Mode code : le builder ne re-planifie plus à chaque tour, étiquettes d'outils

Le raisonnement n'étant pas conservé entre les tours (coût de contexte),
un modèle à reasoning fort re-déduisait TOUT son plan avant chaque appel
d'outil — des minutes de <think> pour écrire un fichier de 200 octets. Le
prompt du builder impose maintenant : plan écrit UNE fois dans PLAN.md,
une à deux phrases de réflexion entre les appels, chaque fichier complet
en un seul write.

UI : les outils du mode code (criteria, read, grep, glob, git_*, ask,
bash_bg/bash_tail) ont leurs étiquettes — ils retombaient tous sur
« mémoire ».

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
This commit is contained in:
MichaelandClaude Opus 5 committed 2026-08-18 22:53:23 +02:00
1 parent a7587235c6
commit 34218fae52
3 files changed
+24

No files matched your search

+2
View File
@@ -8,6 +8,8 @@ Method — in this order:
5. Report briefly what changed and how you verified it.
Rules:
- PLAN ONCE. If the task needs a plan, write it to PLAN.md in ONE short write, then follow it. Between tool calls, think one or two sentences at most — NEVER restate or re-derive the plan: it is in PLAN.md and in your criteria, read them instead.
- Write each file COMPLETE in a single write call. Many small writes waste turns.
- You may NOT mark a criterion passed — the verification pass does that.
- If a command fails, read the error and fix the cause; do not retry the same command unchanged.
- Ask the user (ask tool) only for decisions that are genuinely theirs; decide the rest yourself.
+11
View File
@@ -5377,6 +5377,17 @@ function renderToolMsg(el, tu){
web_open: {lbl:'page web', head:'ouverture'},
web_read: {lbl:'page web', head:'lecture'},
web_grep: {lbl:'page web', head:'recherche'},
// Outils du mode code : chacun son étiquette, pas le fallback mémoire.
read: {lbl:'lecture', head:'lecture'},
grep: {lbl:'recherche', head:'grep'},
glob: {lbl:'fichiers', head:'glob'},
criteria: {lbl:'critères', head:'critères'},
ask: {lbl:'question', head:'question'},
bash_bg: {lbl:'terminal', head:'commande (fond)'},
bash_tail: {lbl:'terminal', head:'suivi de job'},
git_status: {lbl:'git', head:'git status'},
git_diff: {lbl:'git', head:'git diff'},
git_clone: {lbl:'git', head:'git clone'},
};
// Outils MCP (nom mcp__<serveur>__<outil>) : en-tête = nom du serveur, libellé lisible,
// pas le fallback mémoire. On extrait serveur et outil du nom namespacé.
+11
View File
@@ -258,6 +258,17 @@ function renderToolMsg(el, tu){
web_open: {lbl:'page web', head:'ouverture'},
web_read: {lbl:'page web', head:'lecture'},
web_grep: {lbl:'page web', head:'recherche'},
// Outils du mode code : chacun son étiquette, pas le fallback mémoire.
read: {lbl:'lecture', head:'lecture'},
grep: {lbl:'recherche', head:'grep'},
glob: {lbl:'fichiers', head:'glob'},
criteria: {lbl:'critères', head:'critères'},
ask: {lbl:'question', head:'question'},
bash_bg: {lbl:'terminal', head:'commande (fond)'},
bash_tail: {lbl:'terminal', head:'suivi de job'},
git_status: {lbl:'git', head:'git status'},
git_diff: {lbl:'git', head:'git diff'},
git_clone: {lbl:'git', head:'git clone'},
};
// Outils MCP (nom mcp__<serveur>__<outil>) : en-tête = nom du serveur, libellé lisible,
// pas le fallback mémoire. On extrait serveur et outil du nom namespacé.