// agent.el — Autonomous agent. Entirely in Engram.
//
// The daemon calls on_startup() in a background thread at boot.
// on_startup() delegates immediately to loop_main() from loop.el —
// the six-tier pacemaker IS the agent loop. loop manages WHEN to tick;
// agent_tick() manages WHAT to do on each tick (called from loop_do_tick).
//
// LLM calls go through native_http_post — the agent builds the full
// Anthropic request itself. No logic is baked into Rust.
//
// Priority → model mapping:
//   P0 (critical) → claude-opus-4-5   (highest quality for blocking issues)
//   P1 (high)     → claude-sonnet-4-5 (balanced quality/speed)
//   P2 (medium)   → claude-haiku-4-5  (fast, cheap for routine work)
//   P3 (low)      → NEURON_MODEL env  (can be a local model)

from types import {
    NeuronError,
}

// ── Constants ─────────────────────────────────────────────────────────────────

// ANTHROPIC_API_BASE — override with NEURON_INFERENCE_URL env to point at a
// local proxy or a different provider (OpenAI-compat).
let ANTHROPIC_API_BASE = "https://api.anthropic.com/v1/messages"
let TICK_INTERVAL_MS = 30000

// ── Priority helpers ──────────────────────────────────────────────────────────

@accessor
fn pick_model(priority: String) -> String {
    if priority == "P0" {
        return "claude-opus-4-5"
    }
    if priority == "P1" {
        return "claude-sonnet-4-5"
    }
    if priority == "P2" {
        return "claude-haiku-4-5"
    }
    "claude-haiku-4-5"
}

@accessor
fn priority_rank(priority: String) -> Int {
    if priority == "P0" {
        return 0
    }
    if priority == "P1" {
        return 1
    }
    if priority == "P2" {
        return 2
    }
    3
}

@accessor
fn pick_highest(items: [Any]) -> Any {
    let best = items[0]
    let best_rank = priority_rank(best.priority)
    for item in items {
        let rank = priority_rank(item.priority)
        if rank < best_rank {
            let best = item
            let best_rank = rank
        }
    }
    best
}

// ── LLM call ─────────────────────────────────────────────────────────────────

// call_llm — send a single-turn message to the Anthropic Messages API.
//
// Returns the assistant's text response, or an error string if the call fails.
// The agent builds the full request — no logic is in Rust.
@accessor
fn call_llm(model: String, system: String, message: String) -> String {
    let api_key = native_shell_exec({"cmd": "echo $ANTHROPIC_API_KEY"})
    let key = api_key.stdout

    let response = native_http_post({
        "url": ANTHROPIC_API_BASE,
        "headers": {
            "x-api-key": key,
            "anthropic-version": "2023-06-01",
            "Content-Type": "application/json",
        },
        "body": {
            "model": model,
            "system": system,
            "messages": [{"role": "user", "content": message}],
            "max_tokens": 4096,
        },
    })

    if response.ok {
        let body = response.body
        let content = body.content
        let first = content[0]
        first.text
    } else {
        "error: " + response.err
    }
}

// ── Agent tick ────────────────────────────────────────────────────────────────

// agent_tick — one cycle of the autonomous agent.
//
// Idempotent: if there is nothing ready, returns {status: "idle"} with no
// side effects.
@manager
fn agent_tick(params: Map<String, Any>) -> Result<Map<String, Any>, NeuronError> {
    let items = native_list_backlog({"status": "ready"})

    if items == [] {
        return Ok({"status": "idle"})
    }

    let item = pick_highest(items)

    let item_id = item.id
    let item_title = item.title
    let item_description = item.description
    let item_priority = item.priority

    let model = pick_model(item_priority)

    let system = "You are Neuron, an autonomous AI agent with persistent memory and a mission. You are processing a backlog task autonomously. Be direct, specific, and actionable. Provide concrete output the agent can store and act on."

    let message = "Task: " + item_title + "\n\nDescription: " + item_description + "\n\nPriority: " + item_priority + "\n\nAnalyze this task and provide your best response."

    let response_text = call_llm(model, system, message)

    native_update_backlog_status({"id": item_id, "status": "in_progress"})

    native_store_memory({
        "content": "Agent processed: [" + item_title + "] (priority: " + item_priority + ", model: " + model + "). Response: " + response_text,
        "tags": ["agent-loop", "autonomous", item_priority],
        "importance": "normal",
    })

    native_emit("agent.tick_completed", {
        "task_id": item_id,
        "task_title": item_title,
        "priority": item_priority,
        "model_used": model,
    })

    Ok({
        "status": "completed",
        "task_id": item_id,
        "model_used": model,
    })
}

// ── Startup hook ──────────────────────────────────────────────────────────────

// on_startup — called by the daemon at boot in a dedicated background thread.
//
// Delegates to loop_main() from loop.el — the six-tier pacemaker is the
// agent loop. The naive while-true sleep loop has been replaced by the
// self-pacing, bell-aware, tier-stepping loop in loop.el.
//
// loop_do_tick() calls agent_tick() on each tick, so agent cognitive work
// continues to happen — just paced by the loop tier instead of a fixed interval.
//
// This function never returns (loop_main() is tail-recursive).
@manager
fn on_startup(params: Map<String, Any>) -> Result<Map<String, Any>, NeuronError> {
    native_emit("agent.started", {"module": "agent"})
    loop_main()
    Ok({"status": "stopped"})
}