mirror of
https://github.com/R0m1k3/Loki.git
synced 2026-10-12 01:37:06 +02:00
- accès web de l'IA via un serveur Crawl4AI configurable (outils
web_search / web_open / web_read / web_grep), réglage CRAWL4AI_URL +
toggle, actifs seulement si le mode agent et le serveur sont joignables.
CLI `jean internet [on|off|status|url <url>]`.
- mémoire réglable en 3 modes indépendants du mode agent : off / sur
demande / auto (proactive). CLI `jean memory [off|ondemand|always|status]`.
- fix : prompt et outils gouvernés par la même condition (plus de
web_search tapé en bash quand le crawl est injoignable) ; consigne
année de recherche durcie ; icône 🌐 des outils web dans l'UI.
197 lines
8.2 KiB
Go
197 lines
8.2 KiB
Go
package main
|
|
|
|
import (
|
|
"context"
|
|
"fmt"
|
|
"os"
|
|
"os/exec"
|
|
"os/user"
|
|
"runtime"
|
|
"strconv"
|
|
"strings"
|
|
"time"
|
|
)
|
|
|
|
const (
|
|
toolDefaultTimeout = 30
|
|
toolMaxTimeout = 300
|
|
toolMaxOutput = 8000 // characters of stdout/stderr returned to the model
|
|
)
|
|
|
|
// baseSystemPrompt is the always-on system preamble. Structured like pi's
|
|
// proven prompt (identity → method → guidelines → concision → date): a concrete,
|
|
// procedural prompt gives the model rails so it stops deliberating forever
|
|
// ("Wait, let me check… Wait, I'll just run it…") and commits to an action.
|
|
// The per-tool "Outil disponible" sections live in machine/skills prompts so
|
|
// they only appear when the matching feature is on.
|
|
func baseSystemPrompt(caps Caps) string {
|
|
hasMem := caps.Mem != MemOff
|
|
// No tool access at all → no agentic preamble. A plain chat model told to
|
|
// "call tools immediately" hallucinates textual tool calls (e.g.
|
|
// default_api:bash) that leak into the answer. Let the user's own system
|
|
// prompt stand alone. (Internet requiert l'agent, donc pas testé ici.)
|
|
if !caps.Agent && !hasMem {
|
|
return ""
|
|
}
|
|
var b strings.Builder
|
|
// Prompt VOLONTAIREMENT court. Un préambule verbeux (longue liste de
|
|
// « guidelines », surtout des méta-instructions sur la réflexion) fait
|
|
// sur-raisonner les modèles à reasoning (Qwen3) : ils émettent leur <think>
|
|
// puis le token de fin SANS appeler d'outil (~25-45 % de tours « morts »
|
|
// mesurés). Une version courte et directe ramène ça à 0 %. NE PAS regonfler.
|
|
b.WriteString("You are jean, an expert assistant operating directly on this machine with real tools.")
|
|
if caps.Mem == MemAlways {
|
|
b.WriteString(" You evolve with every conversation: you actively maintain a persistent memory so nothing useful is lost between sessions.")
|
|
}
|
|
b.WriteString("\n\nTools:\n")
|
|
if caps.Agent {
|
|
b.WriteString("- bash — run a shell command on this machine (inspect files, processes, logs, run scripts).\n")
|
|
b.WriteString("- edit — patch a file by exact replacement (old → new, old must be unique).\n")
|
|
}
|
|
if hasMem {
|
|
b.WriteString("- mem_search / mem_read / mem_add / mem_edit — your persistent Markdown memory under MEMORY/.\n")
|
|
}
|
|
// Politique d'usage de la mémoire selon le mode.
|
|
switch caps.Mem {
|
|
case MemAlways:
|
|
b.WriteString("\nManaging your memory is part of the job, not optional:\n")
|
|
b.WriteString("- When the user tells you to remember something, or shares a preference, fact, decision, or how-to worth keeping, save it with mem_add (or mem_edit to update an existing page) — do it on your own, without being asked.\n")
|
|
b.WriteString("- Before doing any task or answering, first call mem_search to see whether your memory already holds the answer or how to do it, then mem_read the best page. Do this even when the request has new specifics like a name, a place or a value — your saved method still applies, only the parameter changes.\n")
|
|
case MemOnDemand:
|
|
b.WriteString("\nMemory is ON-DEMAND: you have the mem_* tools but do NOT read or write memory on your own. Call mem_search/mem_read only when the user explicitly asks you to recall or look something up, and mem_add/mem_edit only when the user explicitly asks you to remember something. Otherwise leave memory untouched and answer directly.\n")
|
|
}
|
|
if caps.Agent {
|
|
b.WriteString("\nFor anything about the system or files, use bash instead of guessing. Act immediately — call the right tool, then answer. Never end your turn after only thinking. Be concise.\n")
|
|
if caps.Mem == MemAlways {
|
|
b.WriteString("Before answering any question about yourself or this machine, always call mem_search first — even trivial-seeming ones. Testing with a tool never replaces this: memory may hold context the tool won't reveal. Search memory, then verify, then answer.\n")
|
|
}
|
|
}
|
|
if caps.Internet {
|
|
b.WriteString("\nWeb access (Crawl4AI): web_search (DuckDuckGo), web_open (fetch a URL → metadata + outline), web_read (read a line range of an opened URL), web_grep (regex in an opened URL). Workflow: web_open first, then web_read/web_grep.\n")
|
|
year := time.Now().Format("2006")
|
|
b.WriteString("Your training data is stale. For ANY question about recent/latest/current things (releases, versions, news, prices, scores, fixtures, 'since when') call web_search BEFORE writing any date or version. Your answer must match the dates/facts you actually read.\n")
|
|
b.WriteString("SEARCH QUERY YEAR RULE: today is in " + year + ". If your query includes a year, use ONLY " + year + " — NEVER write a past year like " + prevYear(year) + " that you remember from training; it silently biases results toward stale pages. Default: put no year at all and let the freshest result win. Don't hedge ('probably', 'I think') about a fact a tool can verify — search instead.\n")
|
|
}
|
|
b.WriteString("\nDate: " + time.Now().Format("2006-01-02"))
|
|
return b.String()
|
|
}
|
|
|
|
// prevYear returns the year before the given "2006"-formatted year string, used
|
|
// to name explicitly the stale year the model must NOT put in search queries.
|
|
func prevYear(year string) string {
|
|
n, err := strconv.Atoi(year)
|
|
if err != nil {
|
|
return year
|
|
}
|
|
return strconv.Itoa(n - 1)
|
|
}
|
|
|
|
// machineSystemPrompt returns a short briefing about the host the model is
|
|
// running on, so that when machine access is enabled it knows *which* machine
|
|
// run_shell acts upon (and doesn't claim it has no access to "your PC").
|
|
// Returns "" when machine access is off.
|
|
func machineSystemPrompt(caps Caps) string {
|
|
if !caps.Agent {
|
|
return ""
|
|
}
|
|
host, _ := os.Hostname()
|
|
if host == "" {
|
|
host = "unknown"
|
|
}
|
|
who := ""
|
|
if u, err := user.Current(); err == nil {
|
|
who = u.Username
|
|
}
|
|
cwd, _ := os.Getwd()
|
|
|
|
var b strings.Builder
|
|
b.WriteString(fmt.Sprintf("Machine: host=%s, %s/%s", host, runtime.GOOS, runtime.GOARCH))
|
|
if who != "" {
|
|
b.WriteString(", user=" + who)
|
|
}
|
|
if cwd != "" {
|
|
b.WriteString(", cwd=" + cwd)
|
|
}
|
|
b.WriteString(".")
|
|
return b.String()
|
|
}
|
|
|
|
// runShell executes a command via the platform shell (bash -c on Unix, cmd /C
|
|
// on Windows — see newShellCmd in platform_*.go) with a clamped timeout,
|
|
// returning a single string formatted "exit: N\n\nstdout:\n...\n\nstderr:\n..."
|
|
// truncated to keep tool output bounded.
|
|
func runShell(command string, timeoutSec int) string {
|
|
if timeoutSec <= 0 {
|
|
timeoutSec = toolDefaultTimeout
|
|
}
|
|
if timeoutSec > toolMaxTimeout {
|
|
timeoutSec = toolMaxTimeout
|
|
}
|
|
ctx, cancel := context.WithTimeout(context.Background(), time.Duration(timeoutSec)*time.Second)
|
|
defer cancel()
|
|
cmd := newShellCmd(ctx, command)
|
|
var stdout, stderr strings.Builder
|
|
cmd.Stdout = &stdout
|
|
cmd.Stderr = &stderr
|
|
err := cmd.Run()
|
|
if ctx.Err() == context.DeadlineExceeded {
|
|
return fmt.Sprintf("[timeout après %ds]", timeoutSec)
|
|
}
|
|
exit := 0
|
|
if err != nil {
|
|
if ee, ok := err.(*exec.ExitError); ok {
|
|
exit = ee.ExitCode()
|
|
} else {
|
|
return fmt.Sprintf("[erreur: %v]", err)
|
|
}
|
|
}
|
|
out := tailRunes(stdout.String(), toolMaxOutput)
|
|
errOut := tailRunes(stderr.String(), toolMaxOutput)
|
|
parts := []string{fmt.Sprintf("exit: %d", exit)}
|
|
if out != "" {
|
|
parts = append(parts, "stdout:\n"+out)
|
|
}
|
|
if errOut != "" {
|
|
parts = append(parts, "stderr:\n"+errOut)
|
|
}
|
|
return strings.Join(parts, "\n\n")
|
|
}
|
|
|
|
// fileEdit applies a single exact-text replacement to a file on disk: oldText
|
|
// must appear EXACTLY once (otherwise it errors), so the model can patch a file
|
|
// without rewriting it whole. Returns a short status string for the tool result.
|
|
func fileEdit(path, oldText, newText string) string {
|
|
if strings.TrimSpace(path) == "" {
|
|
return "[erreur] chemin vide"
|
|
}
|
|
if oldText == "" {
|
|
return "[erreur] old vide"
|
|
}
|
|
b, err := os.ReadFile(path)
|
|
if err != nil {
|
|
return "[erreur] " + err.Error()
|
|
}
|
|
content := string(b)
|
|
n := strings.Count(content, oldText)
|
|
if n == 0 {
|
|
return "[erreur] old introuvable dans le fichier"
|
|
}
|
|
if n > 1 {
|
|
return fmt.Sprintf("[erreur] old apparaît %d fois — ajoute du contexte pour le rendre unique", n)
|
|
}
|
|
updated := strings.Replace(content, oldText, newText, 1)
|
|
if err := os.WriteFile(path, []byte(updated), 0o644); err != nil {
|
|
return "[erreur] " + err.Error()
|
|
}
|
|
return fmt.Sprintf("[ok] %s modifié (1 remplacement)", path)
|
|
}
|
|
|
|
// tailRunes returns the last n runes of s (used to cap tool output).
|
|
func tailRunes(s string, n int) string {
|
|
r := []rune(s)
|
|
if len(r) <= n {
|
|
return s
|
|
}
|
|
return string(r[len(r)-n:])
|
|
}
|