fix: am ai avec aichat >= 0.30 — rôles index.yaml + agents.txt, contexte filtré, generation --code (v1.1.5)

- roles.rs: aichat 0.30 a remplace les agents markdown agents/<name>.md
  par functions/agents/<name>/index.yaml + registre functions/agents.txt;
  generate_all ecrit les DEUX formats (compat <= 0.29 et >= 0.30)
- ai_cmd.rs: role_args detecte la definition index.yaml en premier
  ('Unknown agent am-copilot' corrige) ; contexte par defaut (dossier
  courant) filtre par am (.git, target, node_modules, binaires non-UTF-8,
  .env*, > 64 KiB, budget 128 KiB, 40 fichiers max) — aichat ne filtre
  rien et echouait sur .git/index ; exec mode utilise le role am-do
  par defaut (commande brute)
- shell_ai.rs: generation via --code au lieu de --execute --dry-run
  (aichat 0.30: --dry-run echo la requete sans appeler l'API)
- version 1.1.5, man pages et ROADMAP maj
This commit is contained in:
2026-08-21 12:22:18 -04:00
parent 66a3eebbd6
commit e637c066fe
8 changed files with 256 additions and 44 deletions
+10 -3
View File
@@ -7,8 +7,9 @@
//! - conversational mode (no `--exec`): `aichat -f <ctx> "<prompt>"` is
//! spawned in the foreground, exactly as the user would run it;
//! - exec mode (`--exec`): aichat is asked to generate the command with
//! `--dry-run` (it prints the command and never runs it), then the
//! generated command is classified (safe / risky), a confidence score
//! `--code` (it prints only the command and never runs it — aichat is
//! invoked without `--execute`), then the generated command is
//! classified (safe / risky), a confidence score
//! is estimated, the configured safety policy is applied (dry-run by
//! default), and only then — after confirmation when required — is the
//! command executed through the user's shell.
@@ -283,7 +284,13 @@ pub fn generate(
role_args: &[String],
) -> Result<String> {
let agent = require_agent(app, "aichat")?;
let mut extra_args: Vec<String> = vec!["--exec".to_string(), "--dry-run".to_string()];
// aichat >= 0.30 renamed --exec to --execute AND changed --dry-run to
// skip the API call entirely (it echoes the request), so the old
// `--execute --dry-run` no longer yields the generated command. Plain
// chat with `--code` extracts and prints only the command (same output
// shape, works on every aichat version). aichat NEVER executes here —
// am enforces the safety policy before running the command itself.
let mut extra_args: Vec<String> = vec!["--code".to_string()];
extra_args.extend(role_args.iter().cloned());
for f in files {
extra_args.push("-f".to_string());