vivac 0.6.4

//! The two session hooks. `ROADMAP.md` §4.
//!
//! `session start` injects the brief and `session end` leaves an automatic
//! stop. They are the **seams of the session**, the way `push`/`pop` are the
//! seams of the work: they ask for no judgement of relevance, they just happen.
//!
//! Both exit 0 and say nothing when there is no `.vivac/`. A hook that fails
//! in every directory without a tree gets switched off within two days.

use crate::args::Args;
use crate::event::VivacKind;
use crate::failure::{Failure, R};
use crate::output::outln;

/// Wraps text in the envelope Claude Code injects into the context. It is a
/// single JSON line, with no external dependency and no `jq` in between: a
/// hook with a pipe is a hook that breaks on the first different machine.
fn envelope(event_name: &str, text: &str) -> String {
    serde_json::json!({
        "hookSpecificOutput": {
            "hookEventName": event_name,
            "additionalContext": text,
        }
    })
    .to_string()
}

/// What the hook is handed on stdin.
///
/// Two fields are read and the rest is left where it is. `transcript_path`
/// travels in this same payload and it is the tempting one --it is what really
/// links the tree to the conversation-- but it carries the user's home
/// directory, and the security pillar vetoes that without negotiation. An
/// opaque identifier yes; a path into somebody's filesystem no.
struct HookInput {
    source: String,
    session: Option<String>,
}

impl HookInput {
    fn read() -> HookInput {
        use std::io::IsTerminal;
        let mut raw = String::new();
        // A terminal has no payload to give, and reading one would hang the
        // hook waiting for an EOF that never comes.
        if !std::io::stdin().is_terminal() {
            use std::io::Read;
            std::io::stdin().read_to_string(&mut raw).ok();
        }
        let v: serde_json::Value = serde_json::from_str(&raw).unwrap_or(serde_json::Value::Null);
        HookInput {
            // `unknown` and not an empty string, so that reading it later tells
            // "it did not say" apart from "we did not look".
            source: v
                .get("source")
                .and_then(|s| s.as_str())
                .unwrap_or("unknown")
                .to_string(),
            session: v
                .get("session_id")
                .and_then(|s| s.as_str())
                .map(str::to_string),
        }
    }
}

pub fn start(ctx: &mut crate::ops::Ctx, a: &Args, project: &str) -> R {
    if !a.has("hook") {
        return crate::brief::brief(&ctx.tree, ctx.anchor.as_ref(), a, project);
    }
    // In hook mode the brief is captured and emitted inside the envelope. No
    // loose noise on stdout: what is not in the envelope, the agent never sees.
    let text = crate::brief::to_text(&ctx.tree, ctx.anchor.as_ref(), a, project)?;
    outln!("{}", envelope("SessionStart", &text));
    // The brief goes out **first**, and the write cannot take it down. A
    // failure that left the agent with no brief would turn a hole in the
    // instrument into blindness in the product, which is a far worse trade: a
    // log missing an opening shows up on reading, an agent missing its brief
    // does not show up until the thread is already lost.
    let hook = HookInput::read();
    crate::ops::session_started(ctx, &hook.source, hook.session).ok();
    Ok(())
}

pub fn end(ctx: &mut crate::ops::Ctx, a: &Args) -> R {
    // With no stack there is no thread to close, and an empty vivac is just
    // noise to be pruned later.
    if ctx.tree.stack.is_empty() {
        if !a.has("hook") {
            outln!("  Empty stack: no stop worth saving.");
        }
        return Ok(());
    }
    // Nor with nothing new. Claude Code's `Stop` hook runs **on every turn**,
    // not when the session closes: there is no end-of-session event (`f35`).
    // Without this guard it would be forty identical stops a day, and a stop
    // that repeats is not a stop, it is a log.
    if ctx.tree.seq_change <= ctx.tree.seq_vivac {
        if !a.has("hook") {
            outln!("  Nothing changed since the last stop.");
        }
        return Ok(());
    }
    let next = a.opt_or("next");
    let label = segment_label(&ctx.tree);
    let num = ctx.tree.next_vivac_num.max(1);
    crate::ops::auto_vivac(ctx, VivacKind::Auto, &next, &label)?;
    if !a.has("hook") {
        outln!("  v{num}  automatic stop at session close");
    }
    Ok(())
}

/// What the segment being closed contained, counted off the seams.
///
/// The other four kinds of stop are written by somebody who knows what they
/// were doing, and they all carry a `next_intent`. The automatic one is
/// written by a hook that was never told: asking the agent for the intent is
/// the judgement of relevance `DX` already measured at zero uses. So it
/// carries what it can know without asking --how much the segment held-- and
/// leaves `next_intent` honestly empty (`f59`).
fn segment_label(t: &crate::model::Tree) -> String {
    let mut parts = Vec::new();
    if t.seg_new > 0 {
        parts.push(format!("{} new", t.seg_new));
    }
    if t.seg_closed > 0 {
        parts.push(format!("{} closed", t.seg_closed));
    }
    if t.seg_notes == 1 {
        parts.push("1 note".to_string());
    } else if t.seg_notes > 1 {
        parts.push(format!("{} notes", t.seg_notes));
    }
    if parts.is_empty() && t.seg_events > 0 {
        parts.push(if t.seg_events == 1 {
            "1 change".to_string()
        } else {
            format!("{} changes", t.seg_events)
        });
    }
    parts.join(", ")
}

pub fn dispatch(ctx: &mut crate::ops::Ctx, a: &Args, project: &str) -> R {
    match a.positional(0) {
        Some("start") => start(ctx, a, project),
        Some("end") => end(ctx, a),
        _ => Err(Failure::usage("usage: vivac session start|end [--hook]")),
    }
}

/// `vivac hooks` — prints what to paste, and touches nobody's configuration.
/// Writing to the user's settings is an action you ask for, not one that
/// happens by surprise.
pub fn hooks() -> R {
    outln!(
        r#"
  Paste this into the project's .claude/settings.json:

  {{
    "hooks": {{
      "SessionStart": [
        {{ "matcher": "*", "hooks": [{{ "type": "command", "command": "vivac session start --hook" }}] }}
      ],
      "Stop": [
        {{ "hooks": [{{ "type": "command", "command": "vivac session end --hook" }}] }}
      ]
    }}
  }}

  SessionStart injects the brief into the agent's context, and fires again
  after a compaction -- exactly when the brief has the most work to do
  picking the thread back up.
  Stop leaves an automatic stop with the stack as it stood.

  Both stay quiet and exit 0 where there is no .vivac/, so they can be left
  in the global configuration without getting in the way of other projects.
"#
    );
    Ok(())
}