Lesson 79 — Intermediate

Tool Context Management

Har tool call do tarah se context khata hai — definitions (schemas) har call par, results har turn par. Is lesson mein dono kharche control karte hain: dynamic exposure, output filtering aur observation decay.

Tools ka apna budget

Yeh level ki sabse common mistake: "tools toh free hain." Nahi — schema har call jaata hai aur har result turn mein baithta hai. Dono ka audit karo (L6.1 wala budget yahin lagta hai).

Dynamic tool exposure

Model ko saare tools dene ka matlab har call par poora schema bhejna. Relevant subset hi expose karo:

• Tools ko tags se group karo (research, data, comms) — task se matching subset.
• Task badal toh tool set bhi badal — definitions turn ke saath chalein.
• Discovery ke liye ek chhota list_tools meta-tool — poora registry hamesha loaded nahi.
• Tool use ho chuka toh definition hata do; summary mein naam + usage rehne do.

REGISTRY = {
    "search": Tool(tags={"research"}),
    "sql": Tool(tags={"data"}),
    "email": Tool(tags={"comms"}),
}

def tools_for(task):
    need = set(task.tags)                  # sirf is step ke tools
    return [t for t in REGISTRY.values() if t.tags & need]

Bade outputs filter karo

• Top-k + fields: sirf relevant rows, sirf maangi columns.
• Dedupe: repeat entries hatao; sort by relevance/date.
• Caps: max items/lines per result — cap ke baad cursor.
• Errors: status + short message; stack trace sirf log mein.
• Full result external store: model ko path/pointer + slim view.

def handle(result):
    full = store.put(result)                    # bahar, pura saved
    return {"path": full, "view": result.items[:5],
            "total": len(result.items), "next": result.cursor}

Observation build-up rokna

Lambe loop mein har step ka observation window mein rehta jaata hai — purane stale ho jaate hain, naye kabhi chadh nahi paate. Yeh silent bloat hai.

• Har N steps par raw observations summarize karke facts bana do.
• Facts nikal chuke raw outputs hata do (source pointer rakho).
• Per-tool observation cap — ek hi tool ka repeat dump baar baar nahi.
• Long-term notes mein move karo (L6.4) — window mein sirf active rehne do.

Exercise

Common mistakes

• Sab tools hamesha on: 50 schemas har call = poora ek budget section khali. Dynamic exposure se turant bachat hota hai.
• Raw dump wapas bhejna: Ek hi API call ne window aadha khaya — slim preprocessor har tool result par mandatory hai.
• Observations kabhi hataana nahi: Loop chal chal ke junk pile hota hai. Periodic summarize + caps ke bina agent puraane notes mein dub jaata hai.

active = tools_for(step)           # sirf zaroori tools exposed
ctx.add(slim(run(step), cap=50))   # result cap ke saath
if step_no % 5 == 0:
    ctx.summarize_observations()   # hafta waali safai