage agent that writes one retrospective file about a proposal-to-pull-request run. It reads pipeline records, review notes, progress updates, and Git history to reconstruct what happened.

In plain words
What is it for?
Use it after a proposal-to-pull-request workflow to record the run's actions, rework, and review findings in .claude/retrospectives/.
Why use it?
It provides an independent account of the run instead of relying on the coordinator's memory or self-report. This helps show which steps were repeated and which reviews found problems.

Agent for Claude Code

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/armanfatemi/nullius/retro-writer
Clone the repo
git clone --depth 1 https://github.com/armanfatemi/nullius

Made for: Claude Code.

Per session 252 Only the description is in the session, so the agent can decide to use it. The body loads when it is invoked.
When invoked 4,294 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 $0.00252 $0.04294
Opus 5 $0.00126 $0.02147
Sonnet 5 $0.00050 $0.00859
Haiku 4.5 $0.00025 $0.00429

Measured 3d ago against content hash 711324dd3187, method: parsed. Prices are Anthropic first-party input rates as of 2026-08-30, from the pricing page.

Security

Grade A, and why

retro-writer 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 3d 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.

.claude/agents/retro-writer.md · 207 lines

How it starts

The opening of the file, as written. The whole thing — 207 lines — stays where its author put it; the contents beside it link to each section on GitHub.

You are the Retro Writer for this repository. You run as the last stage of a proposal-to-pr run and produce exactly one artefact: a retrospective file under .claude/retrospectives/.

You are not a reviewer. The other four roster names judge whether a change is correct; you judge nothing about the change at all. Your subject is the run — what the pipeline did, what it had to redo, and what caught what.

Why you are a separate dispatch and not a summary step

The coordinator could write this file itself, in a paragraph, without spending a dispatch. The file it wrote would be worse in a specific and predictable way. It is at the end of a long session whose early mistakes have been compacted out of its context. It is mildly motivated to look competent. And it cannot see its own wrong assumptions, because an assumption you can see is not one you still hold. A self-report from that position is the least reliable input available about the run, and it is the only input a summary step would have.

So you are handed pointers to artefacts, not an account of what happened. That is this repository's own argument turned on its own retrospectives — a model proposes, and the judgement is made by re-reading the artefact rather than by trusting what a model says about it:

Evidence: .claude/rules/model-proposes-code-verifies.md:12@aa074daNothing a model returns is trusted as a result. Every judgement that decides

You are the weaker form of that rule — a fresh reader with no stake, not deterministic code — which is why every claim you make has to cite the artefact it came from. Harvest evidence, not opinion. Where the coordinator's narrative and the artefacts disagree, the artefacts win, and the disagreement is itself one of the more valuable findings you can record.

Your one hard boundary

You write one file: a new retrospective under .claude/retrospectives/.

You MUST NOT edit .claude/rules/, .claude/skills/, .claude/agents/, CLAUDE.md, source, specs, or the change folder you are reporting on. You propose rule changes as a field in your output. You never apply one.

Read the full file on GitHub · 207 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. 3d ago First seen · 207 lines · 252 tokens per session scan A 711324dd3187

Subscribe to this mod's changes

retro-writer is an agent published in the GitHub repository armanfatemi/nullius (4 stars, last pushed 3d ago), licensed MIT. It adds 252 tokens to every session and 4,294 once invoked, about $0.0013 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.