Files
ai-app/server/src/session/driver.rs
T
iris-aiandClaude Opus 5 81ab564a09 Declare provider settings, and give the context figure a denominator
Two things a session could not say, and one it was saying wrongly.

**Every provider setting is reachable.** `-np 1`, the MTP draft depth, the
tool set, the sampling parameters -- most were hardcoded to what measured
best on this machine, which is right as a default and wrong as a constant:
the next machine has a different GPU and a different core count, and
nobody running this app can edit the source. `DriverKind::params` now
declares what a provider takes -- key, label, shape, what blank means, and
whether a change waits for a restart -- and the phone renders whatever
arrives, on the spawn form and in the session settings dialog. Adding a
setting to a driver is one entry in that table and no app change.
`POST /sessions/{id}/params` takes the whole map, so an absent key is the
instruction to unset; the sampling half applies at once and the session is
told in words which of the rest are waiting for a restart.

`tools` is one of them, because it is the biggest lever on a tight
context: the seven built-in definitions are ~1,300 tokens of every prompt
(2,191 against 887 with none). `"none"` omits the flag rather than passing
it on, since `--tools none` is `unknown tool "none"` and a server that
exits.

**The context figure has a denominator.** `Event::ContextWindow` carries
it, read from `llama-server`'s `/props` once the model is up -- the
measurement rather than the request, since a session that named no context
size gets the model's own. Neither coding CLI states its window, so those
keep the bare figure: "2,042" and "2,042 / 8,192" are deliberately
different-looking, and a missing ceiling is never drawn as a proportion of
an assumed one.

**And the numerator was wrong**, by the length of the last reply: it was
the prompt alone, so a five-word answer reported 2,042 against a slot
holding 2,355. It is the turn's total now, which matches `llama-server`'s
own `n_tokens` to within a token.

Two defects the review found, both of which would have shipped: changing
settings on a *stopped* session reported "no process running, so it can't
take new settings", when a stopped session is exactly when you would set
them for the next start; and `GET /tools` answers **403** rather than an
empty list on a server started without `--tools`, so reading it as a
failure made the no-tools session one that never started.

Verified against real models: settings spawned and changed live, the
restart note, a session with two tools and one with none, and the counter
checked against the server's own slot occupancy each time.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
2026-09-19 13:58:08 -04:00

909 lines
40 KiB
Rust

//! The common event model and the `Driver` trait -- the one abstraction
//! everything hangs off (see PLAN.md).
//!
//! A driver translates its child process's JSONL dialect into [`Event`]s
//! and accepts the small inbound vocabulary below. The transcript, the SSE
//! stream, and the phone UI work purely in this model; nothing downstream
//! of a driver may branch on the session kind.
use serde::{Deserialize, Serialize};
use tokio::sync::mpsc;
/// The name a session's image is stored and served under -- minted for an
/// upload or for one a tool produced, and fetched back from
/// `/sessions/{id}/files/{ref}`. One id both directions, so the transcript
/// renders them identically.
pub type ImageRef = String;
/// Stores image bytes where the files route serves them and returns their transcript reference.
/// Both CLI dialects produce images in different envelopes; the durable file and naming rule are
/// part of the common event model and must not vary with that envelope.
pub(in crate::session) fn store_image(
session_dir: &std::path::Path,
media_type: &str,
bytes: &[u8],
) -> Option<ImageRef> {
let extension = crate::media::extension_for(media_type).unwrap_or("png");
let name = format!("{}.{extension}", super::random_hex());
let dir = session_dir.join("files");
if let Err(err) = wg_app_link::private::create_dir(&dir)
.map_err(std::io::Error::other)
.and_then(|()| std::fs::write(dir.join(&name), bytes))
{
tracing::error!("couldn't save produced image: {err}");
return None;
}
Some(name)
}
/// Prefixed to a message that was typed while a turn was already running.
///
/// A steer goes to the CLI the moment it arrives, but *when the model reads
/// it* is not ours to decide: it lands at the next model call if there is
/// one, and a turn that ends first delivers it as the opening line of the
/// next turn instead -- Claude's from the fifo, Codex's requeued after
/// `activeTurnNotSteerable`. Read there it looks like a reply to the answer
/// just given, so the model acts as if the person had seen that answer, which
/// is exactly what they had not. Nothing else distinguishes the two cases by
/// the time the model sees them, so the note is the only thing that can carry
/// the fact.
const STEERING_NOTE: &str = "[Sent while you were still working, so it was written without \
having seen the rest of that turn. Treat it as steering the work in progress, not as a \
reply to anything you said after it was sent.]";
/// The text a CLI receives for one message: the note above when this is a
/// steer, the typed words, and the paths of any attachment the model has to
/// read from disk rather than being handed inline.
///
/// The paths come after the words because a model reading a list of files
/// before the request treats the list as the request.
pub(in crate::session) fn message_body(
text: &str,
files: &[std::path::PathBuf],
steering: bool,
) -> String {
let mut body = String::new();
if steering {
body.push_str(STEERING_NOTE);
}
for part in std::iter::once(text.to_string()).chain(
files
.iter()
.map(|path| format!("Attached file: {}", path.display())),
) {
if part.is_empty() {
continue;
}
if !body.is_empty() {
body.push_str("\n\n");
}
body.push_str(&part);
}
body
}
/// The name an upload is stored and served under: an image is
/// `<hex>.<extension>` and is an [`ImageRef`] like any other; any other file
/// keeps its own name after the hex, `<hex>-<name>`, because the name is what
/// the reader attached and what the session is told. Told apart by
/// `crate::media::media_type_for`.
pub type AttachmentRef = String;
/// One choice offered in answer to a [`Event::Question`]. More than a label
/// because the reader is deciding rather than confirming: what an option
/// means, and what picking it would produce, are what decide it.
#[derive(Debug, Clone, PartialEq, Serialize, Deserialize)]
#[serde(rename_all = "camelCase")]
pub struct QuestionOption {
pub label: String,
/// A sentence about what this option means.
#[serde(default, skip_serializing_if = "Option::is_none")]
pub description: Option<String>,
/// A block to show as written -- a mockup, a diff, a config file.
#[serde(default, skip_serializing_if = "Option::is_none")]
pub preview: Option<String>,
}
impl QuestionOption {
pub fn plain(label: impl Into<String>) -> Self {
Self {
label: label.into(),
description: None,
preview: None,
}
}
}
/// Everything a session can tell the outside world. Every event is appended
/// to the transcript with a sequence number, then fanned out to SSE
/// subscribers, so reconnecting is just "events after seq N" -- no separate
/// history path to drift from the live one.
#[derive(Debug, Clone, PartialEq, Serialize, Deserialize)]
// `rename_all` renames the variants; `rename_all_fields` renames what is
// inside them. Both are needed and only the first is obvious: every field
// here was one lowercase word until `pre_tokens` arrived, so a multi-word
// field went out as snake_case, the app looked for camelCase and found
// nothing, and the event still rendered -- as the "no counts reported" case,
// which is a state it is allowed to be in.
#[serde(
tag = "type",
rename_all = "camelCase",
rename_all_fields = "camelCase"
)]
pub enum Event {
/// What the user sent, written into the transcript by the manager (not by
/// drivers) so every device renders the conversation from one stream.
/// Recorded when the session reads it, which is what `MessageTaken` reports.
UserMessage {
/// The [`Event::MessageQueued`] this resolves, when it waited. The
/// phone has a bubble on screen for the waiting message and needs to
/// know *which* one this is, rather than matching on the text and
/// clearing the wrong one when the same thing was sent twice.
#[serde(default, skip_serializing_if = "Option::is_none")]
id: Option<String>,
text: String,
/// What was attached, by the ref the files route serves. On the
/// message rather than beside it: these used to be their own `Image`
/// events just before, which left the phone deciding from adjacency
/// which message an image belonged to. `images` on disk until
/// 2026-09-03, when files joined them; the alias reads the older rows.
#[serde(default, alias = "images", skip_serializing_if = "Vec::is_empty")]
attachments: Vec<AttachmentRef>,
},
/// A message accepted from the phone that the session cannot read yet.
///
/// Recorded, unlike the message itself, and that difference is the point:
/// the message belongs in the transcript where the session read it, but
/// something has to say it is waiting, and it has to be the server. The
/// phone used to remember its own outgoing messages, so leaving the
/// screen showed nothing pending when something was.
///
/// Carries no row of its own; resolved by the `UserMessage` bearing the
/// same id, as `CommandQueued` is resolved by `CommandSent`.
MessageQueued {
id: String,
text: String,
/// Carried for the same reason [`Event::UserMessage`] carries it,
/// and it matters more here: a waiting message is on screen for as
/// long as the turn runs, so its attachment has nowhere else to be.
#[serde(default, alias = "images", skip_serializing_if = "Vec::is_empty")]
attachments: Vec<AttachmentRef>,
},
/// A message taken out of the queue before the session read it.
///
/// Recorded for the same reason `MessageQueued` is: the queue is the
/// server's, so what is waiting has to be answerable from the transcript
/// alone. Without it a phone that reconnects replays the `MessageQueued`
/// and puts back a bubble nothing will ever resolve -- the `UserMessage`
/// that normally does is exactly what is not coming.
///
/// Only ever sent for a message that had not been handed over; see
/// [`Unqueued::AlreadySent`].
MessageDropped {
id: String,
},
/// A driver has taken one of the user's messages and started reading it.
/// The manager turns this into the `UserMessage` above, so it never
/// reaches a phone itself.
///
/// It exists because sending and being read are not the same moment. A
/// message sent into a running turn waits, and recording it among things
/// already read puts it in the transcript above output that predates it.
MessageTaken {
/// The `MessageQueued` this answers, or `None` when it never waited.
/// Carried through onto the `UserMessage`.
id: Option<String>,
text: String,
#[serde(default, alias = "images", skip_serializing_if = "Vec::is_empty")]
attachments: Vec<AttachmentRef>,
},
/// Streaming assistant text; the phone renders the concatenation as
/// markdown.
AssistantText {
delta: String,
},
/// The authoritative text of the assistant message whose deltas immediately
/// precede this event.
///
/// Some providers stream a provisional rendering and revise it before the
/// item completes. This stays append-only like every other transcript
/// correction: readers replace the open message rather than editing an old
/// line, and replay therefore reaches the same text as the live stream.
AssistantTextFinal {
text: String,
},
ToolStart {
id: String,
tool: String,
input: serde_json::Value,
},
ToolUpdate {
id: String,
output: String,
},
ToolEnd {
id: String,
output: String,
},
/// An image the session produced or was sent, saved under the session
/// dir and referenced by id; the phone fetches it by URL.
Image {
#[serde(rename = "ref")]
image: ImageRef,
/// The tool call whose result carried it, when one did. A screenshot
/// belongs under the call that took it, not floating beside it -- the
/// reader has to pair them by position otherwise, and position is
/// exactly what a page boundary breaks.
#[serde(default, skip_serializing_if = "Option::is_none")]
about: Option<String>,
},
/// Anything the session needs a human for: AskUserQuestion, and
/// permission requests, are the same shape with different options.
Question {
id: String,
prompt: String,
/// A few words naming what the question is about, when the asker
/// offered one. `None` for a permission, which is about the call
/// above it.
header: Option<String>,
options: Vec<QuestionOption>,
/// Whether several options may be chosen at once. Here rather than
/// left for a phone to work out from the dialect underneath: how many
/// answers a question takes is a fact about the question, and the
/// alternative was Claude Code's tool-input schema written out a
/// second time in Kotlin, where no other dialect could reach it.
#[serde(default, skip_serializing_if = "std::ops::Not::not")]
multi_select: bool,
/// The tool call this is permission for, when it is one, so a phone
/// can draw the ask on the tool's own row rather than as a second
/// card repeating its input. `None` for anything not about a tool.
#[serde(default, skip_serializing_if = "Option::is_none")]
about: Option<String>,
},
/// A message another agent sent this session.
///
/// Its own kind rather than a `UserMessage`, because it is not something
/// the reader said and a transcript that renders it in their voice is
/// claiming they did. It also explains what would otherwise be
/// inexplicable: a session working on something nobody here asked for.
PeerMessage {
/// The sending session's own name, which is what the reader
/// recognises it by -- the socket path it came from is not.
from: String,
text: String,
/// The seq of the `Status::Running` that opened the turn this message
/// started, so a reader can draw it above that turn.
///
/// The CLI says nothing about a peer message until the turn's
/// `result`, so the event is appended after everything it caused, and
/// an append-only transcript cannot go back and insert it. Carrying
/// the position instead keeps one order on the wire and one on screen.
///
/// Filled in by the pump, the only place that knows a seq, and only
/// where a turn was open: `None` for a message replayed by `import`,
/// which already has it in the right place.
#[serde(default, skip_serializing_if = "Option::is_none")]
turn_start: Option<u64>,
},
/// **Retired on 2026-09-06, hours after it was added.** Kept only so the
/// transcripts written while it existed still read: a session that ran a
/// background task in that window has these lines for ever, and a
/// transcript is append-only, so there is no pass that could remove them.
///
/// Never constructed. It reported a background task finishing, and drawing
/// a row per one turned out to be a screen of notices about work the
/// reader was not asking after -- see PLAN.md's "Two turns must never be
/// drawn as one". The phone folds it to no row at all, which is what makes
/// keeping it cheap.
///
/// Deleting the variant instead is what broke every live session, and
/// [`Event::Unreadable`] is the reason that cannot happen again. This is
/// still here rather than left to that: an unreadable line draws a
/// placeholder, correctly, and one per background task is the same wall
/// the row was removed for.
TaskNote {
about: String,
#[serde(default, skip_serializing_if = "Option::is_none")]
title: Option<String>,
status: String,
#[serde(default, skip_serializing_if = "Option::is_none")]
summary: Option<String>,
},
/// A line in a transcript that this build cannot read: a kind a newer
/// server wrote, a kind an older one wrote that has since been dropped, or
/// a line whose contents do not fit the kind it names.
///
/// **Only ever made when reading, never sent by a driver**, and it is the
/// reason a transcript can outlive a change to this enum. See
/// `Indexed::parse_at` for the incident: removing a variant after
/// transcripts had recorded it made every read of those files fail, so
/// every session in them lost its status, its history and its ability to
/// be sent to.
///
/// It carries the word the line called itself so a reader is told what
/// they are missing rather than that something is missing. `kind` is not
/// an enum for the obvious reason: the whole point of this variant is the
/// words that are not in one.
Unreadable {
kind: String,
},
/// The manager's record of a question being answered, so a rendered
/// question card resolves on every device rather than only the one that
/// answered.
///
/// A list because a question can take several answers, and one that took
/// one is the list of length one rather than a different shape.
Answered {
id: String,
answers: Vec<String>,
},
/// How many background tasks the provider says are alive now.
///
/// State rather than a transcript row: the phone draws it beside the
/// session status. A distinct event keeps the count current while a
/// session is open; `GET /sessions` supplies the opening snapshot.
BackgroundTasks {
count: usize,
},
Status {
state: SessionStatus,
},
/// What the session is set to, as the session itself reports it.
///
/// Asking for a change and having one are different things, and only this
/// is a measurement: a model name the dialect does not know, a mode it
/// refuses, or a driver whose model is fixed at startup all leave a
/// request that was sent and nothing that changed. Reporting from the
/// request put the answer on the phone before the question was answered.
///
/// Either field alone, because the two are confirmed separately.
Settings {
#[serde(default, skip_serializing_if = "Option::is_none")]
model: Option<String>,
#[serde(default, skip_serializing_if = "Option::is_none")]
permission_mode: Option<String>,
},
/// How much context this session's model has to hold a conversation in.
///
/// The denominator the phone draws [`Event::UsageDelta`]'s `context`
/// against, and its own event rather than a field on that one because it
/// is not a per-turn measurement: it is fixed when the process starts and
/// changes only when a different one is started, which is what a model
/// change does. Reported the moment it is known, so the figure and what it
/// is out of arrive together rather than the first turn drawing a
/// numerator with no denominator.
///
/// **Only ever sent by a driver that actually knows.** A window nobody has
/// measured is not an unlimited one: llama.cpp answers it exactly, because
/// the number is a flag the server was started with and `/props` reads it
/// back, while a coding CLI's context is the vendor's business and
/// nothing in either control protocol states it. Those send nothing, the
/// session has no limit, and the phone draws the figure on its own -- see
/// `SessionSummary::context_limit`.
ContextWindow {
tokens: u64,
},
/// Per-turn token counts, where the dialect reports them.
UsageDelta {
/// What this turn cost: the tokens it was charged for.
tokens: u64,
/// What the model was holding when the turn ended -- see
/// [`context_tokens`].
///
/// Carried rather than summed by whoever is reading, because it is
/// not a sum: context goes *down* at a compaction and a clear, so
/// adding turns up would report a figure the session stopped being
/// true of long ago.
///
/// `None` where the dialect did not say, which every reader has to be
/// able to draw.
#[serde(default, skip_serializing_if = "Option::is_none")]
context: Option<u64>,
},
/// A compaction that finished, and how much context it recovered.
///
/// The counts are the point, and a spinner is not. They are optional
/// because the record has shipped without them, and "the compaction
/// happened, we don't know by how much" is a state this has to be able to
/// say -- a plausible number would be indistinguishable from a counted one.
Compacted {
#[serde(default, skip_serializing_if = "Option::is_none")]
pre_tokens: Option<u64>,
#[serde(default, skip_serializing_if = "Option::is_none")]
post_tokens: Option<u64>,
/// What asked for it, in the dialect's own word -- `auto` when the
/// session compacted on its own. Carried rather than reduced to a bool
/// so an unrecognised trigger stays unrecognised: an automatic
/// compaction is the one worth naming, because it explains a wait
/// nobody asked for.
#[serde(default, skip_serializing_if = "Option::is_none")]
trigger: Option<String>,
},
/// A command the session was asked to run on itself, held because it
/// cannot run yet. These are not messages: `/compact` and `/rename` are
/// instructions about the session, and a session mid-turn reads a line
/// written to it as something the model should see. So they wait, and
/// this is what a phone draws while they do.
CommandQueued {
id: String,
text: String,
},
/// The same command, now handed to the session. Its [`CommandQueued`]
/// stops being pending when this arrives, matched by `id`; a command
/// that ran immediately has only this.
CommandSent {
id: String,
text: String,
},
/// The conversation was cleared: everything above this is still in the
/// record but is no longer in the session's context.
///
/// Nothing is deleted. A transcript is the thing a person scrolls back
/// through, so this is a divider, not a truncation.
///
/// **Load-bearing, not decorative.** For any driver that rebuilds its
/// conversation from the transcript, this marker decides what the model
/// is given -- dropping it, or treating it as something only the phone
/// draws, silently puts a cleared conversation back in front of the model
/// at full cost. Today `llama::conversation` is the only fold that reads
/// it, which is why this is written down rather than left to be inferred
/// from a second example that does not exist.
Cleared,
/// The account behind this session has no quota left, so the turn stopped
/// without finishing.
///
/// Its own event rather than an [`Event::Error`] carrying the dialect's
/// sentence, because two things act on it that cannot read English: the
/// transcript draws it as a state the session is in rather than as a
/// failure of something it did, and `crate::resume` schedules the message
/// that picks the work back up. Recognising it belongs to the driver, which
/// is the only layer that knows its dialect's wording -- above here nothing
/// matches on strings.
///
/// `resets_at` is epoch seconds, and `None` is a real state: the dialect
/// said the limit was hit without saying when it lifts. Nothing here
/// invents one -- what the wait is actually decided against is the usage
/// endpoint, and this is the hint that starts the waiting.
LimitReached {
#[serde(default, skip_serializing_if = "Option::is_none")]
resets_at: Option<f64>,
},
/// The provider refused the turn because its login can no longer be used.
///
/// Recognised by the driver for the same reason [`Event::LimitReached`] is:
/// only that layer knows the provider's dialect, and the phone needs a
/// state it can act on without matching error prose.
AuthenticationRequired {
message: String,
},
Error {
message: String,
},
}
/// The common presentation of a file change, whichever driver produced it.
pub(super) fn patch_start(id: String, diff: String) -> Event {
Event::ToolStart {
id,
tool: "Patch".to_string(),
input: serde_json::json!({"diff": diff}),
}
}
pub(super) fn prefixed_lines(prefix: char, text: &str) -> String {
text.split_inclusive('\n')
.map(|line| {
if line.ends_with('\n') {
format!("{prefix}{line}")
} else {
format!("{prefix}{line}\n")
}
})
.collect()
}
/// How much the model was holding, from the three figures a turn reports:
/// the input side only, prompt plus both cache figures. A cached token is
/// cheaper but it is still one the model was given; output is what the turn
/// produced rather than what continuing has to carry.
///
/// One function so the definition cannot drift, because it is extracted two
/// quite different ways -- the live translators have the usage object parsed,
/// and `import::context_tokens` scans it out of a raw line without parsing.
pub fn context_tokens(input: u64, cache_creation: u64, cache_read: u64) -> u64 {
input + cache_creation + cache_read
}
/// The context window after `event`, given what it was before.
///
/// Beside [`context_after`] because it has the same three readers and the same
/// hazard: a figure that outlives what made it true. A model change replaces
/// the process, so the window it reports replaces the old one -- and until the
/// new one says, there is no answer rather than the previous model's.
pub fn context_limit_after(current: Option<u64>, event: &Event) -> Option<u64> {
match event {
Event::ContextWindow { tokens } => Some(*tokens),
// The window belongs to the process, and a stopped one has none. Left
// standing, a restarted session on a different model would draw its
// occupancy against the previous model's window.
Event::Status {
state: SessionStatus::Exited,
} => None,
_ => current,
}
}
/// The context after `event`, given what it was before.
///
/// The whole rule in one place, because three readers need the same answer:
/// the pump keeping a live session's figure, the transcript seeding it at
/// startup, and the phone folding the same events into what it draws.
///
/// The two that *lower* it are the point. A clear takes the conversation away
/// and a compaction replaces it with a summary, so a figure measured before
/// either stopped being true at that moment -- and carrying it forward is how
/// a session that had just been cleared went on reporting the context it no
/// longer had.
///
/// `None` is "we don't know", which each of them can reach.
pub fn context_after(current: Option<u64>, event: &Event) -> Option<u64> {
match event {
// `or`, so a turn the dialect reported no usage for leaves the last
// measurement standing: stale by a turn, which every context figure
// is, rather than wrong.
Event::UsageDelta { context, .. } => context.or(current),
Event::Compacted { post_tokens, .. } => *post_tokens,
Event::Cleared => None,
_ => current,
}
}
/// Something a session can be asked to do to itself.
///
/// A closed set rather than a string, because the two that are not
/// dialect-specific have to reach every provider: compaction is a capability
/// an llama session may one day have, and a name is this server's own. `Raw`
/// is the escape for a dialect's own commands, which only the thing running
/// the session can interpret.
#[derive(Debug, Clone, PartialEq)]
pub enum SessionCommand {
Compact,
Clear,
SetTitle(String),
Raw(String),
}
impl SessionCommand {
/// What a person would have typed to ask for this, which is what a phone
/// shows while it waits.
pub fn label(&self) -> String {
match self {
Self::Compact => "/compact".to_string(),
Self::Clear => "/clear".to_string(),
Self::SetTitle(title) => format!("/rename {title}"),
Self::Raw(text) => text.clone(),
}
}
/// Runs it. Called only at a boundary -- see [`Event::CommandQueued`].
pub fn apply(&self, driver: &dyn Driver) {
match self {
Self::Compact => driver.compact(),
Self::Clear => driver.clear(),
Self::SetTitle(title) => driver.set_title(title),
Self::Raw(text) => driver.run_command(text),
}
}
}
#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
#[serde(rename_all = "camelCase")]
pub enum SessionStatus {
Idle,
Running,
AwaitingInput,
Compacting,
/// The session's process is up but cannot be spoken to yet.
///
/// Its own state because the two it would otherwise borrow are both
/// wrong in ways somebody notices. `Running` means the session is
/// answering, so a model taking a minute to load looks like a model
/// thinking for a minute -- and there is no way to tell from the screen
/// that the first message will be refused. `Idle` invites that message
/// and then loses it.
///
/// It exists for `llama-server`, which reads a multi-gigabyte file off
/// disk before it answers anything, and it is general because the
/// condition is: a process that is started and not yet ready is a state
/// any driver may have to report. Nothing is queued *because* of this
/// state -- a driver that reports it is responsible for holding what it
/// is sent until it can deliver it -- but this is what says so on screen.
Loading,
/// The session's own turn is over, but work it started is still going:
/// a backgrounded subagent, or a command left running.
///
/// Its own state rather than `Idle` because the two differ in kind and
/// only one of them is an invitation. `Idle` means the session is
/// waiting for a person; this means it is waiting for itself, and a
/// notification saying the work had finished would have been wrong. It
/// is also not `Running`: nothing is being written to the transcript,
/// the reply that ended the turn is finished, and a spinner on a session
/// that will not speak again until a task reports back is a promise
/// nobody can keep.
Waiting,
Exited,
/// There is a process recorded for this session and the machine will not
/// say whether it is still running.
///
/// Its own state rather than the nearest of the others, because both
/// neighbours are lies with consequences: `Exited` invites starting a
/// second process against a conversation that may already have one, and
/// `Idle` claims a session is waiting for you when nobody has checked.
Unknown,
}
/// What became of a request to take a queued message back.
///
/// Three states rather than a bool because the two failures are not the same
/// fact. A driver that writes into its session the moment a message arrives
/// -- which is what `ClaudeDriver` does, so a steer reaches the model at the
/// next tool boundary -- can never take one back, and a phone told only "no"
/// would have to guess whether it asked too late or asked about nothing.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum Unqueued {
/// Out of the queue; the session will never read it.
Dropped,
/// Already handed to the session, so there is nothing left to take back.
AlreadySent,
/// Nothing is waiting under that id.
Unknown,
}
/// Where a driver reports events. Unbounded because producers are child
/// processes a slow phone must never be able to stall; the transcript file
/// is the backpressure-free buffer of record.
pub type EventSink = mpsc::UnboundedSender<Event>;
/// The inbound half of a session. Deliberately small; see PLAN.md for the
/// per-driver mapping of each method onto its dialect.
///
/// `send_user_message` during a run is the point of the whole app: a dialect
/// with live input injects it at the next tool boundary, while a turn-at-a-time
/// dialect queues it for the next child process.
pub trait Driver: Send + Sync {
/// The provider's latest measured number of live background tasks.
/// `None` means it has not reported one, not that the count is zero.
fn background_tasks(&self) -> Option<usize> {
None
}
/// Takes a message, now or once the session is free for it.
///
/// Every driver owes exactly one `MessageTaken` per message, at the moment
/// it actually starts reading it: that event is what puts the message in
/// the transcript, so a driver that never sends it drops the message from
/// the conversation entirely.
fn send_user_message(&self, text: String, attachments: Vec<AttachmentRef>);
/// Takes back a message that is still waiting, named by the id its
/// [`Event::MessageQueued`] carried.
///
/// Answering is the whole of the contract: a driver that drops the message
/// owes an [`Event::MessageDropped`], and one that cannot must say which
/// of the two reasons it is -- "the session has already been told" is
/// worth knowing, and "there is nothing under that id" means the bubble on
/// screen is stale. The default is the honest answer for a driver with no
/// queue at all.
fn unqueue(&self, _id: &str) -> Unqueued {
Unqueued::Unknown
}
/// Answers one question with everything that was chosen, in the order it
/// was offered. A driver whose dialect takes a single value joins them
/// where it writes it.
fn answer_question(&self, id: &str, answers: &[String]);
/// Stop mid-run; the session survives.
fn interrupt(&self);
fn set_model(&self, model: &str);
/// How much the session asks about before acting. Live rather than
/// spawn-only: the answer changes with what is being done, and a phone is
/// the worst place to answer "may I run this?" forty times.
fn set_permission_mode(&self, mode: &str);
// Both of the above are requests, and neither reports the outcome by
// returning. A driver that changes the setting owes an [`Event::Settings`]
// once it has -- that event, not the request, is what the manager and the
// phone read. One that cannot owes an [`Event::Error`] saying why.
/// Tells the process what this conversation is called, when it has
/// somewhere to put it.
///
/// Unlike the two above, this is not a request that can fail: the rename
/// has already happened in this server's config, which is what a phone
/// lists. So a driver whose process has no notion of a name does nothing
/// and says nothing. Claude Code has one: `--name` at creation and
/// `/rename` afterwards, which is what puts the same name in its own
/// session picker and in what other agents see.
fn set_title(&self, title: &str);
/// Takes the session's provider settings, whole.
///
/// The whole map because it is a form's contents -- an absent key means
/// "unset", not "unchanged". A driver applies what it can apply now and
/// says so about the rest: the map is also on disk by the time this is
/// called, so a setting that only takes effect at the next start is not
/// lost, it is waiting. The default is right for a driver with no settings
/// of its own, which is every one but llama.cpp -- see
/// [`crate::config::DriverKind::params`].
fn set_params(&self, _params: &std::collections::BTreeMap<String, String>) {}
/// Runs a command this session's own dialect understands, verbatim --
/// `/context`, `/usage`, anything a CLI adds next month. A driver with no
/// such vocabulary says so with an [`Event::Error`] rather than sending it
/// as a message, which would put a line meant for the session in front of
/// the model.
///
/// Called only when the session is between turns; the waiting is done
/// above, once, for every driver.
fn run_command(&self, text: &str);
/// llama: not built, and refused; claude: `/compact`.
fn compact(&self);
/// Drops the conversation so far without ending the session.
///
/// The cheap half of managing a long session, and why it is a driver
/// operation rather than a manager one: compaction *reads* the whole
/// conversation in order to summarise it, so on a large context it is
/// itself one of the most expensive requests the session will make --
/// measured at 1.7 million tokens for one automatic compaction on
/// 2026-08-29. Clearing costs nothing, because nothing is sent.
///
/// Every implementation emits [`Event::Cleared`] so the transcript carries
/// the divider whatever the dialect did behind it.
fn clear(&self);
/// Stop attending to the process but leave it running, because this
/// server is going away and means to adopt it again.
///
/// Deliberately not a shutdown: a backend restart must not end a turn that
/// is in flight, so a session's process outlives the server that started
/// it and is found again through `session::process`.
///
/// Its counterpart is [`Driver::stop`]. Every driver owes exactly one of
/// the two on the way out, and which one is the difference between "back
/// shortly" and "this conversation is over".
/// Whether a line written *now* would start a turn of its own, rather than
/// landing inside one already in flight.
///
/// Asked of the driver because the driver is the only thing that knows: it
/// updates this the instant it writes rather than when output returns. The
/// manager's `SessionStatus` is built from what has been *recorded*, so
/// between writing a line and the CLI's first output it still reads idle,
/// and a second line sent in that gap lands inside the turn the first one
/// started. For a command that is the difference between being executed
/// and being read to the model as text, which is silent both ways.
///
/// Defaults to true for a driver with no turn of its own to be inside.
fn between_turns(&self) -> bool {
true
}
fn detach(&self);
/// End the process for good, because it must not survive this. The path
/// out for everything [`Driver::detach`] preserves.
///
/// Two callers, differing only in what is being ended: a session being
/// deleted, whose conversation goes with it, and a throwaway session at a
/// server's exit, whose transcript stays and whose process does not.
fn stop(&self);
}
#[cfg(test)]
mod tests {
use super::*;
/// A steer says so to the model and nowhere else: the note goes only into
/// what the CLI is handed, while the transcript keeps what was typed.
#[test]
fn only_a_steer_carries_the_note() {
let files = [std::path::PathBuf::from("/tmp/x/trace.txt")];
let plain = message_body("do the thing", &files, false);
assert_eq!(plain, "do the thing\n\nAttached file: /tmp/x/trace.txt");
let steer = message_body("do the thing", &files, true);
assert_eq!(steer, format!("{STEERING_NOTE}\n\n{plain}"));
// Attachments with nothing typed still need the note, and must not
// arrive with a blank line where the words would have been.
assert_eq!(
message_body("", &files, true),
format!("{STEERING_NOTE}\n\nAttached file: /tmp/x/trace.txt")
);
}
/// A tripwire for the wire format, not for serde. The app reads these
/// names, and getting one wrong does not fail loudly: a field the app
/// cannot find reads as a field the server chose not to send, which
/// several of them are allowed to be.
#[test]
fn multi_word_fields_go_out_in_camel_case() {
let json = serde_json::to_value(Event::Compacted {
pre_tokens: Some(28719),
post_tokens: Some(1125),
trigger: Some("manual".to_string()),
})
.expect("serialize");
assert_eq!(
json,
serde_json::json!({
"type": "compacted",
"preTokens": 28719,
"postTokens": 1125,
"trigger": "manual",
})
);
}
/// The two events that take the context *down* are the point of the fold:
/// a figure measured before a compaction or a clear stopped being true at
/// that moment, and carrying it forward is how a session that had just
/// been cleared went on reporting the context it no longer had.
#[test]
fn a_compaction_and_a_clear_move_the_context_a_turn_cannot() {
let after = |current, event| context_after(current, &event);
assert_eq!(
after(
Some(500),
Event::UsageDelta {
tokens: 12,
context: Some(30_100),
}
),
Some(30_100)
);
assert_eq!(
after(
Some(128_402),
Event::Compacted {
pre_tokens: Some(128_402),
post_tokens: Some(9_617),
trigger: Some("auto".to_string()),
}
),
Some(9_617)
);
assert_eq!(after(Some(9_617), Event::Cleared), None);
// A compaction that did not say how much it recovered leaves the
// context unknown rather than stale: it definitely moved.
assert_eq!(
after(
Some(128_402),
Event::Compacted {
pre_tokens: None,
post_tokens: None,
trigger: None,
}
),
None
);
// A turn the dialect reported no context for is stale by a turn,
// which every context figure is, rather than unknown.
assert_eq!(
after(
Some(30_100),
Event::UsageDelta {
tokens: 12,
context: None,
}
),
Some(30_100)
);
// Everything else leaves it alone.
assert_eq!(
after(
Some(30_100),
Event::Status {
state: SessionStatus::Idle,
}
),
Some(30_100)
);
}
}