mirror of
https://github.com/instructkr/claw-code.git
synced 2026-11-16 00:00:40 +08:00
Merge pull request #3318 from ultraworkers/pr-3292
fix: harden the agent loop against a self-hosted Anthropic-compatible backend
This commit is contained in:
18 files changed
+1940
-464
No files matched your search
@@ -95,10 +95,21 @@ pub fn get_compact_continuation_message(
|
||||
#[must_use]
|
||||
pub fn compact_session(session: &Session, config: CompactionConfig) -> CompactionResult {
|
||||
if !should_compact(session, config) {
|
||||
// "Nothing to drop" is not the same as "nothing to do". A session can exceed the window
|
||||
// while holding fewer messages than `preserve_recent_messages` — a single read of a large
|
||||
// file does it — and message-granular compaction has no move to make there. Trim the
|
||||
// payloads anyway, so the one lever that still works is not gated behind the one that
|
||||
// does not.
|
||||
let mut trimmed_session = session.clone();
|
||||
trimmed_session.messages = session
|
||||
.messages
|
||||
.iter()
|
||||
.map(truncate_oversized_tool_results)
|
||||
.collect();
|
||||
return CompactionResult {
|
||||
summary: String::new(),
|
||||
formatted_summary: String::new(),
|
||||
compacted_session: session.clone(),
|
||||
compacted_session: trimmed_session,
|
||||
removed_message_count: 0,
|
||||
};
|
||||
}
|
||||
@@ -165,7 +176,14 @@ pub fn compact_session(session: &Session, config: CompactionConfig) -> Compactio
|
||||
k
|
||||
};
|
||||
let removed = &session.messages[compacted_prefix_len..keep_from];
|
||||
let preserved = session.messages[keep_from..].to_vec();
|
||||
// Dropping whole messages is not enough on its own. An agentic turn is five messages
|
||||
// (user / assistant+tool_use / tool_result / assistant+tool_use / tool_result), so the
|
||||
// preserved tail IS the pair of tool results that overflowed the window, and the only
|
||||
// message compaction may drop is the short user prompt. Trim the payloads that must stay.
|
||||
let preserved = session.messages[keep_from..]
|
||||
.iter()
|
||||
.map(truncate_oversized_tool_results)
|
||||
.collect::<Vec<_>>();
|
||||
let summary =
|
||||
merge_compact_summaries(existing_summary.as_deref(), &summarize_messages(removed));
|
||||
let formatted_summary = format_compact_summary(&summary);
|
||||
@@ -190,6 +208,78 @@ pub fn compact_session(session: &Session, config: CompactionConfig) -> Compactio
|
||||
}
|
||||
}
|
||||
|
||||
/// Largest tool-result payload compaction will preserve verbatim, in characters.
|
||||
///
|
||||
/// ~16 KB is ~4 k tokens: big enough that ordinary tool output is never touched, small enough
|
||||
/// that several preserved results still fit a small window.
|
||||
const MAX_PRESERVED_TOOL_RESULT_CHARS: usize = 16 * 1024;
|
||||
|
||||
/// Characters kept from the tail of a trimmed payload. The head carries the structure a model
|
||||
/// needs to recognise what it read; the tail is where a file's most recent edits usually are.
|
||||
const PRESERVED_TOOL_RESULT_TAIL_CHARS: usize = 2 * 1024;
|
||||
|
||||
/// Returns `message` with any oversized `ToolResult` payload replaced by a head/tail excerpt.
|
||||
///
|
||||
/// Clones only when something is actually trimmed, so the common path stays cheap.
|
||||
fn truncate_oversized_tool_results(message: &ConversationMessage) -> ConversationMessage {
|
||||
let needs_trim = message.blocks.iter().any(|block| match block {
|
||||
ContentBlock::ToolResult { output, .. } => {
|
||||
output.chars().count() > MAX_PRESERVED_TOOL_RESULT_CHARS
|
||||
}
|
||||
_ => false,
|
||||
});
|
||||
if !needs_trim {
|
||||
return message.clone();
|
||||
}
|
||||
|
||||
let blocks = message
|
||||
.blocks
|
||||
.iter()
|
||||
.map(|block| match block {
|
||||
ContentBlock::ToolResult {
|
||||
tool_use_id,
|
||||
tool_name,
|
||||
output,
|
||||
is_error,
|
||||
} => ContentBlock::ToolResult {
|
||||
tool_use_id: tool_use_id.clone(),
|
||||
tool_name: tool_name.clone(),
|
||||
output: excerpt_tool_result(output),
|
||||
is_error: *is_error,
|
||||
},
|
||||
other => other.clone(),
|
||||
})
|
||||
.collect();
|
||||
|
||||
ConversationMessage {
|
||||
role: message.role,
|
||||
blocks,
|
||||
usage: message.usage,
|
||||
}
|
||||
}
|
||||
|
||||
/// Builds a head + marker + tail excerpt of an oversized tool-result payload.
|
||||
///
|
||||
/// Operates on chars, not bytes, so a multi-byte boundary can never be split.
|
||||
fn excerpt_tool_result(output: &str) -> String {
|
||||
let chars: Vec<char> = output.chars().collect();
|
||||
let total = chars.len();
|
||||
if total <= MAX_PRESERVED_TOOL_RESULT_CHARS {
|
||||
return output.to_string();
|
||||
}
|
||||
|
||||
let tail_len = PRESERVED_TOOL_RESULT_TAIL_CHARS.min(total);
|
||||
let head_len = MAX_PRESERVED_TOOL_RESULT_CHARS.saturating_sub(tail_len);
|
||||
let elided = total.saturating_sub(head_len + tail_len);
|
||||
if elided == 0 {
|
||||
return output.to_string();
|
||||
}
|
||||
|
||||
let head: String = chars[..head_len].iter().collect();
|
||||
let tail: String = chars[total - tail_len..].iter().collect();
|
||||
format!("{head}\n\n[... {elided} characters elided by compaction; re-read this file if you need the omitted portion ...]\n\n{tail}")
|
||||
}
|
||||
|
||||
fn compacted_summary_prefix_len(session: &Session) -> usize {
|
||||
usize::from(
|
||||
session
|
||||
@@ -573,7 +663,7 @@ fn extract_summary_timeline(summary: &str) -> Vec<String> {
|
||||
#[cfg(test)]
|
||||
mod tests {
|
||||
use super::{
|
||||
collect_key_files, compact_session, format_compact_summary,
|
||||
collect_key_files, compact_session, estimate_session_tokens, format_compact_summary,
|
||||
get_compact_continuation_message, infer_pending_work, should_compact, CompactionConfig,
|
||||
};
|
||||
use crate::session::{ContentBlock, ConversationMessage, MessageRole, Session};
|
||||
@@ -763,6 +853,185 @@ mod tests {
|
||||
/// user(ToolResult) pair at the boundary. An orphaned tool-result message
|
||||
/// without the preceding assistant `tool_calls` causes a 400 on the
|
||||
/// OpenAI-compat path (gaebal-gajae repro 2026-04-09).
|
||||
/// The failure this exists for: an agentic turn is 5 messages (user / assistant+tool_use /
|
||||
/// tool_result / assistant+tool_use / tool_result), so `preserve_recent_messages: 4` protects
|
||||
/// exactly the two giant tool results that overflow the window. Message-granular compaction
|
||||
/// can only drop the tiny user prompt, so the payloads themselves must be trimmed.
|
||||
#[test]
|
||||
fn compaction_truncates_oversized_tool_results_it_must_preserve() {
|
||||
let huge = "z".repeat(80_000);
|
||||
let mut session = Session::new();
|
||||
session
|
||||
.push_message(ConversationMessage::user_text("read two files"))
|
||||
.expect("user message");
|
||||
session
|
||||
.push_message(ConversationMessage::assistant(vec![
|
||||
ContentBlock::ToolUse {
|
||||
id: "call-1".to_string(),
|
||||
name: "read_file".to_string(),
|
||||
input: "{\"path\":\"a.rs\"}".to_string(),
|
||||
},
|
||||
]))
|
||||
.expect("assistant tool use");
|
||||
session
|
||||
.push_message(ConversationMessage::tool_result(
|
||||
"call-1",
|
||||
"read_file",
|
||||
huge.clone(),
|
||||
false,
|
||||
))
|
||||
.expect("tool result");
|
||||
session
|
||||
.push_message(ConversationMessage::assistant(vec![
|
||||
ContentBlock::ToolUse {
|
||||
id: "call-2".to_string(),
|
||||
name: "read_file".to_string(),
|
||||
input: "{\"path\":\"b.rs\"}".to_string(),
|
||||
},
|
||||
]))
|
||||
.expect("second assistant tool use");
|
||||
session
|
||||
.push_message(ConversationMessage::tool_result(
|
||||
"call-2",
|
||||
"read_file",
|
||||
huge.clone(),
|
||||
false,
|
||||
))
|
||||
.expect("second tool result");
|
||||
|
||||
let before = estimate_session_tokens(&session);
|
||||
let result = compact_session(
|
||||
&session,
|
||||
CompactionConfig {
|
||||
preserve_recent_messages: 4,
|
||||
max_estimated_tokens: 0,
|
||||
},
|
||||
);
|
||||
let after = estimate_session_tokens(&result.compacted_session);
|
||||
|
||||
assert!(
|
||||
after < before / 2,
|
||||
"compaction must shrink an oversized preserved tool result: {before} -> {after}"
|
||||
);
|
||||
|
||||
let preserved_output = result
|
||||
.compacted_session
|
||||
.messages
|
||||
.iter()
|
||||
.flat_map(|message| &message.blocks)
|
||||
.find_map(|block| match block {
|
||||
ContentBlock::ToolResult { output, .. } => Some(output.clone()),
|
||||
_ => None,
|
||||
})
|
||||
.expect("the tool result must still be present, not deleted");
|
||||
|
||||
assert!(
|
||||
preserved_output.len() < huge.len(),
|
||||
"the payload must actually be trimmed"
|
||||
);
|
||||
assert!(
|
||||
preserved_output.contains("elided"),
|
||||
"the model must be told the view is partial, got: {}",
|
||||
&preserved_output[..preserved_output.len().min(200)]
|
||||
);
|
||||
assert!(
|
||||
preserved_output.starts_with('z'),
|
||||
"the head of the output must be kept"
|
||||
);
|
||||
assert!(
|
||||
preserved_output.ends_with('z'),
|
||||
"the tail of the output must be kept"
|
||||
);
|
||||
}
|
||||
|
||||
/// A session can exceed the window while holding FEWER messages than
|
||||
/// `preserve_recent_messages` — one read of a large file does it. `should_compact` returns
|
||||
/// false there (nothing is droppable), so trimming must not be gated behind it.
|
||||
#[test]
|
||||
fn oversized_tool_results_are_trimmed_even_when_no_message_can_be_dropped() {
|
||||
let huge = "q".repeat(80_000);
|
||||
let mut session = Session::new();
|
||||
session
|
||||
.push_message(ConversationMessage::user_text("read one enormous file"))
|
||||
.expect("user message");
|
||||
session
|
||||
.push_message(ConversationMessage::assistant(vec![
|
||||
ContentBlock::ToolUse {
|
||||
id: "call-1".to_string(),
|
||||
name: "read_file".to_string(),
|
||||
input: "{\"path\":\"huge.rs\"}".to_string(),
|
||||
},
|
||||
]))
|
||||
.expect("assistant tool use");
|
||||
session
|
||||
.push_message(ConversationMessage::tool_result(
|
||||
"call-1",
|
||||
"read_file",
|
||||
huge,
|
||||
false,
|
||||
))
|
||||
.expect("tool result");
|
||||
|
||||
// 3 messages, preserve 4: there is nothing to remove, yet the session is far too big.
|
||||
let before = estimate_session_tokens(&session);
|
||||
let result = compact_session(
|
||||
&session,
|
||||
CompactionConfig {
|
||||
preserve_recent_messages: 4,
|
||||
max_estimated_tokens: 0,
|
||||
},
|
||||
);
|
||||
let after = estimate_session_tokens(&result.compacted_session);
|
||||
|
||||
assert!(
|
||||
after < before / 2,
|
||||
"an untrimmable session must still have its payloads trimmed: {before} -> {after}"
|
||||
);
|
||||
}
|
||||
|
||||
/// Trimming must not touch payloads that already fit — otherwise every ordinary tool result
|
||||
/// would grow an elision marker it does not need.
|
||||
#[test]
|
||||
fn compaction_leaves_small_tool_results_untouched() {
|
||||
let small = "ok".repeat(50);
|
||||
let mut session = Session::new();
|
||||
for index in 0..6 {
|
||||
session
|
||||
.push_message(ConversationMessage::user_text(format!("turn {index}")))
|
||||
.expect("user message");
|
||||
session
|
||||
.push_message(ConversationMessage::tool_result(
|
||||
format!("call-{index}"),
|
||||
"read_file",
|
||||
small.clone(),
|
||||
false,
|
||||
))
|
||||
.expect("tool result");
|
||||
}
|
||||
|
||||
let result = compact_session(
|
||||
&session,
|
||||
CompactionConfig {
|
||||
preserve_recent_messages: 4,
|
||||
max_estimated_tokens: 0,
|
||||
},
|
||||
);
|
||||
|
||||
for block in result
|
||||
.compacted_session
|
||||
.messages
|
||||
.iter()
|
||||
.flat_map(|message| &message.blocks)
|
||||
{
|
||||
if let ContentBlock::ToolResult { output, .. } = block {
|
||||
assert_eq!(
|
||||
output, &small,
|
||||
"a small tool result must pass through as-is"
|
||||
);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn compaction_does_not_split_tool_use_tool_result_pair() {
|
||||
use crate::session::{ContentBlock, Session};
|
||||
|
||||
@@ -18,6 +18,47 @@ use crate::usage::{TokenUsage, UsageTracker};
|
||||
const DEFAULT_AUTO_COMPACTION_INPUT_TOKENS_THRESHOLD: u32 = 100_000;
|
||||
const AUTO_COMPACTION_THRESHOLD_ENV_VAR: &str = "CLAUDE_CODE_AUTO_COMPACT_INPUT_TOKENS";
|
||||
|
||||
const DEFAULT_MAX_STALL_NUDGES: usize = 3;
|
||||
const MAX_STALL_NUDGES_ENV_VAR: &str = "CLAW_MAX_STALL_NUDGES";
|
||||
/// Placeholder recorded in the trace when a turn is resumed rather than started, so the trace
|
||||
/// does not attribute a second user message to a turn that only had one.
|
||||
const RESUMED_TURN_MARKER: &str = "[resumed turn]";
|
||||
/// Operator-declared context window. Kept as a literal rather than imported because `runtime`
|
||||
/// sits below `api` in the dependency graph; `api::context_window_override` reads the same name.
|
||||
const CONTEXT_WINDOW_ENV_VAR: &str = "CLAW_CONTEXT_WINDOW";
|
||||
/// Share of a declared context window at which compaction is triggered, leaving the remainder
|
||||
/// as headroom for the reply plus whatever the next tool result carries.
|
||||
const AUTO_COMPACTION_WINDOW_PERCENT: u32 = 70;
|
||||
|
||||
/// Injected when the model narrates its next step (or asks for a go-ahead mid-task)
|
||||
/// instead of emitting the tool call that would perform it.
|
||||
const STALL_NUDGE_PROMPT: &str = "[auto-continue] You ended the turn without calling a tool, \
|
||||
but the work is not finished. Do not describe the next step and do not ask for \
|
||||
confirmation — issue the tool call now and keep going until the task is done. If \
|
||||
everything asked for is genuinely complete, reply with a one-line summary containing \
|
||||
no forward-looking sentence.";
|
||||
|
||||
/// Phrases that mark text as an announcement of work the model has not performed yet.
|
||||
/// Matched against the final non-empty line of the assistant's reply.
|
||||
const PENDING_ACTION_MARKERS: [&str; 16] = [
|
||||
"let me ",
|
||||
"let's ",
|
||||
"lets ",
|
||||
"i'll ",
|
||||
"i will ",
|
||||
"i'm going to",
|
||||
"i am going to",
|
||||
"going to run",
|
||||
"about to ",
|
||||
"now running",
|
||||
"now checking",
|
||||
"now i",
|
||||
"next i",
|
||||
"next, i",
|
||||
"proceeding to",
|
||||
"moving on to",
|
||||
];
|
||||
|
||||
/// Fully assembled request payload sent to the upstream model client.
|
||||
#[derive(Debug, Clone, PartialEq, Eq)]
|
||||
pub struct ApiRequest {
|
||||
@@ -90,6 +131,7 @@ impl std::error::Error for ToolError {}
|
||||
#[derive(Debug, Clone, PartialEq, Eq)]
|
||||
pub struct RuntimeError {
|
||||
message: String,
|
||||
context_window: bool,
|
||||
}
|
||||
|
||||
impl RuntimeError {
|
||||
@@ -97,8 +139,28 @@ impl RuntimeError {
|
||||
pub fn new(message: impl Into<String>) -> Self {
|
||||
Self {
|
||||
message: message.into(),
|
||||
context_window: false,
|
||||
}
|
||||
}
|
||||
|
||||
/// A failure the caller may be able to clear by shrinking the session.
|
||||
///
|
||||
/// This is a decision the API layer makes and hands up, not one the caller reconstructs by
|
||||
/// matching substrings against the rendered message. Recovering from context exhaustion means
|
||||
/// discarding conversation history, so a transport hiccup misread as an overflow destroys the
|
||||
/// session for no reason — the classification has to come from whoever saw the real error.
|
||||
#[must_use]
|
||||
pub fn context_window(message: impl Into<String>) -> Self {
|
||||
Self {
|
||||
message: message.into(),
|
||||
context_window: true,
|
||||
}
|
||||
}
|
||||
|
||||
#[must_use]
|
||||
pub const fn is_context_window_failure(&self) -> bool {
|
||||
self.context_window
|
||||
}
|
||||
}
|
||||
|
||||
impl Display for RuntimeError {
|
||||
@@ -140,6 +202,7 @@ pub struct ConversationRuntime<C, T> {
|
||||
hook_abort_signal: HookAbortSignal,
|
||||
hook_progress_reporter: Option<Box<dyn HookProgressReporter>>,
|
||||
session_tracer: Option<SessionTracer>,
|
||||
max_stall_nudges: usize,
|
||||
}
|
||||
|
||||
impl<C, T> ConversationRuntime<C, T>
|
||||
@@ -194,9 +257,17 @@ where
|
||||
hook_abort_signal: HookAbortSignal::default(),
|
||||
hook_progress_reporter: None,
|
||||
session_tracer: None,
|
||||
max_stall_nudges: max_stall_nudges_from_env(),
|
||||
}
|
||||
}
|
||||
|
||||
/// Override how many auto-continue nudges a stalled turn may receive. `0` disables them.
|
||||
#[must_use]
|
||||
pub fn with_max_stall_nudges(mut self, max_stall_nudges: usize) -> Self {
|
||||
self.max_stall_nudges = max_stall_nudges;
|
||||
self
|
||||
}
|
||||
|
||||
#[must_use]
|
||||
pub fn with_max_iterations(mut self, max_iterations: usize) -> Self {
|
||||
self.max_iterations = max_iterations;
|
||||
@@ -326,14 +397,33 @@ where
|
||||
}
|
||||
}
|
||||
|
||||
#[allow(clippy::too_many_lines)]
|
||||
pub fn run_turn(
|
||||
&mut self,
|
||||
user_input: impl Into<String>,
|
||||
prompter: Option<&mut dyn PermissionPrompter>,
|
||||
) -> Result<TurnSummary, RuntimeError> {
|
||||
self.drive_turn(Some(user_input.into()), prompter)
|
||||
}
|
||||
|
||||
/// Continues the turn already recorded in the session, without adding a user message.
|
||||
///
|
||||
/// Recovery paths retry a turn that failed part-way through. Calling [`Self::run_turn`] again
|
||||
/// with the original input would append that input a second time and invite the model to redo
|
||||
/// tool calls whose side effects have already landed. Resuming reads the transcript as it
|
||||
/// stands — including the tool results that did complete — and carries on from there.
|
||||
pub fn resume_turn(
|
||||
&mut self,
|
||||
prompter: Option<&mut dyn PermissionPrompter>,
|
||||
) -> Result<TurnSummary, RuntimeError> {
|
||||
self.drive_turn(None, prompter)
|
||||
}
|
||||
|
||||
#[allow(clippy::too_many_lines)]
|
||||
fn drive_turn(
|
||||
&mut self,
|
||||
user_input: Option<String>,
|
||||
mut prompter: Option<&mut dyn PermissionPrompter>,
|
||||
) -> Result<TurnSummary, RuntimeError> {
|
||||
let user_input = user_input.into();
|
||||
|
||||
// ROADMAP #38: Session-health canary - probe if context was compacted
|
||||
if self.session.compaction.is_some() {
|
||||
if let Err(error) = self.run_session_health_probe() {
|
||||
@@ -345,16 +435,24 @@ where
|
||||
}
|
||||
}
|
||||
|
||||
self.record_turn_started(&user_input);
|
||||
self.session
|
||||
.push_user_text(user_input)
|
||||
.map_err(|error| RuntimeError::new(error.to_string()))?;
|
||||
match user_input {
|
||||
Some(user_input) => {
|
||||
self.record_turn_started(&user_input);
|
||||
self.session
|
||||
.push_user_text(user_input)
|
||||
.map_err(|error| RuntimeError::new(error.to_string()))?;
|
||||
}
|
||||
None => self.record_turn_started(RESUMED_TURN_MARKER),
|
||||
}
|
||||
|
||||
let mut assistant_messages = Vec::new();
|
||||
let mut tool_results = Vec::new();
|
||||
let mut prompt_cache_events = Vec::new();
|
||||
let mut iterations = 0;
|
||||
let mut auto_compaction = None;
|
||||
let max_stall_nudges = self.max_stall_nudges;
|
||||
let mut stall_nudges = 0;
|
||||
let mut tools_ran = false;
|
||||
|
||||
loop {
|
||||
iterations += 1;
|
||||
@@ -366,6 +464,23 @@ where
|
||||
return Err(error);
|
||||
}
|
||||
|
||||
// Compact BEFORE the request goes out, not only after a reply lands.
|
||||
//
|
||||
// The usage-based check below is post-hoc: it can only see the context once a call
|
||||
// has returned and reported its counters. Between that measurement and the next
|
||||
// request the session can grow enormously — a reasoning model emitting a long
|
||||
// <think> block, then a tool result carrying a large file. Observed live: a call
|
||||
// measured at 17,858 prompt tokens generated 10,842 more, a read_file added ~19k on
|
||||
// top, and the next request hit the provider at 47,755 tokens and was rejected
|
||||
// outright. No amount of tuning the post-hoc threshold prevents that, because
|
||||
// nothing measures the session in between.
|
||||
//
|
||||
// `estimate_session_tokens` is a local heuristic over the transcript, so this costs
|
||||
// no API call and works before any usage has ever been recorded.
|
||||
if let Some(compaction) = self.maybe_auto_compact_before_request() {
|
||||
auto_compaction = Some(compaction);
|
||||
}
|
||||
|
||||
let request = ApiRequest {
|
||||
system_prompt: self.system_prompt.clone(),
|
||||
messages: self.session.messages.clone(),
|
||||
@@ -381,6 +496,7 @@ where
|
||||
match build_assistant_message(events) {
|
||||
Ok(result) => result,
|
||||
Err(error) => {
|
||||
let error = self.classify_empty_reply(error);
|
||||
self.record_turn_failed(iterations, &error);
|
||||
return Err(error);
|
||||
}
|
||||
@@ -417,8 +533,35 @@ where
|
||||
}
|
||||
|
||||
if pending_tool_uses.is_empty() {
|
||||
// A reply with no tool call usually means the turn is done. On small local
|
||||
// models it just as often means the model narrated the next step — or asked
|
||||
// for a go-ahead mid-task — and then emitted EOS, leaving the announced work
|
||||
// undone. Nudge it back into the loop instead of dropping the user at the
|
||||
// prompt, bounded so a model that insists it is finished still gets to stop.
|
||||
let stalled = stall_nudges < max_stall_nudges
|
||||
&& is_stalled_reply(
|
||||
&assistant_text(
|
||||
assistant_messages
|
||||
.last()
|
||||
.expect("assistant message pushed above"),
|
||||
),
|
||||
tools_ran,
|
||||
);
|
||||
if stalled {
|
||||
stall_nudges += 1;
|
||||
// Pushed as a system turn, not a user one. The model needs to see it, but the
|
||||
// transcript must not claim the user asked for it: this text is persisted,
|
||||
// re-sent on every later request, and folded into the next compaction summary,
|
||||
// where a user-labelled nudge becomes indistinguishable from something the
|
||||
// user actually said.
|
||||
self.session
|
||||
.push_message(ConversationMessage::system_text(STALL_NUDGE_PROMPT))
|
||||
.map_err(|error| RuntimeError::new(error.to_string()))?;
|
||||
continue;
|
||||
}
|
||||
break;
|
||||
}
|
||||
tools_ran = true;
|
||||
|
||||
for (tool_use_id, tool_name, input) in pending_tool_uses {
|
||||
let pre_hook_result = self.run_pre_tool_use_hook(&tool_name, &input);
|
||||
@@ -573,13 +716,62 @@ where
|
||||
self.session
|
||||
}
|
||||
|
||||
/// Compaction check run immediately before a request is sent, based on a locally estimated
|
||||
/// transcript size rather than on provider-reported usage.
|
||||
///
|
||||
/// This is the guard that actually keeps a session inside the context window: it sees growth
|
||||
/// that usage counters cannot, because it runs after tool results and long generations have
|
||||
/// been appended but before the request that would carry them is built.
|
||||
/// Decides whether an empty or unreadable reply was an overflow, by size rather than by text.
|
||||
///
|
||||
/// A stream that produces no content says nothing about why. Some backends answer an
|
||||
/// oversized request that way instead of with a typed error, so the failure is worth
|
||||
/// recovering from — but only when the session was already large enough for that to be the
|
||||
/// plausible cause. Below the compaction threshold it is a blip, and discarding the
|
||||
/// transcript would cost more than the retry could win.
|
||||
fn classify_empty_reply(&self, error: RuntimeError) -> RuntimeError {
|
||||
if error.is_context_window_failure() {
|
||||
return error;
|
||||
}
|
||||
if estimate_session_tokens(&self.session)
|
||||
>= self.auto_compaction_input_tokens_threshold as usize
|
||||
{
|
||||
return RuntimeError::context_window(error.to_string());
|
||||
}
|
||||
error
|
||||
}
|
||||
|
||||
fn maybe_auto_compact_before_request(&mut self) -> Option<AutoCompactionEvent> {
|
||||
let estimated = estimate_session_tokens(&self.session);
|
||||
if estimated < self.auto_compaction_input_tokens_threshold as usize {
|
||||
return None;
|
||||
}
|
||||
self.compact_now()
|
||||
}
|
||||
|
||||
fn maybe_auto_compact(&mut self) -> Option<AutoCompactionEvent> {
|
||||
if self.usage_tracker.cumulative_usage().input_tokens
|
||||
// Gate on the LIVE context size — the most recent call's input counters — not on a
|
||||
// cumulative running total, and count cached input as the context it is.
|
||||
//
|
||||
// The previous gate (`cumulative_usage().input_tokens`) was wrong twice over:
|
||||
// 1. It ignored `cache_read_input_tokens`. With prompt caching a warm call reports
|
||||
// `input_tokens: 2`, so the counter crawled and compaction never fired however
|
||||
// large the context grew.
|
||||
// 2. Summing across calls is not a context measurement, and the sum never decreased,
|
||||
// so once it did trip, every later iteration compacted forever.
|
||||
// Reading the latest turn fixes both: it rises with the real context and falls back
|
||||
// below the threshold as soon as a compaction has done its job.
|
||||
if self.usage_tracker.current_turn_usage().total_input_tokens()
|
||||
< self.auto_compaction_input_tokens_threshold
|
||||
{
|
||||
return None;
|
||||
}
|
||||
|
||||
self.compact_now()
|
||||
}
|
||||
|
||||
/// Compacts the session unconditionally, returning `None` when nothing actually changed.
|
||||
fn compact_now(&mut self) -> Option<AutoCompactionEvent> {
|
||||
let result = compact_session(
|
||||
&self.session,
|
||||
CompactionConfig {
|
||||
@@ -588,7 +780,12 @@ where
|
||||
},
|
||||
);
|
||||
|
||||
if result.removed_message_count == 0 {
|
||||
// `removed_message_count` alone is the wrong test for "did anything happen": compaction
|
||||
// also trims oversized tool-result payloads in messages it must PRESERVE, and on a
|
||||
// session too short to have anything droppable that trimming is the entire result.
|
||||
// Gating on the count would compute the smaller session and then throw it away.
|
||||
let shrank = result.compacted_session.messages != self.session.messages;
|
||||
if result.removed_message_count == 0 && !shrank {
|
||||
return None;
|
||||
}
|
||||
|
||||
@@ -709,19 +906,88 @@ where
|
||||
/// Reads the automatic compaction threshold from the environment.
|
||||
#[must_use]
|
||||
pub fn auto_compaction_threshold_from_env() -> u32 {
|
||||
parse_auto_compaction_threshold(
|
||||
resolve_auto_compaction_threshold(
|
||||
std::env::var(AUTO_COMPACTION_THRESHOLD_ENV_VAR)
|
||||
.ok()
|
||||
.as_deref(),
|
||||
std::env::var(CONTEXT_WINDOW_ENV_VAR).ok().as_deref(),
|
||||
)
|
||||
}
|
||||
|
||||
/// Resolves the threshold from an explicit setting, else from the declared context window.
|
||||
///
|
||||
/// `DEFAULT_AUTO_COMPACTION_INPUT_TOKENS_THRESHOLD` is a 200k-window number. Against a smaller
|
||||
/// server it sits above the window entirely, so the pre-request guard can never fire and the
|
||||
/// first sign of trouble is the backend rejecting the request. Deriving from the declared window
|
||||
/// keeps the guard meaningful without every caller having to compute a second number by hand.
|
||||
#[must_use]
|
||||
fn resolve_auto_compaction_threshold(explicit: Option<&str>, context_window: Option<&str>) -> u32 {
|
||||
if let Some(threshold) = parse_positive_u32(explicit) {
|
||||
return threshold;
|
||||
}
|
||||
parse_positive_u32(context_window).map_or(
|
||||
DEFAULT_AUTO_COMPACTION_INPUT_TOKENS_THRESHOLD,
|
||||
|window| {
|
||||
(u64::from(window) * u64::from(AUTO_COMPACTION_WINDOW_PERCENT) / 100)
|
||||
.try_into()
|
||||
.unwrap_or(DEFAULT_AUTO_COMPACTION_INPUT_TOKENS_THRESHOLD)
|
||||
},
|
||||
)
|
||||
}
|
||||
|
||||
#[must_use]
|
||||
fn parse_auto_compaction_threshold(value: Option<&str>) -> u32 {
|
||||
fn parse_positive_u32(value: Option<&str>) -> Option<u32> {
|
||||
value
|
||||
.and_then(|raw| raw.trim().parse::<u32>().ok())
|
||||
.filter(|threshold| *threshold > 0)
|
||||
.unwrap_or(DEFAULT_AUTO_COMPACTION_INPUT_TOKENS_THRESHOLD)
|
||||
.filter(|parsed| *parsed > 0)
|
||||
}
|
||||
|
||||
/// How many auto-continue nudges a single turn may inject. `0` disables the behaviour.
|
||||
#[must_use]
|
||||
pub fn max_stall_nudges_from_env() -> usize {
|
||||
parse_max_stall_nudges(std::env::var(MAX_STALL_NUDGES_ENV_VAR).ok().as_deref())
|
||||
}
|
||||
|
||||
#[must_use]
|
||||
fn parse_max_stall_nudges(value: Option<&str>) -> usize {
|
||||
value
|
||||
.and_then(|raw| raw.trim().parse::<usize>().ok())
|
||||
.unwrap_or(DEFAULT_MAX_STALL_NUDGES)
|
||||
}
|
||||
|
||||
/// Concatenated text of an assistant message, ignoring thinking and tool blocks.
|
||||
#[must_use]
|
||||
fn assistant_text(message: &ConversationMessage) -> String {
|
||||
message
|
||||
.blocks
|
||||
.iter()
|
||||
.filter_map(|block| match block {
|
||||
ContentBlock::Text { text } => Some(text.as_str()),
|
||||
_ => None,
|
||||
})
|
||||
.collect::<Vec<_>>()
|
||||
.join("\n")
|
||||
}
|
||||
|
||||
/// True when a tool-less reply is a stall rather than a finished turn.
|
||||
///
|
||||
/// Small local models routinely narrate the next tool call in prose and then emit EOS, which
|
||||
/// ends the turn with the announced work never performed. `after_tool_use` additionally treats
|
||||
/// a mid-task question as a stall: once tools have run in this turn, stopping to ask for a
|
||||
/// go-ahead is the same failure wearing a different hat.
|
||||
#[must_use]
|
||||
fn is_stalled_reply(text: &str, after_tool_use: bool) -> bool {
|
||||
let Some(last_line) = text.lines().map(str::trim).rfind(|line| !line.is_empty()) else {
|
||||
return false;
|
||||
};
|
||||
let lowered = last_line.to_lowercase();
|
||||
|
||||
if after_tool_use && last_line.ends_with('?') {
|
||||
return true;
|
||||
}
|
||||
PENDING_ACTION_MARKERS
|
||||
.iter()
|
||||
.any(|marker| lowered.contains(marker))
|
||||
}
|
||||
|
||||
fn build_assistant_message(
|
||||
@@ -853,11 +1119,13 @@ impl ToolExecutor for StaticToolExecutor {
|
||||
#[cfg(test)]
|
||||
mod tests {
|
||||
use super::{
|
||||
build_assistant_message, parse_auto_compaction_threshold, ApiClient, ApiRequest,
|
||||
AssistantEvent, AutoCompactionEvent, ConversationRuntime, PromptCacheEvent, RuntimeError,
|
||||
assistant_text, build_assistant_message, is_stalled_reply, parse_max_stall_nudges,
|
||||
resolve_auto_compaction_threshold, ApiClient, ApiRequest, AssistantEvent,
|
||||
AutoCompactionEvent, ConversationRuntime, PromptCacheEvent, RuntimeError,
|
||||
StaticToolExecutor, ToolExecutor, DEFAULT_AUTO_COMPACTION_INPUT_TOKENS_THRESHOLD,
|
||||
DEFAULT_MAX_STALL_NUDGES,
|
||||
};
|
||||
use crate::compact::CompactionConfig;
|
||||
use crate::compact::{estimate_session_tokens, CompactionConfig};
|
||||
use crate::config::{RuntimeFeatureConfig, RuntimeHookConfig};
|
||||
use crate::permissions::{
|
||||
PermissionMode, PermissionPolicy, PermissionPromptDecision, PermissionPrompter,
|
||||
@@ -1701,19 +1969,246 @@ mod tests {
|
||||
assert_eq!(runtime.session().messages.len(), 2);
|
||||
}
|
||||
|
||||
/// Prompt caching reports the bulk of the context as `cache_read_input_tokens` and leaves
|
||||
/// `input_tokens` at a token or two, so a gate that reads `input_tokens` alone never fires
|
||||
/// on a cached session no matter how large the context has grown.
|
||||
#[test]
|
||||
fn auto_compaction_counts_cached_context_toward_threshold() {
|
||||
struct CachedApi;
|
||||
impl ApiClient for CachedApi {
|
||||
fn stream(
|
||||
&mut self,
|
||||
_request: ApiRequest,
|
||||
) -> Result<Vec<AssistantEvent>, RuntimeError> {
|
||||
Ok(vec![
|
||||
AssistantEvent::TextDelta("done".to_string()),
|
||||
AssistantEvent::Usage(TokenUsage {
|
||||
input_tokens: 2,
|
||||
output_tokens: 4,
|
||||
cache_creation_input_tokens: 1_804,
|
||||
cache_read_input_tokens: 112_639,
|
||||
}),
|
||||
AssistantEvent::MessageStop,
|
||||
])
|
||||
}
|
||||
}
|
||||
|
||||
let mut session = Session::new();
|
||||
session.messages = vec![
|
||||
crate::session::ConversationMessage::user_text("one"),
|
||||
crate::session::ConversationMessage::assistant(vec![ContentBlock::Text {
|
||||
text: "two".to_string(),
|
||||
}]),
|
||||
crate::session::ConversationMessage::user_text("three"),
|
||||
crate::session::ConversationMessage::assistant(vec![ContentBlock::Text {
|
||||
text: "four".to_string(),
|
||||
}]),
|
||||
];
|
||||
|
||||
let mut runtime = ConversationRuntime::new(
|
||||
session,
|
||||
CachedApi,
|
||||
StaticToolExecutor::new(),
|
||||
PermissionPolicy::new(PermissionMode::DangerFullAccess),
|
||||
vec!["system".to_string()],
|
||||
)
|
||||
.with_auto_compaction_input_tokens_threshold(100_000);
|
||||
|
||||
let summary = runtime
|
||||
.run_turn("trigger", None)
|
||||
.expect("turn should succeed");
|
||||
|
||||
assert_eq!(
|
||||
summary.auto_compaction,
|
||||
Some(AutoCompactionEvent {
|
||||
removed_message_count: 2,
|
||||
})
|
||||
);
|
||||
}
|
||||
|
||||
/// The threshold describes the LIVE context, not a monotonic running total. Once compaction
|
||||
/// has shrunk the context back down, later small turns must not keep compacting.
|
||||
#[test]
|
||||
fn auto_compaction_stops_once_the_context_shrinks_again() {
|
||||
struct ShrinkingApi {
|
||||
calls: u32,
|
||||
}
|
||||
impl ApiClient for ShrinkingApi {
|
||||
fn stream(
|
||||
&mut self,
|
||||
_request: ApiRequest,
|
||||
) -> Result<Vec<AssistantEvent>, RuntimeError> {
|
||||
let call = self.calls;
|
||||
self.calls += 1;
|
||||
// The first call is over budget; every call after it is small.
|
||||
let input_tokens = if call == 0 { 150_000 } else { 900 };
|
||||
Ok(vec![
|
||||
AssistantEvent::TextDelta("done".to_string()),
|
||||
AssistantEvent::Usage(TokenUsage {
|
||||
input_tokens,
|
||||
output_tokens: 4,
|
||||
cache_creation_input_tokens: 0,
|
||||
cache_read_input_tokens: 0,
|
||||
}),
|
||||
AssistantEvent::MessageStop,
|
||||
])
|
||||
}
|
||||
}
|
||||
|
||||
let mut session = Session::new();
|
||||
session.messages = vec![
|
||||
crate::session::ConversationMessage::user_text("one"),
|
||||
crate::session::ConversationMessage::assistant(vec![ContentBlock::Text {
|
||||
text: "two".to_string(),
|
||||
}]),
|
||||
crate::session::ConversationMessage::user_text("three"),
|
||||
crate::session::ConversationMessage::assistant(vec![ContentBlock::Text {
|
||||
text: "four".to_string(),
|
||||
}]),
|
||||
];
|
||||
|
||||
let mut runtime = ConversationRuntime::new(
|
||||
session,
|
||||
ShrinkingApi { calls: 0 },
|
||||
// Every turn after a compaction runs the session health probe, which calls
|
||||
// `glob_search`; without it the probe fails the turn before the assertion is reached.
|
||||
StaticToolExecutor::new().register("glob_search", |_| Ok(String::new())),
|
||||
PermissionPolicy::new(PermissionMode::DangerFullAccess),
|
||||
vec!["system".to_string()],
|
||||
)
|
||||
.with_auto_compaction_input_tokens_threshold(100_000);
|
||||
|
||||
let first = runtime.run_turn("trigger", None).expect("first turn");
|
||||
assert!(
|
||||
first.auto_compaction.is_some(),
|
||||
"the over-budget turn should compact"
|
||||
);
|
||||
|
||||
let second = runtime.run_turn("again", None).expect("second turn");
|
||||
assert_eq!(
|
||||
second.auto_compaction, None,
|
||||
"a small turn after compaction must not compact again"
|
||||
);
|
||||
let third = runtime.run_turn("and again", None).expect("third turn");
|
||||
assert_eq!(
|
||||
third.auto_compaction, None,
|
||||
"compaction must not latch on once the threshold has been crossed"
|
||||
);
|
||||
}
|
||||
|
||||
/// Usage-based compaction is post-hoc: it only sees the context AFTER a call returns. A
|
||||
/// large tool result (or a long reasoning block) appended after that measurement can blow
|
||||
/// the window on the very next request, which the provider rejects outright — so there must
|
||||
/// also be a check BEFORE the request goes out, based on a locally estimated session size.
|
||||
#[test]
|
||||
fn auto_compaction_fires_before_the_first_request_when_the_session_is_already_oversized() {
|
||||
struct RecordingApi {
|
||||
largest_request_messages: std::rc::Rc<std::cell::Cell<usize>>,
|
||||
}
|
||||
impl ApiClient for RecordingApi {
|
||||
fn stream(&mut self, request: ApiRequest) -> Result<Vec<AssistantEvent>, RuntimeError> {
|
||||
self.largest_request_messages.set(
|
||||
self.largest_request_messages
|
||||
.get()
|
||||
.max(request.messages.len()),
|
||||
);
|
||||
Ok(vec![
|
||||
AssistantEvent::TextDelta("done".to_string()),
|
||||
// Deliberately no Usage event: nothing has been measured yet, which is
|
||||
// exactly the situation the pre-send check has to cover.
|
||||
AssistantEvent::MessageStop,
|
||||
])
|
||||
}
|
||||
}
|
||||
|
||||
// Six messages that are individually large — the shape of a session that has just
|
||||
// absorbed a couple of big file reads.
|
||||
let bulk = "x".repeat(40_000);
|
||||
let mut session = Session::new();
|
||||
for _ in 0..3 {
|
||||
session
|
||||
.messages
|
||||
.push(crate::session::ConversationMessage::user_text(bulk.clone()));
|
||||
session
|
||||
.messages
|
||||
.push(crate::session::ConversationMessage::assistant(vec![
|
||||
ContentBlock::Text { text: bulk.clone() },
|
||||
]));
|
||||
}
|
||||
let before = estimate_session_tokens(&session);
|
||||
assert!(
|
||||
before > 24_000,
|
||||
"test fixture must start over the threshold, got {before}"
|
||||
);
|
||||
|
||||
let largest = std::rc::Rc::new(std::cell::Cell::new(0usize));
|
||||
let mut runtime = ConversationRuntime::new(
|
||||
session,
|
||||
RecordingApi {
|
||||
largest_request_messages: std::rc::Rc::clone(&largest),
|
||||
},
|
||||
StaticToolExecutor::new().register("glob_search", |_| Ok(String::new())),
|
||||
PermissionPolicy::new(PermissionMode::DangerFullAccess),
|
||||
vec!["system".to_string()],
|
||||
)
|
||||
.with_auto_compaction_input_tokens_threshold(24_000);
|
||||
|
||||
let summary = runtime
|
||||
.run_turn("trigger", None)
|
||||
.expect("turn should succeed");
|
||||
|
||||
assert!(
|
||||
summary.auto_compaction.is_some(),
|
||||
"an already-oversized session must compact before the request is sent"
|
||||
);
|
||||
assert!(
|
||||
estimate_session_tokens(runtime.session()) < before,
|
||||
"compaction must actually shrink the session"
|
||||
);
|
||||
assert!(
|
||||
largest.get() < 8,
|
||||
"the request must be sent with the compacted history, not the original 7 messages"
|
||||
);
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn auto_compaction_threshold_defaults_and_parses_values() {
|
||||
assert_eq!(
|
||||
parse_auto_compaction_threshold(None),
|
||||
resolve_auto_compaction_threshold(None, None),
|
||||
DEFAULT_AUTO_COMPACTION_INPUT_TOKENS_THRESHOLD
|
||||
);
|
||||
assert_eq!(parse_auto_compaction_threshold(Some("4321")), 4321);
|
||||
assert_eq!(resolve_auto_compaction_threshold(Some("4321"), None), 4321);
|
||||
assert_eq!(
|
||||
parse_auto_compaction_threshold(Some("0")),
|
||||
resolve_auto_compaction_threshold(Some("0"), None),
|
||||
DEFAULT_AUTO_COMPACTION_INPUT_TOKENS_THRESHOLD
|
||||
);
|
||||
assert_eq!(
|
||||
parse_auto_compaction_threshold(Some("not-a-number")),
|
||||
resolve_auto_compaction_threshold(Some("not-a-number"), None),
|
||||
DEFAULT_AUTO_COMPACTION_INPUT_TOKENS_THRESHOLD
|
||||
);
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn auto_compaction_threshold_derives_from_a_declared_context_window() {
|
||||
// 70% of 65536, so the guard sits inside the window instead of 34k above it.
|
||||
assert_eq!(
|
||||
resolve_auto_compaction_threshold(None, Some("65536")),
|
||||
45875
|
||||
);
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn an_explicit_threshold_still_wins_over_the_declared_window() {
|
||||
assert_eq!(
|
||||
resolve_auto_compaction_threshold(Some("48000"), Some("65536")),
|
||||
48000
|
||||
);
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn an_unparseable_context_window_falls_back_to_the_default_threshold() {
|
||||
assert_eq!(
|
||||
resolve_auto_compaction_threshold(None, Some("plenty")),
|
||||
DEFAULT_AUTO_COMPACTION_INPUT_TOKENS_THRESHOLD
|
||||
);
|
||||
}
|
||||
@@ -1885,6 +2380,146 @@ mod tests {
|
||||
assert_eq!(error.to_string(), "unknown tool: missing");
|
||||
}
|
||||
|
||||
/// Replays a fixed script of stream outcomes, one per call.
|
||||
struct ScriptedApi {
|
||||
steps: Vec<Result<Vec<AssistantEvent>, String>>,
|
||||
calls: usize,
|
||||
}
|
||||
|
||||
impl ScriptedApi {
|
||||
fn new(steps: Vec<Result<Vec<AssistantEvent>, String>>) -> Self {
|
||||
Self { steps, calls: 0 }
|
||||
}
|
||||
}
|
||||
|
||||
impl ApiClient for ScriptedApi {
|
||||
fn stream(&mut self, _request: ApiRequest) -> Result<Vec<AssistantEvent>, RuntimeError> {
|
||||
let step = self
|
||||
.steps
|
||||
.get(self.calls)
|
||||
.cloned()
|
||||
.unwrap_or_else(|| Err("script exhausted".to_string()));
|
||||
self.calls += 1;
|
||||
step.map_err(RuntimeError::new)
|
||||
}
|
||||
}
|
||||
|
||||
fn tool_call_events(id: &str) -> Vec<AssistantEvent> {
|
||||
vec![
|
||||
AssistantEvent::ToolUse {
|
||||
id: id.to_string(),
|
||||
name: "echo".to_string(),
|
||||
input: "payload".to_string(),
|
||||
},
|
||||
AssistantEvent::MessageStop,
|
||||
]
|
||||
}
|
||||
|
||||
fn scripted_runtime(
|
||||
steps: Vec<Result<Vec<AssistantEvent>, String>>,
|
||||
) -> ConversationRuntime<ScriptedApi, StaticToolExecutor> {
|
||||
ConversationRuntime::new(
|
||||
Session::new(),
|
||||
ScriptedApi::new(steps),
|
||||
StaticToolExecutor::new().register("echo", |input| Ok(input.to_string())),
|
||||
PermissionPolicy::new(PermissionMode::DangerFullAccess),
|
||||
vec!["system".to_string()],
|
||||
)
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn a_failed_turn_keeps_the_results_of_tools_that_already_ran() {
|
||||
// given: a turn whose first iteration runs a tool, and whose second call dies
|
||||
let mut runtime = scripted_runtime(vec![
|
||||
Ok(tool_call_events("tool-1")),
|
||||
Err("upstream failed".to_string()),
|
||||
]);
|
||||
|
||||
// when
|
||||
runtime
|
||||
.run_turn("do the thing", None)
|
||||
.expect_err("the second call fails");
|
||||
|
||||
// then: the executed tool's result is still in the transcript. The side effect happened;
|
||||
// a caller that discards this record will let the model run the same tool again.
|
||||
let tool_results = runtime
|
||||
.session()
|
||||
.messages
|
||||
.iter()
|
||||
.filter(|message| {
|
||||
message
|
||||
.blocks
|
||||
.iter()
|
||||
.any(|block| matches!(block, ContentBlock::ToolResult { .. }))
|
||||
})
|
||||
.count();
|
||||
assert_eq!(tool_results, 1);
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn resume_turn_continues_without_appending_another_user_message() {
|
||||
// given: a session that already carries a user turn and a completed tool result
|
||||
let mut runtime = scripted_runtime(vec![
|
||||
Ok(tool_call_events("tool-1")),
|
||||
Err("upstream failed".to_string()),
|
||||
Ok(vec![
|
||||
AssistantEvent::TextDelta("done".to_string()),
|
||||
AssistantEvent::MessageStop,
|
||||
]),
|
||||
]);
|
||||
runtime.run_turn("do the thing", None).expect_err("fails");
|
||||
let user_messages_before = runtime
|
||||
.session()
|
||||
.messages
|
||||
.iter()
|
||||
.filter(|message| message.role == MessageRole::User)
|
||||
.count();
|
||||
|
||||
// when
|
||||
runtime
|
||||
.resume_turn(None)
|
||||
.expect("the resumed turn succeeds");
|
||||
|
||||
// then: the prompt was not repeated
|
||||
let user_messages_after = runtime
|
||||
.session()
|
||||
.messages
|
||||
.iter()
|
||||
.filter(|message| message.role == MessageRole::User)
|
||||
.count();
|
||||
assert_eq!(user_messages_before, user_messages_after);
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn a_stall_nudge_is_recorded_as_a_system_turn_not_a_user_one() {
|
||||
// given: a reply that announces work and stops, so the nudge fires once
|
||||
let mut runtime = nudge_runtime(
|
||||
vec![
|
||||
text_only("Let me check the logs."),
|
||||
text_only("The logs were clean."),
|
||||
],
|
||||
1,
|
||||
);
|
||||
runtime
|
||||
.run_turn("check the logs", None)
|
||||
.expect("the nudged turn completes");
|
||||
|
||||
// then: exactly one user message — the real one
|
||||
let user_messages = runtime
|
||||
.session()
|
||||
.messages
|
||||
.iter()
|
||||
.filter(|message| message.role == MessageRole::User)
|
||||
.count();
|
||||
assert_eq!(user_messages, 1);
|
||||
assert!(runtime
|
||||
.session()
|
||||
.messages
|
||||
.iter()
|
||||
.any(|message| message.role == MessageRole::System
|
||||
&& assistant_text(message).contains("[auto-continue]")));
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn run_turn_errors_when_max_iterations_is_exceeded() {
|
||||
struct LoopingApi;
|
||||
@@ -1956,4 +2591,136 @@ mod tests {
|
||||
// then
|
||||
assert_eq!(error.to_string(), "upstream failed");
|
||||
}
|
||||
|
||||
fn text_only(text: &str) -> Vec<AssistantEvent> {
|
||||
vec![
|
||||
AssistantEvent::TextDelta(text.to_string()),
|
||||
AssistantEvent::MessageStop,
|
||||
]
|
||||
}
|
||||
|
||||
struct CountingApi {
|
||||
calls: usize,
|
||||
replies: Vec<Vec<AssistantEvent>>,
|
||||
}
|
||||
|
||||
impl ApiClient for CountingApi {
|
||||
fn stream(&mut self, _request: ApiRequest) -> Result<Vec<AssistantEvent>, RuntimeError> {
|
||||
let reply = self
|
||||
.replies
|
||||
.get(self.calls)
|
||||
.cloned()
|
||||
.unwrap_or_else(|| text_only("Now running the next probe:"));
|
||||
self.calls += 1;
|
||||
Ok(reply)
|
||||
}
|
||||
}
|
||||
|
||||
fn nudge_runtime(
|
||||
replies: Vec<Vec<AssistantEvent>>,
|
||||
max_stall_nudges: usize,
|
||||
) -> ConversationRuntime<CountingApi, StaticToolExecutor> {
|
||||
ConversationRuntime::new(
|
||||
Session::new(),
|
||||
CountingApi { calls: 0, replies },
|
||||
StaticToolExecutor::new().register("probe", |_| Ok("probe output".to_string())),
|
||||
PermissionPolicy::new(PermissionMode::DangerFullAccess),
|
||||
vec!["system".to_string()],
|
||||
)
|
||||
.with_max_stall_nudges(max_stall_nudges)
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn announced_but_uncalled_tool_is_auto_continued() {
|
||||
// given: the model narrates the next probe instead of calling it
|
||||
let mut runtime = nudge_runtime(
|
||||
vec![
|
||||
text_only("Now running the Ford EEC-V tuner probe (ATSP1):"),
|
||||
vec![
|
||||
AssistantEvent::ToolUse {
|
||||
id: "tool-1".to_string(),
|
||||
name: "probe".to_string(),
|
||||
input: "atsp1".to_string(),
|
||||
},
|
||||
AssistantEvent::MessageStop,
|
||||
],
|
||||
text_only("No response on ATSP1, which is expected here."),
|
||||
],
|
||||
3,
|
||||
);
|
||||
|
||||
// when
|
||||
let summary = runtime.run_turn("run the probes", None).expect("turn runs");
|
||||
|
||||
// then: the nudge pulled it back into the loop and the tool actually ran
|
||||
assert_eq!(summary.iterations, 3);
|
||||
assert_eq!(summary.tool_results.len(), 1);
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn nudges_are_bounded_so_a_stuck_model_still_stops() {
|
||||
// given: every reply is an announcement and nothing is ever called
|
||||
let mut runtime = nudge_runtime(Vec::new(), 2);
|
||||
|
||||
// when
|
||||
let summary = runtime.run_turn("run the probes", None).expect("turn runs");
|
||||
|
||||
// then: initial reply plus exactly two nudged retries
|
||||
assert_eq!(summary.iterations, 3);
|
||||
assert!(summary.tool_results.is_empty());
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn a_plain_completion_is_not_nudged() {
|
||||
// given
|
||||
let mut runtime = nudge_runtime(vec![text_only("The DTC scan came back clean.")], 3);
|
||||
|
||||
// when
|
||||
let summary = runtime.run_turn("scan for codes", None).expect("turn runs");
|
||||
|
||||
// then
|
||||
assert_eq!(summary.iterations, 1);
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn zero_budget_disables_auto_continue() {
|
||||
// given
|
||||
let mut runtime = nudge_runtime(vec![text_only("Let me run the DTC scan.")], 0);
|
||||
|
||||
// when
|
||||
let summary = runtime.run_turn("scan for codes", None).expect("turn runs");
|
||||
|
||||
// then
|
||||
assert_eq!(summary.iterations, 1);
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn a_question_stalls_only_after_a_tool_has_run() {
|
||||
assert!(is_stalled_reply("Should I run the DTC scan next?", true));
|
||||
assert!(!is_stalled_reply("Which port is the adapter on?", false));
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn stall_detection_reads_the_last_line_not_the_whole_reply() {
|
||||
// an announcement earlier in the reply does not make a finished turn look stalled
|
||||
assert!(!is_stalled_reply(
|
||||
"Let me check the port.\nThe scan finished with no codes stored.",
|
||||
true
|
||||
));
|
||||
assert!(is_stalled_reply(
|
||||
"The scan finished.\nNext I'll read the freeze-frame data.",
|
||||
true
|
||||
));
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn stall_nudge_budget_parses_from_env_value() {
|
||||
// asserted against the literal, not the constant: comparing the parser to the same
|
||||
// constant it falls back to would pass no matter what the default became
|
||||
assert_eq!(DEFAULT_MAX_STALL_NUDGES, 3);
|
||||
assert_eq!(parse_max_stall_nudges(None), 3);
|
||||
assert_eq!(parse_max_stall_nudges(Some("not a number")), 3);
|
||||
assert_eq!(parse_max_stall_nudges(Some(" 5 ")), 5);
|
||||
assert_eq!(parse_max_stall_nudges(Some("0")), 0);
|
||||
}
|
||||
}
|
||||
@@ -11,8 +11,48 @@ use serde::{Deserialize, Serialize};
|
||||
use walkdir::{DirEntry, WalkDir};
|
||||
|
||||
/// Maximum file size that can be read (10 MB).
|
||||
///
|
||||
/// This is a *process* guard — it stops us slurping a huge or binary blob into memory. It is
|
||||
/// not a context guard: 10 MB is on the order of 2.5 M tokens, so it never fires on the files
|
||||
/// that actually blow a model's context window. `DEFAULT_READ_LINES` / `MAX_READ_CHARS` below
|
||||
/// are the context guard.
|
||||
const MAX_READ_SIZE: u64 = 10 * 1024 * 1024;
|
||||
|
||||
/// Lines returned by `read_file` when the caller does not pass an explicit `limit`.
|
||||
///
|
||||
/// This is a *paging* size, not a refusal. An unbounded read is the easiest way for a model to
|
||||
/// destroy its own context: one 100 KB source file is ~25 k tokens, which does not fit alongside
|
||||
/// a system prompt in a 32 k window. When that happens the backend rejects the whole request and
|
||||
/// the session dies — so reading everything at once is what makes it stop. Returning a page plus
|
||||
/// `nextOffset` means the read always succeeds and the model can keep going.
|
||||
const DEFAULT_READ_LINES: usize = 2000;
|
||||
|
||||
/// Hard ceiling on the characters `read_file` returns, applied even when `limit` is explicit.
|
||||
///
|
||||
/// `DEFAULT_READ_LINES` alone does not bound the payload: minified bundles, JSON blobs and
|
||||
/// generated sources routinely put megabytes on a handful of lines. ~64 KB is ~16 k tokens.
|
||||
///
|
||||
/// That default assumes a large (200 k) context window. Against a small self-hosted backend it
|
||||
/// is far too generous: two reads at this size put a request over 47 k tokens, which a 32 k
|
||||
/// window rejects outright — and compaction cannot recover it, because the preserved recent
|
||||
/// messages ARE the oversized tool results. Override with [`MAX_READ_CHARS_ENV_VAR`].
|
||||
const DEFAULT_MAX_READ_CHARS: usize = 64 * 1024;
|
||||
|
||||
/// Environment override for [`DEFAULT_MAX_READ_CHARS`], in characters.
|
||||
const MAX_READ_CHARS_ENV_VAR: &str = "CLAW_MAX_READ_CHARS";
|
||||
|
||||
/// Effective character ceiling for one `read_file` page.
|
||||
///
|
||||
/// Read per call rather than cached: a long-lived process should pick up a changed budget, and
|
||||
/// this runs once per read, not per line.
|
||||
fn max_read_chars() -> usize {
|
||||
std::env::var(MAX_READ_CHARS_ENV_VAR)
|
||||
.ok()
|
||||
.and_then(|raw| raw.trim().parse::<usize>().ok())
|
||||
.filter(|chars| *chars > 0)
|
||||
.unwrap_or(DEFAULT_MAX_READ_CHARS)
|
||||
}
|
||||
|
||||
/// Maximum file size that can be written (10 MB).
|
||||
const MAX_WRITE_SIZE: usize = 10 * 1024 * 1024;
|
||||
|
||||
@@ -80,6 +120,17 @@ pub struct TextFilePayload {
|
||||
pub start_line: usize,
|
||||
#[serde(rename = "totalLines")]
|
||||
pub total_lines: usize,
|
||||
/// True when the file has more lines past this page. Signals the caller that it has NOT seen
|
||||
/// the whole file, so it does not silently reason over a partial view.
|
||||
#[serde(default)]
|
||||
pub truncated: bool,
|
||||
/// Line index to pass back as `offset` to continue reading. Absent once the file is exhausted.
|
||||
#[serde(
|
||||
rename = "nextOffset",
|
||||
default,
|
||||
skip_serializing_if = "Option::is_none"
|
||||
)]
|
||||
pub next_offset: Option<usize>,
|
||||
}
|
||||
|
||||
/// Output envelope for the `read_file` tool.
|
||||
@@ -228,10 +279,28 @@ pub fn read_file(
|
||||
let content = fs::read_to_string(&absolute_path)?;
|
||||
let lines: Vec<&str> = content.lines().collect();
|
||||
let start_index = offset.unwrap_or(0).min(lines.len());
|
||||
let end_index = limit.map_or(lines.len(), |limit| {
|
||||
start_index.saturating_add(limit).min(lines.len())
|
||||
});
|
||||
// An absent `limit` pages rather than reading to EOF: see DEFAULT_READ_LINES.
|
||||
let requested_end = start_index
|
||||
.saturating_add(limit.unwrap_or(DEFAULT_READ_LINES))
|
||||
.min(lines.len());
|
||||
|
||||
// Apply the character ceiling on top of the line window, so a file with very long lines
|
||||
// cannot blow the budget the line cap was meant to enforce. Always yield at least one line,
|
||||
// otherwise an oversized single line would return nothing and the caller could not advance.
|
||||
let mut end_index = start_index;
|
||||
let mut chars = 0usize;
|
||||
let char_ceiling = max_read_chars();
|
||||
for line in &lines[start_index..requested_end] {
|
||||
let next = chars + line.chars().count() + 1;
|
||||
if next > char_ceiling && end_index > start_index {
|
||||
break;
|
||||
}
|
||||
chars = next;
|
||||
end_index += 1;
|
||||
}
|
||||
|
||||
let selected = lines[start_index..end_index].join("\n");
|
||||
let truncated = end_index < lines.len();
|
||||
|
||||
Ok(ReadFileOutput {
|
||||
kind: String::from("text"),
|
||||
@@ -241,6 +310,8 @@ pub fn read_file(
|
||||
num_lines: end_index.saturating_sub(start_index),
|
||||
start_line: start_index.saturating_add(1),
|
||||
total_lines: lines.len(),
|
||||
truncated,
|
||||
next_offset: truncated.then_some(end_index),
|
||||
},
|
||||
})
|
||||
}
|
||||
@@ -812,11 +883,18 @@ mod tests {
|
||||
|
||||
use super::{
|
||||
component_contains_glob, derive_glob_walk_root, edit_file, expand_braces, glob_search,
|
||||
grep_search, is_symlink_escape, normalize_for_comparison, read_file,
|
||||
grep_search, is_symlink_escape, max_read_chars, normalize_for_comparison, read_file,
|
||||
read_file_in_workspace, write_file, write_file_in_workspace, GrepSearchInput,
|
||||
MAX_WRITE_SIZE,
|
||||
DEFAULT_MAX_READ_CHARS, DEFAULT_READ_LINES, MAX_READ_CHARS_ENV_VAR, MAX_WRITE_SIZE,
|
||||
};
|
||||
|
||||
/// Serializes the tests that mutate `CLAW_MAX_READ_CHARS`; env vars are process-global.
|
||||
fn env_lock() -> std::sync::MutexGuard<'static, ()> {
|
||||
static LOCK: std::sync::Mutex<()> = std::sync::Mutex::new(());
|
||||
LOCK.lock()
|
||||
.unwrap_or_else(std::sync::PoisonError::into_inner)
|
||||
}
|
||||
|
||||
fn temp_path(name: &str) -> std::path::PathBuf {
|
||||
let unique = SystemTime::now()
|
||||
.duration_since(UNIX_EPOCH)
|
||||
@@ -837,6 +915,132 @@ mod tests {
|
||||
assert_eq!(read_output.file.content, "two");
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn unbounded_read_of_a_large_file_pages_instead_of_returning_everything() {
|
||||
// Regression: an unbounded read used to return the whole file, which on a small-context
|
||||
// backend overflows the window and the request is rejected outright — the session dies.
|
||||
let path = temp_path("paged-read.txt");
|
||||
let total = DEFAULT_READ_LINES + 500;
|
||||
let body = (0..total)
|
||||
.map(|i| format!("line {i}"))
|
||||
.collect::<Vec<_>>()
|
||||
.join("\n");
|
||||
write_file(path.to_string_lossy().as_ref(), &body).expect("write should succeed");
|
||||
|
||||
let first = read_file(path.to_string_lossy().as_ref(), None, None).expect("first page");
|
||||
assert_eq!(first.file.num_lines, DEFAULT_READ_LINES);
|
||||
assert_eq!(first.file.total_lines, total);
|
||||
assert!(first.file.truncated, "caller must be told it is partial");
|
||||
assert_eq!(first.file.next_offset, Some(DEFAULT_READ_LINES));
|
||||
assert!(first.file.content.starts_with("line 0\n"));
|
||||
|
||||
// The advertised next_offset must actually resume where the last page stopped.
|
||||
let second = read_file(
|
||||
path.to_string_lossy().as_ref(),
|
||||
first.file.next_offset,
|
||||
None,
|
||||
)
|
||||
.expect("second page");
|
||||
assert_eq!(second.file.num_lines, 500);
|
||||
assert!(!second.file.truncated, "the tail is the final page");
|
||||
assert_eq!(second.file.next_offset, None);
|
||||
assert!(second
|
||||
.file
|
||||
.content
|
||||
.starts_with(&format!("line {DEFAULT_READ_LINES}\n")));
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn char_ceiling_bounds_files_whose_lines_are_enormous() {
|
||||
// A line cap alone does not bound the payload: minified bundles put megabytes on a few
|
||||
// lines, so the byte ceiling has to apply too — and must still yield forward progress.
|
||||
let _guard = env_lock();
|
||||
std::env::remove_var(MAX_READ_CHARS_ENV_VAR);
|
||||
let path = temp_path("long-lines.txt");
|
||||
let huge = "x".repeat(DEFAULT_MAX_READ_CHARS);
|
||||
let body = format!("{huge}\n{huge}\n{huge}");
|
||||
write_file(path.to_string_lossy().as_ref(), &body).expect("write should succeed");
|
||||
|
||||
let page = read_file(path.to_string_lossy().as_ref(), None, None).expect("first page");
|
||||
assert_eq!(page.file.num_lines, 1, "one oversized line still advances");
|
||||
assert!(page.file.truncated);
|
||||
assert_eq!(page.file.next_offset, Some(1));
|
||||
|
||||
// An explicit limit must not be able to defeat the ceiling.
|
||||
let greedy = read_file(path.to_string_lossy().as_ref(), None, Some(3)).expect("greedy");
|
||||
assert_eq!(greedy.file.num_lines, 1);
|
||||
assert!(greedy.file.truncated);
|
||||
}
|
||||
|
||||
/// The 64 KB default is a 200 k-context number. A small self-hosted backend must be able to
|
||||
/// shrink it, or two reads overflow a 32 k window in a single turn.
|
||||
#[test]
|
||||
fn read_char_ceiling_is_configurable_via_the_environment() {
|
||||
let _guard = env_lock();
|
||||
let path = temp_path("configurable-ceiling.txt");
|
||||
// 40 lines of 100 chars = ~4 KB: far under the default ceiling, over a 1 KB one.
|
||||
let body = (0..40)
|
||||
.map(|_| "y".repeat(100))
|
||||
.collect::<Vec<_>>()
|
||||
.join("\n");
|
||||
write_file(path.to_string_lossy().as_ref(), &body).expect("write should succeed");
|
||||
|
||||
std::env::remove_var(MAX_READ_CHARS_ENV_VAR);
|
||||
let whole = read_file(path.to_string_lossy().as_ref(), None, None).expect("default read");
|
||||
assert_eq!(
|
||||
whole.file.num_lines, 40,
|
||||
"default ceiling returns the file whole"
|
||||
);
|
||||
assert!(!whole.file.truncated);
|
||||
|
||||
std::env::set_var(MAX_READ_CHARS_ENV_VAR, "1024");
|
||||
let clipped = read_file(path.to_string_lossy().as_ref(), None, None).expect("clipped read");
|
||||
std::env::remove_var(MAX_READ_CHARS_ENV_VAR);
|
||||
assert!(
|
||||
clipped.file.num_lines < 40,
|
||||
"a smaller ceiling must actually clip, got {} lines",
|
||||
clipped.file.num_lines
|
||||
);
|
||||
assert!(
|
||||
clipped.file.truncated,
|
||||
"a clipped page must report truncated"
|
||||
);
|
||||
assert_eq!(
|
||||
clipped.file.next_offset,
|
||||
Some(clipped.file.num_lines),
|
||||
"the caller must be told where to resume"
|
||||
);
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn read_char_ceiling_ignores_junk_and_zero_values() {
|
||||
let _guard = env_lock();
|
||||
for value in ["", "0", "-5", "not-a-number"] {
|
||||
std::env::set_var(MAX_READ_CHARS_ENV_VAR, value);
|
||||
assert_eq!(
|
||||
max_read_chars(),
|
||||
DEFAULT_MAX_READ_CHARS,
|
||||
"{value:?} must fall back to the default rather than disabling reads"
|
||||
);
|
||||
}
|
||||
std::env::set_var(MAX_READ_CHARS_ENV_VAR, "24000");
|
||||
assert_eq!(max_read_chars(), 24_000);
|
||||
std::env::remove_var(MAX_READ_CHARS_ENV_VAR);
|
||||
assert_eq!(max_read_chars(), DEFAULT_MAX_READ_CHARS);
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn small_files_are_returned_whole_and_not_marked_truncated() {
|
||||
let path = temp_path("small-read.txt");
|
||||
write_file(path.to_string_lossy().as_ref(), "one\ntwo\nthree").expect("write");
|
||||
|
||||
let output = read_file(path.to_string_lossy().as_ref(), None, None).expect("read");
|
||||
assert_eq!(output.file.content, "one\ntwo\nthree");
|
||||
assert_eq!(output.file.num_lines, 3);
|
||||
assert!(!output.file.truncated);
|
||||
assert_eq!(output.file.next_offset, None);
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn edits_file_contents() {
|
||||
let path = temp_path("edit.txt");
|
||||
|
||||
@@ -42,7 +42,25 @@ pub const SYSTEM_PROMPT_DYNAMIC_BOUNDARY: &str = "__SYSTEM_PROMPT_DYNAMIC_BOUNDA
|
||||
pub const FRONTIER_MODEL_NAME: &str = "Claude Opus 4.6";
|
||||
const MAX_INSTRUCTION_FILE_CHARS: usize = 4_000;
|
||||
const MAX_TOTAL_INSTRUCTION_CHARS: usize = 12_000;
|
||||
const MAX_GIT_DIFF_CHARS: usize = 50_000;
|
||||
/// Character budget for the git diff snapshot embedded in the system prompt.
|
||||
///
|
||||
/// This lives in the *system prompt*, so its cost is paid on every session whose
|
||||
/// working tree has changed — and a diff changes on essentially every edit. The
|
||||
/// previous 50_000 budget was four times the entire instruction-file budget and
|
||||
/// could add ~15_000 tokens to the prompt from a single modified file, which is
|
||||
/// affordable against a hosted frontier model and ruinous against a local one.
|
||||
///
|
||||
/// Override with `CLAW_MAX_GIT_DIFF_CHARS`. `0` drops the diff snapshot entirely;
|
||||
/// git status and the recent-commit list are unaffected.
|
||||
const DEFAULT_MAX_GIT_DIFF_CHARS: usize = 4_000;
|
||||
const MAX_GIT_DIFF_CHARS_ENV_VAR: &str = "CLAW_MAX_GIT_DIFF_CHARS";
|
||||
|
||||
fn max_git_diff_chars() -> usize {
|
||||
std::env::var(MAX_GIT_DIFF_CHARS_ENV_VAR)
|
||||
.ok()
|
||||
.and_then(|value| value.trim().parse().ok())
|
||||
.unwrap_or(DEFAULT_MAX_GIT_DIFF_CHARS)
|
||||
}
|
||||
|
||||
/// Neutral identity for the model family line in generated prompts.
|
||||
#[derive(Debug, Clone, Copy, Default, PartialEq, Eq)]
|
||||
@@ -498,8 +516,15 @@ fn read_git_diff(cwd: &Path) -> Option<String> {
|
||||
}
|
||||
|
||||
fn truncate_diff(mut diff: String) -> String {
|
||||
if diff.len() > MAX_GIT_DIFF_CHARS {
|
||||
let mut end = MAX_GIT_DIFF_CHARS;
|
||||
let budget = max_git_diff_chars();
|
||||
// A zero budget means "omit the diff", not "truncate to nothing" -- returning
|
||||
// an empty string here would still push an empty "Git diff snapshot:" heading
|
||||
// into the prompt.
|
||||
if budget == 0 {
|
||||
return String::new();
|
||||
}
|
||||
if diff.len() > budget {
|
||||
let mut end = budget;
|
||||
while !diff.is_char_boundary(end) {
|
||||
end -= 1;
|
||||
}
|
||||
@@ -758,6 +783,7 @@ fn get_simple_doing_tasks_section() -> String {
|
||||
"Do not add speculative abstractions, compatibility shims, or unrelated cleanup.".to_string(),
|
||||
"Do not create files unless they are required to complete the task.".to_string(),
|
||||
"If an approach fails, diagnose the failure before switching tactics.".to_string(),
|
||||
"Act in the turn you plan in: when the next step is a tool call, make the call. Never end a turn describing a step you have not taken, and do not stop to ask for a go-ahead on work already requested.".to_string(),
|
||||
"Be careful not to introduce security vulnerabilities such as command injection, XSS, or SQL injection.".to_string(),
|
||||
"Report outcomes faithfully: if verification fails or was not run, say so explicitly.".to_string(),
|
||||
]);
|
||||
@@ -782,7 +808,8 @@ mod tests {
|
||||
collapse_blank_lines, display_context_path, normalize_instruction_content,
|
||||
render_instruction_content, render_instruction_files, truncate_diff,
|
||||
truncate_instruction_content, ContextFile, ModelFamilyIdentity, ProjectContext,
|
||||
SystemPromptBuilder, MAX_GIT_DIFF_CHARS, SYSTEM_PROMPT_DYNAMIC_BOUNDARY,
|
||||
SystemPromptBuilder, DEFAULT_MAX_GIT_DIFF_CHARS, MAX_GIT_DIFF_CHARS_ENV_VAR,
|
||||
SYSTEM_PROMPT_DYNAMIC_BOUNDARY,
|
||||
};
|
||||
use crate::config::ConfigLoader;
|
||||
use std::fs;
|
||||
@@ -1444,33 +1471,69 @@ mod tests {
|
||||
assert!(rendered.contains("Project rules"));
|
||||
}
|
||||
|
||||
/// Serializes the tests that read or mutate `CLAW_MAX_GIT_DIFF_CHARS`; env vars
|
||||
/// are process-global and the default-budget tests fail if another test leaves
|
||||
/// an override set.
|
||||
fn diff_env_lock() -> std::sync::MutexGuard<'static, ()> {
|
||||
static LOCK: std::sync::Mutex<()> = std::sync::Mutex::new(());
|
||||
LOCK.lock()
|
||||
.unwrap_or_else(std::sync::PoisonError::into_inner)
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn truncate_diff_preserves_short_content() {
|
||||
let _guard = diff_env_lock();
|
||||
std::env::remove_var(MAX_GIT_DIFF_CHARS_ENV_VAR);
|
||||
let short = "a".repeat(1_000);
|
||||
let result = truncate_diff(short.clone());
|
||||
assert_eq!(result, short);
|
||||
assert!(!result.contains("[diff truncated"));
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn truncate_diff_honors_env_override() {
|
||||
let _guard = diff_env_lock();
|
||||
std::env::set_var(MAX_GIT_DIFF_CHARS_ENV_VAR, "100");
|
||||
let result = truncate_diff("z".repeat(1_000));
|
||||
std::env::remove_var(MAX_GIT_DIFF_CHARS_ENV_VAR);
|
||||
let marker = "\n\n... [diff truncated — too large for system prompt]";
|
||||
assert!(result.ends_with(marker));
|
||||
assert_eq!(result.len() - marker.len(), 100);
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn truncate_diff_zero_budget_omits_the_snapshot_entirely() {
|
||||
// Zero must yield nothing at all -- a truncated-to-empty body would still
|
||||
// drag a "Git diff snapshot:" heading and the truncation marker into the
|
||||
// prompt, which is the opposite of what the operator asked for.
|
||||
let _guard = diff_env_lock();
|
||||
std::env::set_var(MAX_GIT_DIFF_CHARS_ENV_VAR, "0");
|
||||
let result = truncate_diff("z".repeat(1_000));
|
||||
std::env::remove_var(MAX_GIT_DIFF_CHARS_ENV_VAR);
|
||||
assert!(result.is_empty());
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn truncate_diff_caps_oversized_content() {
|
||||
let large = "x".repeat(MAX_GIT_DIFF_CHARS + 5_000);
|
||||
let _guard = diff_env_lock();
|
||||
std::env::remove_var(MAX_GIT_DIFF_CHARS_ENV_VAR);
|
||||
let large = "x".repeat(DEFAULT_MAX_GIT_DIFF_CHARS + 5_000);
|
||||
let result = truncate_diff(large);
|
||||
assert!(result.contains("... [diff truncated — too large for system prompt]"));
|
||||
// The body before the marker must be at most MAX_GIT_DIFF_CHARS bytes
|
||||
// The body before the marker must be at most DEFAULT_MAX_GIT_DIFF_CHARS bytes
|
||||
let marker = "\n\n... [diff truncated — too large for system prompt]";
|
||||
let body_len = result.len() - marker.len();
|
||||
assert!(body_len <= MAX_GIT_DIFF_CHARS);
|
||||
assert!(body_len <= DEFAULT_MAX_GIT_DIFF_CHARS);
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn truncate_diff_respects_utf8_char_boundaries() {
|
||||
// Build a string where MAX_GIT_DIFF_CHARS falls in the middle of a
|
||||
// Build a string where DEFAULT_MAX_GIT_DIFF_CHARS falls in the middle of a
|
||||
// multi-byte character (U+1F600 = 4 bytes in UTF-8).
|
||||
let prefix_len = MAX_GIT_DIFF_CHARS - 2;
|
||||
let prefix_len = DEFAULT_MAX_GIT_DIFF_CHARS - 2;
|
||||
let mut input = "a".repeat(prefix_len);
|
||||
// Append a 4-byte emoji so bytes [prefix_len..prefix_len+4] are the
|
||||
// emoji. MAX_GIT_DIFF_CHARS lands at prefix_len+2, inside the emoji.
|
||||
// emoji. DEFAULT_MAX_GIT_DIFF_CHARS lands at prefix_len+2, inside the emoji.
|
||||
input.push('\u{1F600}');
|
||||
input.push_str(&"b".repeat(10_000));
|
||||
|
||||
@@ -1482,7 +1545,7 @@ mod tests {
|
||||
// inside it would be invalid UTF-8.
|
||||
let marker = "\n\n... [diff truncated — too large for system prompt]";
|
||||
let body = &result[..result.len() - marker.len()];
|
||||
assert!(body.len() <= MAX_GIT_DIFF_CHARS);
|
||||
assert!(body.len() <= DEFAULT_MAX_GIT_DIFF_CHARS);
|
||||
assert!(body.is_char_boundary(body.len()));
|
||||
}
|
||||
}
|
||||
@@ -718,6 +718,20 @@ impl ConversationMessage {
|
||||
}
|
||||
}
|
||||
|
||||
/// Text the harness injected on its own initiative, not something the user typed.
|
||||
///
|
||||
/// Reaches the model as an ordinary turn — the wire protocol has nowhere else to put it — but
|
||||
/// stays labelled in the transcript, so a replay, an export, or a compaction summary does not
|
||||
/// present the harness's own prompting back as the user's words.
|
||||
#[must_use]
|
||||
pub fn system_text(text: impl Into<String>) -> Self {
|
||||
Self {
|
||||
role: MessageRole::System,
|
||||
blocks: vec![ContentBlock::Text { text: text.into() }],
|
||||
usage: None,
|
||||
}
|
||||
}
|
||||
|
||||
#[must_use]
|
||||
pub fn assistant(blocks: Vec<ContentBlock>) -> Self {
|
||||
Self {
|
||||
|
||||
@@ -29,7 +29,7 @@ impl Default for TridentConfig {
|
||||
}
|
||||
|
||||
/// Statistics from a Trident compaction run.
|
||||
#[derive(Debug, Clone, PartialEq, Eq)]
|
||||
#[derive(Debug, Clone, PartialEq, Eq, Default)]
|
||||
pub struct TridentStats {
|
||||
pub superseded_count: usize,
|
||||
pub collapsed_chains: usize,
|
||||
@@ -41,21 +41,6 @@ pub struct TridentStats {
|
||||
pub final_message_count: usize,
|
||||
}
|
||||
|
||||
impl Default for TridentStats {
|
||||
fn default() -> Self {
|
||||
Self {
|
||||
superseded_count: 0,
|
||||
collapsed_chains: 0,
|
||||
messages_collapsed: 0,
|
||||
clusters_found: 0,
|
||||
messages_clustered: 0,
|
||||
tokens_saved_estimate: 0,
|
||||
original_message_count: 0,
|
||||
final_message_count: 0,
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
impl TridentStats {
|
||||
pub fn format_report(&self) -> String {
|
||||
let compression = if self.final_message_count > 0 {
|
||||
@@ -192,7 +177,7 @@ fn stage1_supersede(messages: &[ConversationMessage]) -> (Vec<ConversationMessag
|
||||
|
||||
let mut obsolete_indices: BTreeSet<usize> = BTreeSet::new();
|
||||
|
||||
for (_path, ops) in &file_ops {
|
||||
for ops in file_ops.values() {
|
||||
if ops.len() < 2 {
|
||||
continue;
|
||||
}
|
||||
@@ -205,10 +190,8 @@ fn stage1_supersede(messages: &[ConversationMessage]) -> (Vec<ConversationMessag
|
||||
|
||||
if let Some(last_write) = last_write_idx {
|
||||
for op in ops {
|
||||
if op.op_type == FileOp::Read && op.index < last_write {
|
||||
obsolete_indices.insert(op.index);
|
||||
} else if (op.op_type == FileOp::Write || op.op_type == FileOp::Edit)
|
||||
&& op.index < last_write
|
||||
if op.index < last_write
|
||||
&& matches!(op.op_type, FileOp::Read | FileOp::Write | FileOp::Edit)
|
||||
{
|
||||
obsolete_indices.insert(op.index);
|
||||
}
|
||||
@@ -325,7 +308,7 @@ fn stage2_collapse(
|
||||
usage: None,
|
||||
});
|
||||
} else {
|
||||
result.extend(buffer.drain(..));
|
||||
result.append(&mut buffer);
|
||||
}
|
||||
buffer.clear();
|
||||
result.push(msg.clone());
|
||||
@@ -476,7 +459,7 @@ fn stage3_cluster(
|
||||
}
|
||||
|
||||
let total_clustered: usize = cluster_assignments.len();
|
||||
let clusters_found = cluster_id as usize;
|
||||
let clusters_found = cluster_id;
|
||||
|
||||
let mut result: Vec<ConversationMessage> = Vec::new();
|
||||
let mut cluster_buffers: BTreeMap<usize, Vec<usize>> = BTreeMap::new();
|
||||
@@ -677,7 +660,7 @@ fn truncate_text(text: &str, max_chars: usize) -> String {
|
||||
mod tests {
|
||||
use super::*;
|
||||
use crate::compact::CompactionConfig;
|
||||
use crate::session::{ContentBlock, ConversationMessage, MessageRole, Session};
|
||||
use crate::session::{ContentBlock, ConversationMessage, Session};
|
||||
|
||||
#[test]
|
||||
fn stage1_removes_obsolete_file_reads() {
|
||||
@@ -736,7 +719,7 @@ mod tests {
|
||||
fn stage2_collapses_chatty_messages() {
|
||||
let mut messages = vec![];
|
||||
for i in 0..6 {
|
||||
messages.push(ConversationMessage::user_text(&format!("ok {i}")));
|
||||
messages.push(ConversationMessage::user_text(format!("ok {i}")));
|
||||
messages.push(ConversationMessage::assistant(vec![ContentBlock::Text {
|
||||
text: format!("got {i}"),
|
||||
}]));
|
||||
@@ -767,9 +750,9 @@ mod tests {
|
||||
},
|
||||
]));
|
||||
messages.push(ConversationMessage::tool_result(
|
||||
&format!("read_{i}"),
|
||||
format!("read_{i}"),
|
||||
"read_file",
|
||||
&format!(r#"{{"path":"src/{i}.rs","content":"data {i}"}}"#),
|
||||
format!(r#"{{"path":"src/{i}.rs","content":"data {i}"}}"#),
|
||||
false,
|
||||
));
|
||||
}
|
||||
|
||||
@@ -89,6 +89,19 @@ impl TokenUsage {
|
||||
+ self.cache_read_input_tokens
|
||||
}
|
||||
|
||||
/// Total tokens the provider had to read to serve this call, i.e. the live size of the
|
||||
/// context window.
|
||||
///
|
||||
/// `input_tokens` alone is NOT that number. With prompt caching a warm call reports
|
||||
/// `input_tokens: 2` and puts the real bulk in `cache_read_input_tokens`, so anything
|
||||
/// budgeting against the context window must sum all three input counters.
|
||||
#[must_use]
|
||||
pub fn total_input_tokens(self) -> u32 {
|
||||
self.input_tokens
|
||||
.saturating_add(self.cache_creation_input_tokens)
|
||||
.saturating_add(self.cache_read_input_tokens)
|
||||
}
|
||||
|
||||
#[must_use]
|
||||
pub fn estimate_cost_usd(self) -> UsageCostEstimate {
|
||||
self.estimate_cost_usd_with_pricing(ModelPricing::default_sonnet_tier())
|
||||
@@ -243,6 +256,41 @@ mod tests {
|
||||
assert_eq!(tracker.cumulative_usage().total_tokens(), 48);
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn total_input_tokens_includes_cached_context() {
|
||||
// The shape a warm prompt-cached call actually reports: input_tokens is ~nothing and
|
||||
// the real context size is in the cache counters.
|
||||
let cached = TokenUsage {
|
||||
input_tokens: 2,
|
||||
output_tokens: 1_133,
|
||||
cache_creation_input_tokens: 1_453,
|
||||
cache_read_input_tokens: 111_186,
|
||||
};
|
||||
assert_eq!(cached.total_input_tokens(), 112_641);
|
||||
// Output tokens are not part of the input budget.
|
||||
assert_ne!(cached.total_input_tokens(), cached.total_tokens());
|
||||
|
||||
// A backend without prompt caching reports the whole prompt as input_tokens.
|
||||
let uncached = TokenUsage {
|
||||
input_tokens: 29_056,
|
||||
output_tokens: 73,
|
||||
cache_creation_input_tokens: 0,
|
||||
cache_read_input_tokens: 0,
|
||||
};
|
||||
assert_eq!(uncached.total_input_tokens(), 29_056);
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn total_input_tokens_saturates_instead_of_overflowing() {
|
||||
let usage = TokenUsage {
|
||||
input_tokens: u32::MAX,
|
||||
output_tokens: 0,
|
||||
cache_creation_input_tokens: u32::MAX,
|
||||
cache_read_input_tokens: u32::MAX,
|
||||
};
|
||||
assert_eq!(usage.total_input_tokens(), u32::MAX);
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn computes_cost_summary_lines() {
|
||||
let usage = TokenUsage {
|
||||
|
||||
Reference in new issue
Block a user