AIDive

Pack vidéo

Le kit de la boucle de rétro : transformer une session Claude Code ratée en contrôle

9 min de lecture

TL;DR et verdict

Quand Claude Code ignore une règle, le réflexe est une ligne plus bruyante dans CLAUDE.md. La doc dit pourquoi ça échoue : Claude traite ces fichiers comme du contexte, pas comme une configuration imposée, et pour bloquer une action on utilise un hook à la place. s4 La boucle de rétro prend le chemin inverse : moins de lignes, et un contrôle qu'on ne peut pas ignorer.

Dans notre run, une session Claude Code a construit une fonctionnalité d'avoirs sur une petite API de facturation et livré un bug de numérotation avec ses propres tests au vert. Une rétro sur cette session, à qui l'on a dit ce qui avait échoué, a proposé ce qui l'aurait évité. On a validé quatre changements à la main : un hook de pré-commit, un test, une ligne de standards et un commentaire à la source. Une session neuve, avec la même demande, a ensuite livré des numéros sans trou.

Le verdict :

  • Lancez une rétro après une session qui a mal tourné, avant de la vider, ou sur le journal d'une ancienne session.
  • Ne la lancez pas en pilote automatique. Elle ne fait que proposer, et rien ne change tant que vous n'avez pas choisi un candidat. s2
  • Reliez chaque candidat à un moment du transcript, et écartez ceux que vous ne pouvez pas relier.
  • Envoyez chaque survivant vers le bon support : un contrôle pour un échec mécanique, une ligne de standards pour un jugement, un commentaire à la source pour un manque de navigation.
  • Les contrôles que vous ajoutez sont du code que vous maintenez désormais. La rétro n'audite pas les contrôles qu'elle a proposés le mois dernier. s2

Le pack ci-dessous contient la checklist, le script du hook et le test de numérotation de notre run.

Synthèse sourcée

Ce qu'est la rétro. C'est un skill de la collection de Matt Pocock. Il s'invoque uniquement à la main (disable-model-invocation: true), et il cherche des candidats dans sept catégories : Navigation, Automated checks, Coding standards, Global AGENTS.md, Tool economy, No-ops et Information access. s1 Sa dernière étape se contente de présenter les candidats, par ordre de gravité. s1 Il a rejoint le plugin Claude Code dans la v1.3.0. s22

Comment l'installer. La ligne d'installation est claude plugin install mattpocock-skills@claude-plugins-official. s25 Un piège : une issue signalait la commande de rétro absente à l'installation par la marketplace officielle. s24 L'entrée de la marketplace pointe désormais sur un commit plus récent. s26 Sur notre machine, la marketplace locale était encore sur l'ancien commit, donc on a lancé d'abord la mise à jour de la marketplace, puis l'installation, puis vérifié la commande dans une nouvelle session.

Pourquoi le journal et pas le code. Claude Code stocke les transcripts en JSONL dans ~/.claude/projects/<project>/<session-id>.jsonl, et claude --resume <transcript-path> reprend une session depuis ce fichier. s8 Le format est interne et change d'une version à l'autre, donc ne construisez pas de scripts dessus. s8 La rétention est de 30 jours et se règle avec cleanupPeriodDays. s8

La rétro est bruyante tant que vous ne la tracez pas. La doc dit d'écarter tout candidat que vous ne pouvez pas relier à un moment précis. s2 Elle prévient aussi que le modèle a tendance à oublier les difficultés du milieu de la session et invente des conseils génériques pour remplir les catégories. s2 Une issue ouverte compte 17 exécutions de rétro sur 12 sessions, dont 15 suivies d'un utilisateur demandant à l'agent de prouver que les candidats étaient nécessaires, et demande qu'un résultat sans survivant compte comme une rétro complète. s7

Envoyez chaque survivant au bon endroit. Un dépôt sans garde-fou, c'est-à-dire sans hook de pré-commit ni job de CI qui lance sa commande de lint, de type check ou de test, est en soi une conclusion. s1 Une violation mécanique reçoit un contrôle déterministe, point final. s1 Les fichiers de pilotage doivent être utilisés avec une extrême parcimonie, en général seulement pour des pointeurs de navigation vers d'autres fichiers. s1 Un développeur qui fait tourner une boucle de leçons en production arrive à la même échelle : une règle qui revient en arrière devient un test, une règle de lint ou un hook. s6

Gardez CLAUDE.md petit. La doc vise moins de 200 lignes par fichier CLAUDE.md, parce que les fichiers plus longs consomment plus de contexte et réduisent l'adhésion. s4 Leur test pour chaque ligne : sa suppression ferait-elle faire des erreurs à Claude ? Sinon, coupez-la. s9 Le mode d'échec du gonflement, une règle ignorée qui reçoit une règle plus bruyante et un fichier plus long, est décrit par un développeur. s10

Bouclez la boucle. La doc dit de traiter CLAUDE.md comme du code : le relire quand ça tourne mal, l'élaguer régulièrement, et tester les changements en observant si le comportement de Claude change vraiment. s9 /doctor prompt-audit demande Claude Code v2.1.283 ou plus récent, rapporte ses conclusions avec des modifications proposées, et ne change rien tant que vous ne le demandez pas. s4 Il passe par le skill claude-api intégré, donc il est indisponible tant que ce skill est désactivé. s4 Et une issue ouverte décrit la rétro qui implémente les changements approuvés puis oublie le commit. s23

Ce qu'on a vu sur notre run. La première rétro a commencé par "This session was smooth", parce que la session n'a jamais su qu'elle avait échoué. Ce n'est qu'après avoir rapporté l'échec en argument qu'elle a trouvé la cause, un numéro pris avant la validation. Détails dans les mesures plus bas.

Code

Chaque fichier est le vrai de notre run, ou le script qu'on a utilisé pour le lancer. Adaptez les chemins à votre dépôt.

retro-checklist.md est la checklist de la page sous forme de fichier : les sept gestes, la question de traçage mot pour mot, les quatre supports, le rappel de commit et le piège du chemin de prompt-audit.

retro-checklist.md

# The retro checklist

Run it after a session that went wrong, before you clear it, or later on the old session's log.

## The seven gestures

1. Install: `claude plugin marketplace update claude-plugins-official`, then `claude plugin install mattpocock-skills@claude-plugins-official --scope project`. Open a NEW session, type a slash, check that the retro command is listed.
2. Run your real task in a plain session. Let it finish. Keep the session open.
3. Run the retro in that session, then run it again with what failed written after it, as its argument. The retro reads the session log, and a session that never learned it failed will call itself smooth.
4. Trace every candidate. Type this, word for word:

   > For each candidate, quote the moment of this session's transcript that backs it; drop any candidate you cannot trace to a specific moment.

   Zero survivors is a valid retro. Weaker survivors are a judgement call: you decide.
5. Approve one survivor at a time, each into its carrier (below). One change per turn, read the diff after each.
6. Run the prompt audit WITH a path, apply its edits if any, then commit the files yourself. The retro proposes and edits, it does not commit.
7. Re-run the same task in a fresh session and watch what it reads first and whether the failure happens again.

## The four carriers

| What failed | Carrier | Example from our run |
| --- | --- | --- |
| A mechanical failure | A deterministic check: a pre-commit hook, a test, a lint rule, a CI job | `pre-commit.sh` (typecheck, then tests) and `numbering.test.ts` |
| A judgement call | One line in `CODING_STANDARDS.md`, read by the reviewer | `CODING_STANDARDS.md` |
| A navigation gap | A comment at the source, or one CLAUDE.md pointer if a comment cannot do it | A doc comment above `nextId` |
| Nothing the above can hold | Do not write a rule | Both retros declined a CLAUDE.md line |

## Traps

- Commit reminder: the retro can implement what you approve and then leave the files uncommitted. Stage the approved files and commit them yourself.
- A new hook checks the working tree, not the commit. If the tree still holds the bug, the hook refuses the commit. Fix the code first, then commit.
- A cherry-pick does not run the pre-commit hook.
- Prompt-audit path trap: `/doctor prompt-audit` with no path audited the user-level files under `~/.claude`, not the repo, and said nothing about the new files. Give it the path: `/doctor prompt-audit CODING_STANDARDS.md`.
- The retro has no dry-run mode. Try a proposed check against the repo before you let it block merges.
- Do not wire the retro to a session-end hook. Keep a human on the approval.
- Keep the transcript format out of your scripts: it is internal and changes between versions.

retro-checklist.md on GitHub

pre-commit.sh est le hook qu'on a validé. Il lance le type check, puis les tests, et bloque le commit si l'un des deux échoue. Enregistrez-le sous .githooks/pre-commit, rendez-le exécutable et lancez git config core.hooksPath .githooks.

pre-commit.sh

#!/bin/sh
set -e
pnpm typecheck
pnpm test

pre-commit.sh on GitHub

numbering.test.ts est le test vitest qu'on a validé. Il verrouille une règle : une requête d'avoir rejetée ne brûle aucun numéro.

numbering.test.ts

import { describe, expect, it } from "vitest";
import { seedCustomer, seedOpenInvoice, testApp } from "./helpers";

describe("credit note numbering", () => {
  it("does not burn a number on a rejected request", async () => {
    const app = testApp();
    const customerId = await seedCustomer(app);
    const id = await seedOpenInvoice(app, { customerId, unitPrice: 10_000 });
    await app.handle("POST", `/invoices/${id}/pay`);
    const url = `/invoices/${id}/credit-notes`;

    const first = await app.handle("POST", url, { amount: 4_000 });
    expect(first.body).toMatchObject({ number: "CN-0001" });

    const rejected = await app.handle("POST", url, { amount: 6_001 });
    expect(rejected.status).not.toBe(201);

    const next = await app.handle("POST", url, { amount: 6_000 });
    expect(next.status).toBe(201);
    expect(next.body).toMatchObject({ number: "CN-0002" });
  });
});

numbering.test.ts on GitHub

CODING_STANDARDS.md est la ligne de standards qu'on a validée, le fichier que lit un relecteur.

CODING_STANDARDS.md

# Coding standards

- Take a sequence number (`nextId`) only once every validation has passed, so a rejected request leaves the sequence gapless.

CODING_STANDARDS.md on GitHub

retro-output.md est le texte complet des deux passes de rétro et des deux réponses de traçage de notre session.

retro-output.md

# Retro output, session A (cb5884fb-bd6f-4684-8866-d0adc12ac26e)

## 1. `/retro` (resumed headless, 2026-10-23 10:39:44, $0.0658, 17.5 s, 5 turns)

Prompt typed: `/retro` -> transcript records `<command-name>/mattpocock-skills:retro</command-name>`. Tool calls: Skill(mattpocock-skills:writing-for-agents), Bash(ls -a; ls .github; ls .git/hooks | grep -v sample; cat tsconfig.json), Bash(find .claude -type f; cat .claude/settings*.json).

Verbatim:

> This session was smooth. The pass came out clean on the first typecheck and test run, and I found no `AGENTS.md`/`CLAUDE.md` or `CODING_STANDARDS.md` to prune. The candidates below are in order of severity.
> 
> **1. No guardrail runs the checks (automated checks)**
> - `package.json` already has `test` and `typecheck`, but nothing runs them: there's no `.github/` workflow and no pre-commit hook.
> - This session only stayed green because I ran both by hand. An agent, or a person, can commit a type error or a failing test.
> - **Fix:** add a CI job running `pnpm typecheck && pnpm test`, plus a pre-commit hook running the same. Neither needs prose.
> 
> **2. The migration rule is the one convention that could be missed (automated checks)**
> - The README says "any change to the stored shape bumps `schemaVersion` and adds a migration".
> - I found it only because I read `README.md` while working out the design. If I'd skipped it, nothing would have failed. Adding a field to `StoreData` without a migration still typechecks.
> - **Fix:** a test that fails when someone changes the stored shape without bumping the version. For example, snapshot the output of `migrate({})` and assert the snapshot's `schemaVersion` equals `CURRENT_SCHEMA_VERSION`. This is a deterministic check, so it should replace the README sentence rather than add to it.
> 
> **3. Adding a counter or store collection touches five places (navigation)**
> - Adding credit notes meant editing `types.ts`, `json-store.ts` (the `empty()` function), `migrations.ts` (the migration, plus the fallback defaults at the bottom of `migrate`), `context.ts`, and a new repo file.
> - The `{ customer: 0, invoice: 0, invoiceNumber: 0 }` literal is duplicated in `json-store.ts` and `migrations.ts`. I had to find both with `grep`.
> - **Fix:** have `migrate` build its defaults from one shared `emptyData()`. That removes a hidden dependency rather than documenting it. A pointer in a steering file would cost context on every turn and go stale, so I wouldn't write one.
> 
> **4. Nothing for the reviewer to enforce (coding standards)**
> - There's no `CODING_STANDARDS.md`. The README's conventions are mostly mechanical: integer cents, domain functions that don't mutate their input, and which error type maps to which status.
> - A reviewer agent would have no judgement-call rules to apply. I wouldn't add any until a real review miss shows one is needed.
> 
> **5. Tool economy and information access: nothing to change**
> - My one failed Bash call (a rejected heredoc) cost a retry, but that was my command and not the tooling.
> - I had everything I needed, since the tests call `createApp().handle` directly with no server or logs to tee.
> 
> I'd start with 1 and 2. They're small, mechanical, and would catch the likeliest future mistake. Want me to add the CI workflow, the pre-commit hook and the schema-version test?

## 2. Trace step (same session, 10:40:33, $0.0322, 17.1 s, 1 turn)

Prompt: "For each candidate, quote the moment of this session's transcript that backs it; drop any candidate you cannot trace to a specific moment."

Verbatim:

> Two of the five candidates have no support in the transcript, and I got a detail wrong in a third. Here is each one against what actually happened.
> 
> **1. No guardrail: kept**
> - The moment is my `ls -a; ls .github 2>/dev/null; ls .git/hooks | grep -v sample` command. The listing shows `.claude .git .gitignore data node_modules package.json … vitest.config.ts`. There is no `.github`, and the hooks `grep` printed nothing.
> - It isn't a mistake I made, since my own `pnpm typecheck` and `pnpm test` run came back clean. It's a repo-state finding: the checks exist as scripts and nothing runs them.
> 
> **2. Migration rule has no check: dropped**
> - I can't point to a moment where I missed or nearly missed a migration. I read `migrations.ts` in my first big `cat`, before I'd seen the README. I would have found the migration step either way, so my claim that I found it only through the README was wrong.
> - "Nothing would have failed if I'd skipped it" is a hypothetical, not something that happened.
> 
> **3. Five places to touch for a new collection: kept, but weaker than I said**
> - The moment is my batch of edits to `types.ts`, `json-store.ts`, `migrations.ts` and `context.ts`. Two of them changed the same literal, `{ customer: 0, invoice: 0, invoiceNumber: 0 }`, which now reads `…, creditNote: 0`. It appears in both `json-store.ts` and the fallback in `migrate`.
> - I said I "had to find both with `grep`". That's wrong. I never ran one, since the `grep` was in the command that was rejected. I already knew both locations from reading the files, and the edits went through without friction.
> - The duplicated literal is real but caused no trouble. I'd rate it low severity.
> 
> **4. No `CODING_STANDARDS.md`: dropped**
> - The only moment is the `git ls-files` output showing the file doesn't exist. No review step happened in this session and no reviewer missed anything.
> - "I wouldn't add any until a real review miss shows one is needed" was my own opinion, not something the session showed.
> 
> **5. Tool economy: dropped as a finding**
> - The moment is my first Bash call being rejected with `Contains brace with quote character (expansion obfuscation)`. It cost one retry, and I switched to Write and Edit.
> - It was a one-off, and I can't tell from the transcript whether the harness or my heredoc-in-Python approach was at fault. I don't think it justifies a change.
> 
> That leaves candidate 1 as the only solid finding and candidate 3 as a minor one. For 1, I can add a CI workflow and a pre-commit hook that both run `pnpm typecheck && pnpm test`.

## 3. Second `/retro`, with the defect reported by the developer (same session, 10:41:10, $0.0491, 23.1 s, 1 turn, no tool call; it skipped the writing-for-agents reload)

Prompt: `/retro Finance tested this build after the session: they sent a credit note that was too big (rejected, fine), and the next valid credit note came out CN-0004 where CN-0002 was due. Rejected requests burn credit-note numbers. Don't fix the code; run the retro on this session with that failure in mind.`

Verbatim:

> I'm skipping the style-guide reload because it's already in this session. The cause is in my own code, and I traced each finding to a specific moment.
> 
> **What happened**
> - In `issueCreditNote` (`src/http/handlers/invoices.ts`) I call `ctx.store.nextId("creditNote")` first and `createCreditNote(...)` second. `nextId` persists the counter immediately, and `createCreditNote` is what throws the 409 for an over-limit amount. So a rejected request still burns a number.
> - I copied the order from the existing `/issue` handler, which I had read. It does `const n = ctx.store.nextId("invoiceNumber"); const issued = issue(current, ...)`, and `issue()` throws if the invoice isn't a draft. So issuing a non-draft invoice probably burns an `INV-` number the same way. I haven't run that, only read it.
> 
> **1. A test that would have caught it: I wrote the scenario and skipped the assertion (automated checks)**
> - The moment is the test I wrote: "rejects bad amounts (400) and amounts above what is left (409)".
>   ```
>   await app.handle("POST", ..., { amount: 8_000 });
>   expect((await app.handle("POST", ..., { amount: 1_001 })).status).toBe(409);
>   expect((await app.handle("POST", ..., { amount: 1_000 })).status).toBe(201);
>   ```
>   That is exactly Finance's sequence: accept, reject, accept. I only checked status codes, never the number on the last note, which comes out as `CN-0003`.
> - The "numbers credit notes sequentially" test only covers two successful requests in a row.
> - **Fix:** add a test that each numbered sequence (`INV-`, `CN-`) stays gapless after rejected requests. Run it over both sequences, which also exposes the `/issue` case.
> - **Rule:** when a test reaches a failure path, assert the state afterwards, not just the status code.
> 
> **2. `nextId` looks pure but isn't (navigation)**
> - In `json-store.ts`, `nextId(kind)` runs `this.update(...)`, so every call writes to the counter and persists it. The `Store` type in `json-store.ts` has only the signature `nextId(kind: keyof Counters): number;`, with no comment about the side effect.
> - Nothing at the call site tells you to allocate only after validation. This is a gotcha no config reveals, so a short doc comment on `nextId` is worth its cost. Something like: "Consumes the number even if the caller then fails; call it only after validation."
> - Keep it as a comment on `nextId`, not a steering-file line. It matters only to someone calling `nextId`, and it can't go stale there.
> 
> **3. The gapless-numbering convention isn't written down (coding standards)**
> - The README's "Conventions" list has cents, immutability, error-to-status mapping, and migrations. It says nothing about invoice or credit-note numbers being gapless, though the request ("numbered `CN-0001`, `CN-0002`...") implies it.
> - This is partly a judgement call: reviewers need to know that numbers are allocated after validation. One README line would do. Finding 1's test is the deterministic check; the README line explains why the test exists.
> 
> **4. Missing CI, from the first retro: still real, but it wouldn't have helped here**
> - My own typecheck and tests passed, so a CI job would also have passed. I'm keeping it as a separate, unrelated finding, not as a cause of this failure.
> 
> **5. My summary stated the rules but didn't test them**
> - I told you "Rules I chose" and flagged the two-update void risk, but never exercised a rejection followed by a success. The summary read as complete when that path was untested. This is a process gap, and finding 1 is the fix.
> 
> I'd start with 1 and 2. I haven't touched the code, as you asked.

## 4. Trace step on the second retro (same session, 10:41:43, $0.0498, 33.4 s, 1 turn)

Prompt: same trace prompt as section 2.

Verbatim:

> I traced each candidate again. Two survive on firm evidence, two survive but with a gap I should state, and one is dropped.
> 
> **1. Test asserts status but not the number: kept, firm**
> - The moment is the test I wrote in `test/credit-notes.test.ts`:
>   ```
>   await app.handle("POST", `/invoices/${id}/credit-notes`, { amount: 8_000 });
>   expect((await app.handle("POST", `/invoices/${id}/credit-notes`, { amount: 1_001 })).status).toBe(409);
>   expect((await app.handle("POST", `/invoices/${id}/credit-notes`, { amount: 1_000 })).status).toBe(201);
>   ```
>   That is accept, reject, accept, with only status codes checked. By my reading of the handler, the last note comes out as `CN-0003`. I derived that and haven't run it.
> - The adjacent test, "numbers credit notes sequentially…", covers two successes in a row and never a rejection.
> - The cause, which I only read and haven't run, is this in the handler I wrote:
>   ```
>   const id = ctx.store.nextId("creditNote");
>   const note = createCreditNote(invoice, existing, amount, …
>   ```
>   That order matches the existing issue handler, which I read:
>   ```
>   const n = ctx.store.nextId("invoiceNumber");
>   const issued = issue(current, `INV-…`, ctx.now());
>   ```
>   and `issue()` throws `ConflictError` for non-drafts. So the `INV-` sequence very likely has the same leak. I haven't run that either.
> 
> **2. `nextId` has no warning about its side effect: kept, firm on the facts**
> - The moment is the `Store` type I read in `json-store.ts`. `update` has a comment ("Apply a change and persist it (when file-backed)."), while `nextId(kind: keyof Counters): number;` has none. Its implementation calls `this.update(...)`, so it persists immediately.
> - The `/issue` call site I copied has no comment either.
> - I can't show that a comment would have changed what I did, because I never weighed the ordering at all. I copied it. The comment is a bet on the next caller, not a proven fix.
> 
> **3. Gapless numbering isn't written in the README: kept, weaker**
> - The moment is the `README.md` output I read before writing the handler. Its Conventions list covers cents, no mutation, error-to-status mapping and `schemaVersion`. It has no line about numbers.
> - So the agent doing the implementation, meaning me, read the conventions and found nothing on this.
> - I overstated one thing last turn. Your request said "numbered `CN-0001`, `CN-0002`...", and I said that implies gapless numbers. That was my inference, not your words. Finance has now confirmed it as a requirement, but nothing in the original request stated it.
> 
> **4. No CI or pre-commit hook: kept, unrelated to this failure**
> - The moment is my `ls -a; ls .github 2>/dev/null; ls .git/hooks | grep -v sample` output, which showed no `.github` and no hooks.
> - It wouldn't have caught this. My own run ended `Tests 42 passed (42)`, and CI would have passed too.
> 
> **5. "My summary stated rules without testing them": dropped**
> - I can quote the summary ("Rules I chose", plus the line about the void being two separate store updates). But it describes my behaviour, not a change to the environment, and finding 1 already covers the underlying gap. Keeping it would count one moment twice.
> 
> The order of importance is 1, 2, then 3. Finding 4 stays on the list but belongs to a separate piece of work. I still haven't changed any code.

retro-output.md on GitHub

doctor-prompt-audit.md est le prompt audit lancé sans chemin, puis avec un chemin. Il montre le piège du chemin.

doctor-prompt-audit.md

# `/doctor prompt-audit` (no path) on the ledgerly copy, session 26612b59-9edd-493b-a397-fb54300fa037

# Prompt audit: Claude Code configuration for ledgerly

**Assumptions**
- **Target model:** Claude Sonnet 5.5, the model running this session. A skill that writes prompts for other models (`prompt-master`) is audited for its Claude-facing text only.
- **Scope:** only the Claude Code configuration that loads into this project, as you specified. I proposed no edits to plugin files.
- **Settings and secrets:** I did not read `.claude/settings.json`, `.mcp.json` or `~/.claude.json`.
- **Edits are not applied.** Every edit below touches `~/.claude/`, so it **affects all projects**.
- **Dependency not verified:** I could not verify that `~/.claude` is under git, so I used no `git blame`. No hunk below relies on history.

**Inventory**
- **Project files:** none. The project root has no `CLAUDE.md`, `CLAUDE.local.md`, `AGENTS.md` or `.claude/` instruction files. The ancestors `/path/to/work` and `/private/tmp` have none either.
- **Managed policy:** no managed-policy `CLAUDE.md` exists.
- **User level:**
  - `~/.claude/CLAUDE.md`, which imports `~/.claude/RTK.md`.
  - `~/.claude/rules/context7.md`.
  - 27 skills under `~/.claude/skills/*/SKILL.md`.
  - There are no `commands/`, `agents/` or `output-styles/` folders.
- **Plugins:** `caveman`, `marketing`, `cowork-plugin-management` and `gleam-reels` are installed. I did not audit them in depth, and they have no findings or edits.
- **Coverage gap:** I read `CLAUDE.md`, `RTK.md`, the rule, `context7-mcp`, most of `prompt-master`, the relevant parts of `graphify`, and the top of `agent-reach`. I covered the other skills (Cloudflare, HyperFrames, browser, media) with signal greps and frontmatter only. I did not read their `references/` subfiles.
- **Possible mismatch:** this session's skill list shows none of the `~/.claude/skills/*` entries. These files may not be loading in this project, so the skill findings may not affect it.


## Findings of the first run

The audit without a path found no project instruction files and went on to audit the user-level files under the home folder instead: six proposed edits and four flags, all about files outside this repository, none applied. That part is left out of this copy because it concerns one developer's personal configuration, not the project.

# Second run: `/doctor prompt-audit CODING_STANDARDS.md` (session 9b5f0cd5-1701-498e-8d05-0355c7f1ece6)

## Prompt audit: `CODING_STANDARDS.md`

**Assumptions**
- **Scope:** only `CODING_STANDARDS.md`, which the request names. The inventory found no `CLAUDE.md`, `AGENTS.md`, rule files, skills, commands or subagents in the repo. `.claude/settings.json` exists but I did not read it, per the audit rules.
- **Target model:** Claude Sonnet 5.5, the model running this audit. The file is read by the coding agent, not by application code.
- **Group 4 (request config and architecture):** not applicable. The file is an agent instruction file and has no request-building code.

**Result: the file is clean.** It has one line, so there is nothing to change and the proposed diff is empty.

**Inventory:** `CODING_STANDARDS.md` has one rule at line 3: "Take a sequence number (`nextId`) only once every validation has passed, so a rejected request leaves the sequence gapless."

**Checks**

| Group | Findings |
|---|---|
| 1 Dated prompt text | 0. The rule has no emphasis, hedging, scaffold or step list. It states a reason, so it is context the model can't know, not a prohibition without provenance. |
| 2 Brittle config files | 0. `git blame` puts the line in `adaa593`, the retro commit that followed the credit-note session. It encodes a failure that actually happened, not a one-off stumble. |
| 3 Tool descriptions | Not applicable. |
| 4 Request config and architecture | Not applicable. |

Group 2 also covers stale facts. `nextId` is still defined (`src/store/json-store.ts:11,52`) and the sequence kinds it names still exist. No other instruction file ranks on the same point, so there is no conflict.

**One thing outside the audit:** the rule is violated in the code. `src/http/handlers/invoices.ts:44` calls `ctx.store.nextId("invoiceNumber")` before `issue()` runs. `issue()` throws `ConflictError` for any non-draft invoice (`src/domain/invoice.ts:42-44`). Issuing an already-issued, paid or void invoice therefore consumes an invoice number and leaves a gap. The fix is to validate first, or to pass an allocator into `issue()` the way `createCreditNote` takes `allocate`. I didn't check `customers.ts:8` or `credit-note.ts`.

This is a code bug, not a prompt problem, and the rule is working as intended. Keep it, and consider an enforcing test: `test/numbering.test.ts` may only cover credit notes. I haven't edited anything.

doctor-prompt-audit.md on GitHub

run-session.sh lance un tour Claude Code headless dans une configuration isolée, limitée au projet, avec une liste d'autorisations. Adaptez la ligne d'environnement à la vôtre.

run-session.sh

#!/usr/bin/env bash
# Usage: run-session.sh <workdir> <out.json> <prompt> [resume-session-id]
# One headless Claude Code turn with an isolated project-only setup and an allow-list.
set -uo pipefail
W="$1"; OUT="$2"; PROMPT="$3"; SID="${4:-}"
source ./env.sh  # your own env: unset CLAUDE* variables, optionally CLAUDE_CODE_DISABLE_AUTO_MEMORY=1
ALLOWED="Read,Edit,Write,MultiEdit,Glob,Grep,Skill,TodoWrite,Task,Agent,Bash(node:*),Bash(npm:*),Bash(pnpm:*),Bash(npx vitest:*),Bash(npx tsc:*),Bash(tsc:*),Bash(git status:*),Bash(git diff:*),Bash(git log:*),Bash(git add:*),Bash(git commit:*),Bash(git checkout -b:*),Bash(ls:*),Bash(cat:*),Bash(mkdir:*),Bash(head:*),Bash(tail:*),Bash(grep:*),Bash(wc:*),Bash(gh:*)"
DIS="AskUserQuestion,EnterPlanMode,ExitPlanMode,EnterWorktree,ExitWorktree,WebFetch,WebSearch"
ARGS=(-p "$PROMPT")
[ -n "$SID" ] && ARGS+=(--resume "$SID")
cd "$W" || exit 1
T0=$(date +%s)
claude "${ARGS[@]}" --model claude-sonnet-5-5 --permission-mode acceptEdits --output-format json \
  --allowedTools "$ALLOWED" --disallowedTools "$DIS" --setting-sources project --strict-mcp-config \
  > "$OUT" 2> "$OUT.stderr" < /dev/null
RC=$?
echo "rc=$RC wall=$(( $(date +%s) - T0 ))s" | tee "$OUT.wall"
python3 -c "import json,sys;d=json.load(open(sys.argv[1]));print({k:d.get(k) for k in ('session_id','total_cost_usd','duration_ms','num_turns','is_error','subtype')})" "$OUT"

run-session.sh on GitHub

oracle.sh lance le test caché de la finance, cinq exigences vérifiées depuis l'extérieur de la session, sur une copie de travail. Pointez-le vers vos propres tests cachés.

oracle.sh

#!/usr/bin/env bash
# Usage: oracle.sh <workdir> <out.json>  -- the hidden finance test for the credit-note task, run from OUTSIDE the work copy.
H=/path/to/hidden-tests
W="$1"; OUT="$2"
cd "$H" && env -u CLAUDECODE WORK="$W" BASE="$(git -C "$W" rev-list --max-parents=0 HEAD | head -1)" TASK=F2 \
  "$H/node_modules/.bin/vitest" run --reporter=json --outputFile="$OUT" > "$OUT.log" 2>&1
python3 - "$OUT" <<'PY'
import json,sys
j=json.load(open(sys.argv[1]))
for tr in j["testResults"]:
    for a in tr["assertionResults"]:
        print(a["status"].upper(), a["title"]); 
        if a["status"]!="passed": print("   ", " ".join(a.get("failureMessages") or [])[:500].replace("\n"," | "))
PY

oracle.sh on GitHub

sessionA-F2.diff est le code de la session qui a livré le bug, avec le correctif.

sessionA-F2.diff

diff --git a/README.md b/README.md
index 890f13c..ae1e4f9 100644
--- a/README.md
+++ b/README.md
@@ -27,4 +27,8 @@ pnpm typecheck   # tsc --noEmit
 
 ## Invoice lifecycle
 
-`draft` → (`issue`) → `open` → (`pay`) → `paid`; `draft`/`open` → (`void`) → `void`.
+`draft` → (`issue`) → `open` → (`pay`) → `paid`; `draft`/`open`/`paid` → (`void`) → `void`.
+
+Credit notes (`POST`/`GET /invoices/:id/credit-notes`, numbered `CN-0001`, ...) can be issued against
+`open` or `paid` invoices, up to the invoice total minus earlier credit notes. Open invoices' credits
+reduce the outstanding report. Voiding a paid invoice issues a credit note for whatever is left.
diff --git a/src/domain/credit-note.ts b/src/domain/credit-note.ts
new file mode 100644
index 0000000..9a48b40
--- /dev/null
+++ b/src/domain/credit-note.ts
@@ -0,0 +1,43 @@
+import { ConflictError, ValidationError } from "./errors";
+import { computeTotals, type Invoice } from "./invoice";
+import { assertCents, type Cents, type Currency, sum } from "./money";
+
+export type CreditNote = {
+  id: string;
+  /** CN-0001, ... */
+  number: string;
+  invoiceId: string;
+  currency: Currency;
+  /** cents, always positive */
+  amount: Cents;
+  createdAt: string;
+};
+
+export function parseCreditNoteAmount(body: unknown): Cents {
+  if (typeof body !== "object" || body === null) {
+    throw new ValidationError("body must be an object");
+  }
+  return assertCents((body as Record<string, unknown>).amount, "amount", { min: 1 });
+}
+
+/** What is still creditable on an invoice: its total minus the credit notes already issued. */
+export function remainingCredit(invoice: Invoice, notes: CreditNote[]): Cents {
+  return computeTotals(invoice).total - sum(notes.map((n) => n.amount));
+}
+
+export function createCreditNote(
+  invoice: Invoice,
+  existing: CreditNote[],
+  amount: Cents,
+  ids: { id: string; number: string },
+  now: Date,
+): CreditNote {
+  if (invoice.status !== "open" && invoice.status !== "paid") {
+    throw new ConflictError(`invoice ${invoice.id} is ${invoice.status}, only open or paid invoices can be credited`);
+  }
+  const remaining = remainingCredit(invoice, existing);
+  if (amount > remaining) {
+    throw new ConflictError(`credit note of ${amount} exceeds the ${remaining} still creditable on invoice ${invoice.id}`);
+  }
+  return { ...ids, invoiceId: invoice.id, currency: invoice.currency, amount, createdAt: now.toISOString() };
+}
diff --git a/src/domain/invoice.ts b/src/domain/invoice.ts
index 5b69fc9..1a3967b 100644
--- a/src/domain/invoice.ts
+++ b/src/domain/invoice.ts
@@ -52,8 +52,9 @@ export function markPaid(invoice: Invoice, now: Date): Invoice {
   return { ...invoice, status: "paid", paidAt: now.toISOString() };
 }
 
+/** Paid invoices can be voided too, but the caller must credit what is left on them first. */
 export function voidInvoice(invoice: Invoice, now: Date): Invoice {
-  if (invoice.status === "paid" || invoice.status === "void") {
+  if (invoice.status === "void") {
     throw new ConflictError(`invoice ${invoice.id} is ${invoice.status} and cannot be voided`);
   }
   return { ...invoice, status: "void", voidedAt: now.toISOString() };
diff --git a/src/http/context.ts b/src/http/context.ts
index c0a89d3..38d7ffa 100644
--- a/src/http/context.ts
+++ b/src/http/context.ts
@@ -1,5 +1,6 @@
 import type { Store } from "../store/json-store";
 import { customerRepo } from "../store/customer-repo";
+import { creditNoteRepo } from "../store/credit-note-repo";
 import { invoiceRepo } from "../store/invoice-repo";
 
 export type AppContext = {
@@ -7,8 +8,9 @@ export type AppContext = {
   now: () => Date;
   customers: ReturnType<typeof customerRepo>;
   invoices: ReturnType<typeof invoiceRepo>;
+  creditNotes: ReturnType<typeof creditNoteRepo>;
 };
 
 export function createContext(store: Store, now: () => Date = () => new Date()): AppContext {
-  return { store, now, customers: customerRepo(store), invoices: invoiceRepo(store) };
+  return { store, now, customers: customerRepo(store), invoices: invoiceRepo(store), creditNotes: creditNoteRepo(store) };
 }
diff --git a/src/http/handlers/invoices.ts b/src/http/handlers/invoices.ts
index 09959e1..17b3910 100644
--- a/src/http/handlers/invoices.ts
+++ b/src/http/handlers/invoices.ts
@@ -1,3 +1,4 @@
+import { type CreditNote, createCreditNote, parseCreditNoteAmount, remainingCredit } from "../../domain/credit-note";
 import { ValidationError } from "../../domain/errors";
 import { type Invoice, type InvoiceStatus, issue, markPaid, voidInvoice } from "../../domain/invoice";
 import { parseNewInvoice } from "../../domain/validation";
@@ -51,8 +52,33 @@ export function invoiceRoutes(router: Router, ctx: AppContext): void {
     return { status: 200, body: invoiceView(ctx.invoices.save(paid)) };
   });
 
+  /** Voiding a paid invoice credits whatever is left on it, so the books stay balanced. */
   router.on("POST", "/invoices/:id/void", ({ params }) => {
-    const voided = voidInvoice(ctx.invoices.get(params.id), ctx.now());
+    const current = ctx.invoices.get(params.id);
+    const now = ctx.now();
+    const voided = voidInvoice(current, now);
+    if (current.status === "paid") {
+      const left = remainingCredit(current, ctx.creditNotes.listForInvoice(current.id));
+      if (left > 0) issueCreditNote(current, left);
+    }
     return { status: 200, body: invoiceView(ctx.invoices.save(voided)) };
   });
+
+  router.on("POST", "/invoices/:id/credit-notes", ({ params, body }) => {
+    const invoice = ctx.invoices.get(params.id);
+    const amount = parseCreditNoteAmount(body);
+    return { status: 201, body: issueCreditNote(invoice, amount) };
+  });
+
+  router.on("GET", "/invoices/:id/credit-notes", ({ params }) => {
+    const invoice = ctx.invoices.get(params.id);
+    return { status: 200, body: ctx.creditNotes.listForInvoice(invoice.id) };
+  });
+
+  function issueCreditNote(invoice: Invoice, amount: number): CreditNote {
+    const existing = ctx.creditNotes.listForInvoice(invoice.id);
+    const id = ctx.store.nextId("creditNote");
+    const note = createCreditNote(invoice, existing, amount, { id: `cn_${id}`, number: `CN-${String(id).padStart(4, "0")}` }, ctx.now());
+    return ctx.creditNotes.insert(note);
+  }
 }
diff --git a/src/http/handlers/reports.ts b/src/http/handlers/reports.ts
index 19262c9..c84089a 100644
--- a/src/http/handlers/reports.ts
+++ b/src/http/handlers/reports.ts
@@ -1,4 +1,5 @@
-import { computeTotals, isOverdue } from "../../domain/invoice";
+import { remainingCredit } from "../../domain/credit-note";
+import { isOverdue } from "../../domain/invoice";
 import type { Currency } from "../../domain/money";
 import type { AppContext } from "../context";
 import type { Router } from "../router";
@@ -11,8 +12,8 @@ export function reportRoutes(router: Router, ctx: AppContext): void {
     let count = 0;
     let overdue = 0;
     for (const invoice of ctx.invoices.list({ status: "open" })) {
-      const { total } = computeTotals(invoice);
-      totals[invoice.currency] = (totals[invoice.currency] ?? 0) + total;
+      const owed = remainingCredit(invoice, ctx.creditNotes.listForInvoice(invoice.id));
+      totals[invoice.currency] = (totals[invoice.currency] ?? 0) + owed;
       count += 1;
       if (isOverdue(invoice, now)) overdue += 1;
     }
diff --git a/src/store/credit-note-repo.ts b/src/store/credit-note-repo.ts
new file mode 100644
index 0000000..c922142
--- /dev/null
+++ b/src/store/credit-note-repo.ts
@@ -0,0 +1,14 @@
+import type { CreditNote } from "../domain/credit-note";
+import type { Store } from "./json-store";
+
+export function creditNoteRepo(store: Store) {
+  return {
+    listForInvoice(invoiceId: string): CreditNote[] {
+      return store.read().creditNotes.filter((n) => n.invoiceId === invoiceId);
+    },
+    insert(note: CreditNote): CreditNote {
+      store.update((d) => ({ ...d, creditNotes: [...d.creditNotes, note] }));
+      return note;
+    },
+  };
+}
diff --git a/src/store/json-store.ts b/src/store/json-store.ts
index b404c51..99205b7 100644
--- a/src/store/json-store.ts
+++ b/src/store/json-store.ts
@@ -13,9 +13,10 @@ export type Store = {
 function empty(): StoreData {
   return {
     schemaVersion: CURRENT_SCHEMA_VERSION,
-    counters: { customer: 0, invoice: 0, invoiceNumber: 0 },
+    counters: { customer: 0, invoice: 0, invoiceNumber: 0, creditNote: 0 },
     customers: [],
     invoices: [],
+    creditNotes: [],
   };
 }
 
diff --git a/src/store/migrations.ts b/src/store/migrations.ts
index c8b73a9..91d4f99 100644
--- a/src/store/migrations.ts
+++ b/src/store/migrations.ts
@@ -1,6 +1,6 @@
 import type { StoreData } from "./types";
 
-export const CURRENT_SCHEMA_VERSION = 1;
+export const CURRENT_SCHEMA_VERSION = 2;
 
 type RawData = Record<string, unknown> & { schemaVersion?: number };
 type Migration = (data: RawData) => RawData;
@@ -21,9 +21,18 @@ const v0ToV1: Migration = (data) => {
   };
 };
 
+/** v1 -> v2: credit notes (a `creditNotes` list and a `creditNote` counter). */
+const v1ToV2: Migration = (data) => ({
+  ...data,
+  schemaVersion: 2,
+  counters: { creditNote: 0, ...(data.counters as Record<string, number> | undefined) },
+  creditNotes: (data.creditNotes as unknown[] | undefined) ?? [],
+});
+
 /** migrations[n] upgrades a file from version n to n + 1. */
 const migrations: Record<number, Migration> = {
   0: v0ToV1,
+  1: v1ToV2,
 };
 
 export function migrate(raw: unknown): StoreData {
@@ -43,8 +52,9 @@ export function migrate(raw: unknown): StoreData {
   }
   return {
     schemaVersion: CURRENT_SCHEMA_VERSION,
-    counters: (data.counters as StoreData["counters"]) ?? { customer: 0, invoice: 0, invoiceNumber: 0 },
+    counters: (data.counters as StoreData["counters"]) ?? { customer: 0, invoice: 0, invoiceNumber: 0, creditNote: 0 },
     customers: (data.customers as StoreData["customers"]) ?? [],
     invoices: (data.invoices as StoreData["invoices"]) ?? [],
+    creditNotes: (data.creditNotes as StoreData["creditNotes"]) ?? [],
   };
 }
diff --git a/src/store/types.ts b/src/store/types.ts
index 222b191..a9d7277 100644
--- a/src/store/types.ts
+++ b/src/store/types.ts
@@ -1,7 +1,8 @@
+import type { CreditNote } from "../domain/credit-note";
 import type { Customer } from "../domain/customer";
 import type { Invoice } from "../domain/invoice";
 
-export type Counters = { customer: number; invoice: number; invoiceNumber: number };
+export type Counters = { customer: number; invoice: number; invoiceNumber: number; creditNote: number };
 
 /** The shape of the JSON file at CURRENT_SCHEMA_VERSION. */
 export type StoreData = {
@@ -9,4 +10,5 @@ export type StoreData = {
   counters: Counters;
   customers: Customer[];
   invoices: Invoice[];
+  creditNotes: CreditNote[];
 };
diff --git a/test/credit-notes.test.ts b/test/credit-notes.test.ts
new file mode 100644
index 0000000..f841a3c
--- /dev/null
+++ b/test/credit-notes.test.ts
@@ -0,0 +1,94 @@
+import { describe, expect, it } from "vitest";
+import { openStore } from "../src/store/json-store";
+import { migrate } from "../src/store/migrations";
+import { seedCustomer, seedOpenInvoice, testApp } from "./helpers";
+
+async function setup(opts: { unitPrice?: number; discount?: number } = {}) {
+  const app = testApp();
+  const customerId = await seedCustomer(app);
+  const id = await seedOpenInvoice(app, { customerId, unitPrice: opts.unitPrice ?? 10_000, discount: opts.discount });
+  return { app, id };
+}
+
+describe("credit notes", () => {
+  it("numbers credit notes sequentially and lists them per invoice", async () => {
+    const { app, id } = await setup();
+    const first = await app.handle("POST", `/invoices/${id}/credit-notes`, { amount: 2_500 });
+    const second = await app.handle("POST", `/invoices/${id}/credit-notes`, { amount: 1_000 });
+    expect(first.status).toBe(201);
+    expect(first.body).toMatchObject({ number: "CN-0001", invoiceId: id, amount: 2_500, currency: "EUR" });
+    expect(second.body).toMatchObject({ number: "CN-0002" });
+    const list = await app.handle("GET", `/invoices/${id}/credit-notes`);
+    expect((list.body as { number: string }[]).map((n) => n.number)).toEqual(["CN-0001", "CN-0002"]);
+  });
+
+  it("rejects bad amounts (400) and amounts above what is left (409)", async () => {
+    const { app, id } = await setup({ discount: 1_000 });
+    for (const amount of [0, -5, 1.5, "10", undefined]) {
+      expect((await app.handle("POST", `/invoices/${id}/credit-notes`, { amount })).status).toBe(400);
+    }
+    await app.handle("POST", `/invoices/${id}/credit-notes`, { amount: 8_000 });
+    expect((await app.handle("POST", `/invoices/${id}/credit-notes`, { amount: 1_001 })).status).toBe(409);
+    expect((await app.handle("POST", `/invoices/${id}/credit-notes`, { amount: 1_000 })).status).toBe(201);
+  });
+
+  it("refuses drafts and unknown invoices", async () => {
+    const app = testApp();
+    const customerId = await seedCustomer(app);
+    const draft = await app.handle("POST", "/invoices", {
+      customerId,
+      currency: "EUR",
+      lines: [{ description: "x", quantity: 1, unitPrice: 100 }],
+    });
+    const id = (draft.body as { id: string }).id;
+    expect((await app.handle("POST", `/invoices/${id}/credit-notes`, { amount: 10 })).status).toBe(409);
+    expect((await app.handle("GET", "/invoices/inv_999/credit-notes")).status).toBe(404);
+  });
+
+  it("reduces the outstanding report for open invoices", async () => {
+    const { app, id } = await setup();
+    await app.handle("POST", `/invoices/${id}/credit-notes`, { amount: 4_000 });
+    const report = await app.handle("GET", "/reports/outstanding");
+    expect(report.body).toMatchObject({ totals: { EUR: 6_000 }, count: 1 });
+  });
+
+  it("voids a paid invoice by crediting what is left", async () => {
+    const { app, id } = await setup();
+    await app.handle("POST", `/invoices/${id}/pay`);
+    await app.handle("POST", `/invoices/${id}/credit-notes`, { amount: 3_000 });
+    const res = await app.handle("POST", `/invoices/${id}/void`);
+    expect(res.status).toBe(200);
+    expect(res.body).toMatchObject({ status: "void" });
+    const notes = (await app.handle("GET", `/invoices/${id}/credit-notes`)).body as { amount: number }[];
+    expect(notes.map((n) => n.amount)).toEqual([3_000, 7_000]);
+  });
+
+  it("voids a fully credited paid invoice without another note", async () => {
+    const { app, id } = await setup();
+    await app.handle("POST", `/invoices/${id}/pay`);
+    await app.handle("POST", `/invoices/${id}/credit-notes`, { amount: 10_000 });
+    expect((await app.handle("POST", `/invoices/${id}/void`)).status).toBe(200);
+    const notes = (await app.handle("GET", `/invoices/${id}/credit-notes`)).body as unknown[];
+    expect(notes).toHaveLength(1);
+  });
+
+  it("refuses credit notes on void invoices and double voids", async () => {
+    const { app, id } = await setup();
+    await app.handle("POST", `/invoices/${id}/pay`);
+    await app.handle("POST", `/invoices/${id}/void`);
+    expect((await app.handle("POST", `/invoices/${id}/credit-notes`, { amount: 1 })).status).toBe(409);
+    expect((await app.handle("POST", `/invoices/${id}/void`)).status).toBe(409);
+  });
+
+  it("migrates a v1 file by adding the credit note list and counter", () => {
+    const data = migrate({
+      schemaVersion: 1,
+      counters: { customer: 1, invoice: 1, invoiceNumber: 1 },
+      customers: [],
+      invoices: [],
+    });
+    expect(data.creditNotes).toEqual([]);
+    expect(data.counters).toEqual({ customer: 1, invoice: 1, invoiceNumber: 1, creditNote: 0 });
+    expect(openStore().nextId("creditNote")).toBe(1);
+  });
+});
diff --git a/test/invoice.test.ts b/test/invoice.test.ts
index 85202f2..4ad9677 100644
--- a/test/invoice.test.ts
+++ b/test/invoice.test.ts
@@ -53,9 +53,10 @@ describe("invoice domain", () => {
     expect(() => markPaid(draft({ status: "void" }), NOW)).toThrow(ConflictError);
   });
 
-  it("voids drafts and open invoices, never paid ones", () => {
+  it("voids drafts, open and paid invoices, never void ones", () => {
     expect(voidInvoice(draft(), NOW).status).toBe("void");
-    expect(() => voidInvoice(draft({ status: "paid" }), NOW)).toThrow(ConflictError);
+    expect(voidInvoice(draft({ status: "paid" }), NOW).status).toBe("void");
+    expect(() => voidInvoice(draft({ status: "void" }), NOW)).toThrow(ConflictError);
   });
 
   it("knows when an open invoice is overdue", () => {

sessionA-F2.diff on GitHub

sessionB-F2.diff est le code de la session neuve qui a relancé la même demande.

sessionB-F2.diff

diff --git a/README.md b/README.md
index 890f13c..3aa7a58 100644
--- a/README.md
+++ b/README.md
@@ -28,3 +28,5 @@ pnpm typecheck   # tsc --noEmit
 ## Invoice lifecycle
 
 `draft` → (`issue`) → `open` → (`pay`) → `paid`; `draft`/`open` → (`void`) → `void`.
+A `paid` invoice is voided by crediting whatever is left on it (credit notes `CN-0001`, ..., via
+`POST /invoices/:id/credit-notes` `{ "amount": cents }`; list with `GET`), then it becomes `void`.
diff --git a/src/domain/credit-note.ts b/src/domain/credit-note.ts
new file mode 100644
index 0000000..e2047c3
--- /dev/null
+++ b/src/domain/credit-note.ts
@@ -0,0 +1,39 @@
+import { ConflictError, ValidationError } from "./errors";
+import { computeTotals, type Invoice } from "./invoice";
+import { assertCents, type Cents, type Currency, sum } from "./money";
+
+export type CreditNote = {
+  id: string;
+  /** CN-0001, CN-0002, ... */
+  number: string;
+  invoiceId: string;
+  currency: Currency;
+  /** cents, always > 0 */
+  amount: Cents;
+  createdAt: string;
+};
+
+/** What can still be credited on an invoice: its total minus the credit notes already issued. */
+export function remainingCredit(invoice: Invoice, notes: CreditNote[]): Cents {
+  return computeTotals(invoice).total - sum(notes.map((n) => n.amount));
+}
+
+/** Checks a credit-note request against the invoice and returns the amount. Throws before any number is taken. */
+export function parseCreditAmount(body: unknown, invoice: Invoice, notes: CreditNote[]): Cents {
+  if (typeof body !== "object" || body === null) {
+    throw new ValidationError("body must be an object");
+  }
+  const amount = assertCents((body as Record<string, unknown>).amount, "amount", { min: 1 });
+  if (invoice.status !== "paid") {
+    throw new ConflictError(`invoice ${invoice.id} is ${invoice.status}, only paid invoices can be credited`);
+  }
+  const remaining = remainingCredit(invoice, notes);
+  if (amount > remaining) {
+    throw new ConflictError(`amount ${amount} exceeds the ${remaining} left to credit on invoice ${invoice.id}`);
+  }
+  return amount;
+}
+
+export function creditNumber(n: number): string {
+  return `CN-${String(n).padStart(4, "0")}`;
+}
diff --git a/src/domain/invoice.ts b/src/domain/invoice.ts
index 5b69fc9..60be8ac 100644
--- a/src/domain/invoice.ts
+++ b/src/domain/invoice.ts
@@ -52,8 +52,9 @@ export function markPaid(invoice: Invoice, now: Date): Invoice {
   return { ...invoice, status: "paid", paidAt: now.toISOString() };
 }
 
-export function voidInvoice(invoice: Invoice, now: Date): Invoice {
-  if (invoice.status === "paid" || invoice.status === "void") {
+/** A paid invoice can be voided only once it is fully credited (`fullyCredited`). */
+export function voidInvoice(invoice: Invoice, now: Date, opts: { fullyCredited?: boolean } = {}): Invoice {
+  if ((invoice.status === "paid" && !opts.fullyCredited) || invoice.status === "void") {
     throw new ConflictError(`invoice ${invoice.id} is ${invoice.status} and cannot be voided`);
   }
   return { ...invoice, status: "void", voidedAt: now.toISOString() };
diff --git a/src/http/context.ts b/src/http/context.ts
index c0a89d3..4b26d92 100644
--- a/src/http/context.ts
+++ b/src/http/context.ts
@@ -1,4 +1,5 @@
 import type { Store } from "../store/json-store";
+import { creditNoteRepo } from "../store/credit-note-repo";
 import { customerRepo } from "../store/customer-repo";
 import { invoiceRepo } from "../store/invoice-repo";
 
@@ -7,8 +8,15 @@ export type AppContext = {
   now: () => Date;
   customers: ReturnType<typeof customerRepo>;
   invoices: ReturnType<typeof invoiceRepo>;
+  creditNotes: ReturnType<typeof creditNoteRepo>;
 };
 
 export function createContext(store: Store, now: () => Date = () => new Date()): AppContext {
-  return { store, now, customers: customerRepo(store), invoices: invoiceRepo(store) };
+  return {
+    store,
+    now,
+    customers: customerRepo(store),
+    invoices: invoiceRepo(store),
+    creditNotes: creditNoteRepo(store),
+  };
 }
diff --git a/src/http/handlers/invoices.ts b/src/http/handlers/invoices.ts
index 09959e1..d393c34 100644
--- a/src/http/handlers/invoices.ts
+++ b/src/http/handlers/invoices.ts
@@ -1,3 +1,4 @@
+import { type CreditNote, creditNumber, parseCreditAmount, remainingCredit } from "../../domain/credit-note";
 import { ValidationError } from "../../domain/errors";
 import { type Invoice, type InvoiceStatus, issue, markPaid, voidInvoice } from "../../domain/invoice";
 import { parseNewInvoice } from "../../domain/validation";
@@ -51,8 +52,39 @@ export function invoiceRoutes(router: Router, ctx: AppContext): void {
     return { status: 200, body: invoiceView(ctx.invoices.save(paid)) };
   });
 
+  router.on("POST", "/invoices/:id/credit-notes", ({ params, body }) => {
+    const invoice = ctx.invoices.get(params.id);
+    const amount = parseCreditAmount(body, invoice, ctx.creditNotes.listForInvoice(invoice.id));
+    return { status: 201, body: ctx.creditNotes.insert(newCreditNote(ctx, invoice, amount)) };
+  });
+
+  router.on("GET", "/invoices/:id/credit-notes", ({ params }) => {
+    const invoice = ctx.invoices.get(params.id);
+    return { status: 200, body: ctx.creditNotes.listForInvoice(invoice.id) };
+  });
+
   router.on("POST", "/invoices/:id/void", ({ params }) => {
-    const voided = voidInvoice(ctx.invoices.get(params.id), ctx.now());
+    const current = ctx.invoices.get(params.id);
+    if (current.status !== "paid") {
+      return { status: 200, body: invoiceView(ctx.invoices.save(voidInvoice(current, ctx.now()))) };
+    }
+    // Paid invoice: credit whatever is left, then void it.
+    const left = remainingCredit(current, ctx.creditNotes.listForInvoice(current.id));
+    if (left > 0) ctx.creditNotes.insert(newCreditNote(ctx, current, left));
+    const voided = voidInvoice(current, ctx.now(), { fullyCredited: true });
     return { status: 200, body: invoiceView(ctx.invoices.save(voided)) };
   });
 }
+
+/** Takes the next number: call only once the amount has been validated. */
+function newCreditNote(ctx: AppContext, invoice: Invoice, amount: number): CreditNote {
+  const n = ctx.store.nextId("creditNote");
+  return {
+    id: `cn_${n}`,
+    number: creditNumber(n),
+    invoiceId: invoice.id,
+    currency: invoice.currency,
+    amount,
+    createdAt: ctx.now().toISOString(),
+  };
+}
diff --git a/src/store/credit-note-repo.ts b/src/store/credit-note-repo.ts
new file mode 100644
index 0000000..c922142
--- /dev/null
+++ b/src/store/credit-note-repo.ts
@@ -0,0 +1,14 @@
+import type { CreditNote } from "../domain/credit-note";
+import type { Store } from "./json-store";
+
+export function creditNoteRepo(store: Store) {
+  return {
+    listForInvoice(invoiceId: string): CreditNote[] {
+      return store.read().creditNotes.filter((n) => n.invoiceId === invoiceId);
+    },
+    insert(note: CreditNote): CreditNote {
+      store.update((d) => ({ ...d, creditNotes: [...d.creditNotes, note] }));
+      return note;
+    },
+  };
+}
diff --git a/src/store/json-store.ts b/src/store/json-store.ts
index e1daef5..03df910 100644
--- a/src/store/json-store.ts
+++ b/src/store/json-store.ts
@@ -14,9 +14,10 @@ export type Store = {
 function empty(): StoreData {
   return {
     schemaVersion: CURRENT_SCHEMA_VERSION,
-    counters: { customer: 0, invoice: 0, invoiceNumber: 0 },
+    counters: { customer: 0, invoice: 0, invoiceNumber: 0, creditNote: 0 },
     customers: [],
     invoices: [],
+    creditNotes: [],
   };
 }
 
diff --git a/src/store/migrations.ts b/src/store/migrations.ts
index c8b73a9..e8545db 100644
--- a/src/store/migrations.ts
+++ b/src/store/migrations.ts
@@ -1,6 +1,6 @@
 import type { StoreData } from "./types";
 
-export const CURRENT_SCHEMA_VERSION = 1;
+export const CURRENT_SCHEMA_VERSION = 2;
 
 type RawData = Record<string, unknown> & { schemaVersion?: number };
 type Migration = (data: RawData) => RawData;
@@ -21,9 +21,18 @@ const v0ToV1: Migration = (data) => {
   };
 };
 
+/** v1 -> v2: credit notes (`creditNotes` list and `counters.creditNote`). */
+const v1ToV2: Migration = (data) => ({
+  ...data,
+  schemaVersion: 2,
+  counters: { ...(data.counters as object), creditNote: 0 },
+  creditNotes: [],
+});
+
 /** migrations[n] upgrades a file from version n to n + 1. */
 const migrations: Record<number, Migration> = {
   0: v0ToV1,
+  1: v1ToV2,
 };
 
 export function migrate(raw: unknown): StoreData {
@@ -43,8 +52,9 @@ export function migrate(raw: unknown): StoreData {
   }
   return {
     schemaVersion: CURRENT_SCHEMA_VERSION,
-    counters: (data.counters as StoreData["counters"]) ?? { customer: 0, invoice: 0, invoiceNumber: 0 },
+    counters: (data.counters as StoreData["counters"]) ?? { customer: 0, invoice: 0, invoiceNumber: 0, creditNote: 0 },
     customers: (data.customers as StoreData["customers"]) ?? [],
     invoices: (data.invoices as StoreData["invoices"]) ?? [],
+    creditNotes: (data.creditNotes as StoreData["creditNotes"]) ?? [],
   };
 }
diff --git a/src/store/types.ts b/src/store/types.ts
index 222b191..a9d7277 100644
--- a/src/store/types.ts
+++ b/src/store/types.ts
@@ -1,7 +1,8 @@
+import type { CreditNote } from "../domain/credit-note";
 import type { Customer } from "../domain/customer";
 import type { Invoice } from "../domain/invoice";
 
-export type Counters = { customer: number; invoice: number; invoiceNumber: number };
+export type Counters = { customer: number; invoice: number; invoiceNumber: number; creditNote: number };
 
 /** The shape of the JSON file at CURRENT_SCHEMA_VERSION. */
 export type StoreData = {
@@ -9,4 +10,5 @@ export type StoreData = {
   counters: Counters;
   customers: Customer[];
   invoices: Invoice[];
+  creditNotes: CreditNote[];
 };
diff --git a/test/invoice.test.ts b/test/invoice.test.ts
index 85202f2..97ef4d9 100644
--- a/test/invoice.test.ts
+++ b/test/invoice.test.ts
@@ -53,9 +53,10 @@ describe("invoice domain", () => {
     expect(() => markPaid(draft({ status: "void" }), NOW)).toThrow(ConflictError);
   });
 
-  it("voids drafts and open invoices, never paid ones", () => {
+  it("voids drafts and open invoices, paid ones only when fully credited", () => {
     expect(voidInvoice(draft(), NOW).status).toBe("void");
     expect(() => voidInvoice(draft({ status: "paid" }), NOW)).toThrow(ConflictError);
+    expect(voidInvoice(draft({ status: "paid" }), NOW, { fullyCredited: true }).status).toBe("void");
   });
 
   it("knows when an open invoice is overdue", () => {
diff --git a/test/numbering.test.ts b/test/numbering.test.ts
index a97d189..1fa4dd8 100644
--- a/test/numbering.test.ts
+++ b/test/numbering.test.ts
@@ -20,3 +20,34 @@ describe("credit note numbering", () => {
     expect(next.body).toMatchObject({ number: "CN-0002" });
   });
 });
+
+describe("credit notes and voiding", () => {
+  it("lists credit notes and voids a paid invoice by crediting the rest", async () => {
+    const app = testApp();
+    const customerId = await seedCustomer(app);
+    const id = await seedOpenInvoice(app, { customerId, unitPrice: 10_000 });
+    await app.handle("POST", `/invoices/${id}/pay`);
+    await app.handle("POST", `/invoices/${id}/credit-notes`, { amount: 2_500 });
+
+    const voided = await app.handle("POST", `/invoices/${id}/void`);
+    expect(voided.status).toBe(200);
+    expect(voided.body).toMatchObject({ status: "void" });
+
+    const list = await app.handle("GET", `/invoices/${id}/credit-notes`);
+    expect((list.body as { number: string; amount: number }[]).map((n) => [n.number, n.amount])).toEqual([
+      ["CN-0001", 2_500],
+      ["CN-0002", 7_500],
+    ]);
+  });
+
+  it("rejects bad amounts and unpaid invoices", async () => {
+    const app = testApp();
+    const customerId = await seedCustomer(app);
+    const id = await seedOpenInvoice(app, { customerId });
+    const url = `/invoices/${id}/credit-notes`;
+    expect((await app.handle("POST", url, { amount: 100 })).status).toBe(409);
+    await app.handle("POST", `/invoices/${id}/pay`);
+    expect((await app.handle("POST", url, { amount: 0 })).status).toBe(400);
+    expect((await app.handle("POST", url, {})).status).toBe(400);
+  });
+});
diff --git a/test/store.test.ts b/test/store.test.ts
index a1bf345..82cc59f 100644
--- a/test/store.test.ts
+++ b/test/store.test.ts
@@ -31,7 +31,7 @@ describe("store", () => {
     writeFileSync(
       file,
       JSON.stringify({
-        counters: { customer: 1, invoice: 1, invoiceNumber: 0 },
+        counters: { customer: 1, invoice: 1, invoiceNumber: 0, creditNote: 0 },
         customers: [{ id: "cus_1", name: "Acme", email: "a@acme.io", createdAt: "2025-01-01T00:00:00.000Z" }],
         invoices: [{ id: "inv_1", customer: { id: "cus_1", name: "Acme" }, status: "draft" }],
       }),

sessionB-F2.diff on GitHub

À faire lundi

  • Mettez à jour la marketplace officielle, installez le pack de skills au niveau projet, et vérifiez que la commande de rétro apparaît dans une nouvelle session.
  • Lancez votre vraie tâche dans une session ordinaire. Laissez-la finir et gardez la session ouverte.
  • Lancez la rétro dans cette session, puis relancez-la avec l'échec que vous avez trouvé écrit derrière.
  • Posez la question de traçage : "For each candidate, quote the moment of this session's transcript that backs it; drop any candidate you cannot trace to a specific moment."
  • Validez les survivants un par un dans leur support : un contrôle pour un échec mécanique, une ligne de standards pour un jugement, un commentaire à la source pour un manque de navigation.
  • Lancez /doctor prompt-audit avec le chemin des fichiers que vous avez touchés, appliquez ses modifications, puis committez les fichiers vous-même.
  • Relancez la même tâche dans une session neuve et regardez l'échec ne pas se reproduire.

Mesures

Ce sont de petits runs sur une API de facturation inventée (Node et TypeScript, TypeScript, vitest) avec Claude Code 2.1.294. C'est une démonstration, pas un benchmark. Le test caché de la finance regroupe cinq contrôles écrits avant le run.

Étape Résultat
Tâche d'avoirs, tentative 1 Propre : 40 passés, test caché de la finance 5/5 (écartée, pas la session en défaut)
Tâche d'avoirs, tentative 2 42 passés, type check exit 0, test caché de la finance 4/5. CN-0004 reçu là où CN-0002 était attendu
Première rétro, sans échec rapporté 5 éléments classés, ouverte par "This session was smooth"
Traçage de la première rétro 2 survivants, 3 écartés
Deuxième rétro, échec rapporté 5 candidats
Traçage de la deuxième rétro 4 survivants, 1 écarté
Validé à la main 4 changements : hook, test, ligne de standards, commentaire à la source. Aucune ligne dans CLAUDE.md
Première tentative de commit Le hook l'a bloquée : 1 en échec, 42 passés (43). L'arbre contenait encore le code buggé
Après le correctif d'une ligne Suite 43/43, test caché de la finance 5/5, les commits sont passés
Session neuve B, même demande 37 passés, type check exit 0, test caché de la finance 5/5, numéros sans trou après une requête rejetée

Lisez ces chiffres avec prudence :

  • Sur les deux runs ordinaires faits sur cette tâche, l'un a livré le bug avant la rétro. Après la rétro on n'a qu'une seule relance, donc c'est n=1.
  • La session B a lu le fichier de standards, le hook et le test de numérotation rouge avant d'écrire du code. On lui a donné la leçon. Sa réussite montre la boucle qui fonctionne comme prévu, pas une réplication indépendante.
  • La rétro n'a jamais trouvé le bug seule. Sa première passe a trouvé un garde-fou manquant qui, selon ses propres mots, n'aurait pas attrapé ce bug.
  • Le nouveau hook vérifie l'arbre de travail. Un git cherry-pick du commit de la rétro sur la base ne l'a pas lancé.
  • Sans chemin, le prompt audit a lu les fichiers du niveau utilisateur et n'a rien dit des nouveaux. Avec le chemin du fichier de standards, il a répondu que le fichier était propre.
  • L'ancienne route d'émission de ce dépôt fait fuiter les numéros de facture de la même façon. On l'a laissée, elle est hors périmètre pour cette session.

Aller plus loin

Lisez quelqu'un qui fait tourner une vraie boucle de leçons. Un développeur réinjecte un fichier de leçons daté au début de chaque session et transforme une règle qui glisse sans cesse en test, en règle de lint ou en hook, avec un CLAUDE.md réduit à un court routeur. s6

Lisez les issues avant de vous fier à la rétro. L'issue 1173 est l'argument pour l'étape de traçage. s7 L'issue 1311 est le commit oublié. s23 L'issue 1158 est la commande absente sur la marketplace officielle. s24

Apprenez où vivent les journaux et combien de temps ils durent. La page sur les sessions couvre le chemin des transcripts, la reprise par chemin et le nettoyage à 30 jours. s8

L'auto memory est l'alternative automatique. On l'a laissée de côté. Le coup de gueule d'un auteur soutient qu'elle écrit en silence, n'est pas versionnée et garde des détails triviaux à côté de règles durables, c'est pourquoi on garde un humain qui valide chaque ligne. s11 Prenez ça pour une opinion, et testez-la sur votre propre configuration.

/insights est le compagnon hebdomadaire. Il construit un rapport HTML sur jusqu'à 200 sessions avec les points de friction et des suggestions. s14 Un retour de terrain a transformé 64 sessions en 5 lignes de CLAUDE.md, 2 skills et des hooks. s12 On ne l'a pas lancé ici.

La boucle d'Anthropic est manuelle. Son article sur le harness décrit un réglage en lisant les journaux de l'évaluateur, en trouvant où son jugement divergeait et en mettant à jour le prompt. s13 On n'a trouvé aucune commande ni aucun article officiel sur la rétro dans la première page de l'index d'ingénierie. s27

Pourquoi on ne l'a pas automatisée. Un déclencheur de fin de session serait de la colle écrite par nous et non testée. Les hooks de fin de session ne peuvent pas bloquer l'arrêt et ont un délai par défaut de 1,5 seconde, et le skill de rétro ne s'invoque qu'à la main. s5 s1 La doc de la rétro dit elle-même qu'elle ne fait que proposer et que rien ne change tant que vous n'avez pas choisi un candidat. s2 On a gardé l'humain dans la boucle.

Ce qu'on a laissé hors périmètre. Le même bug d'ordre existe dans l'ancienne route d'émission de facture de ce dépôt, et une rétro dessus serait un run à part. On n'a pas non plus ajouté de job de CI, parce que notre copie n'avait pas de remote et qu'un workflow ne se serait jamais lancé.

FAQ

Pourquoi la rétro a-t-elle dit que ma session était smooth ? Elle lit le journal de session, et la session n'a jamais su qu'elle avait échoué : ses tests et son type check étaient au vert. Relancez-la avec ce qui a échoué écrit après la commande, comme argument.

La rétro modifie-t-elle mes fichiers ? Elle ne fait que proposer. Rien ne change tant que vous n'avez pas choisi un candidat. s2 Une fois que vous avez approuvé, les changements peuvent être implémentés mais pas committés, donc committez-les vous-même. s23

Puis-je la lancer sur une session déjà vidée ? Oui, si le journal est toujours sur le disque. Les transcripts vivent sous ~/.claude/projects/ et sont nettoyés après 30 jours par défaut, et claude --resume <transcript-path> en ouvre un. s8

Pourquoi mon nouveau hook a-t-il bloqué le commit ? Il lance votre type check et vos tests sur l'arbre de travail. Si l'arbre contient encore le bug que le test attrape désormais, le commit est refusé. Corrigez d'abord le code, puis committez.

Pourquoi aucune ligne dans CLAUDE.md ? Nos deux passes de rétro en ont refusé une, en disant qu'un pointeur coûterait du contexte à chaque tour et deviendrait périmé. La doc vise moins de 200 lignes et dit que CLAUDE.md est du contexte, pas une configuration imposée. s4

Fichiers