repair.rs

  1//! Repair predictions that received poor quality signals.
  2//!
  3//! This module takes examples with predictions, identifies predictions that need
  4//! improvement, and uses an LLM to generate improved predictions. It supports
  5//! two sources of quality signals:
  6//! - QA feedback (reverts_edits or low confidence)
  7//! - Computed scores when QA is unavailable (high reversal_ratio or wrong_editable_region)
  8
  9use crate::{
 10    BatchProvider, PredictionProvider,
 11    anthropic_client::AnthropicClient,
 12    example::{ActualCursor, Example, ExamplePrediction},
 13    format_prompt::{TeacherPrompt, extract_last_codeblock},
 14    metrics::count_patch_token_changes,
 15    openai_client::OpenAiClient,
 16    parse_output::run_parse_output,
 17    paths::LLM_CACHE_DB,
 18    progress::{ExampleProgress, Step},
 19    word_diff::unified_to_word_diff,
 20};
 21use anyhow::{Context as _, Result};
 22use std::sync::OnceLock;
 23
 24const KEEP_PREVIOUS: &str = "KEEP_PREVIOUS";
 25
 26/// Print a summary report of repair results across all examples.
 27pub fn print_report(examples: &[Example], confidence_threshold: u8) {
 28    let total = examples.len();
 29    let mut no_repair_needed = 0;
 30    let mut repaired = 0;
 31    let mut repair_failed = 0;
 32
 33    for example in examples {
 34        if !needs_repair(example, confidence_threshold) {
 35            no_repair_needed += 1;
 36            continue;
 37        }
 38
 39        if has_successful_repair(example) {
 40            repaired += 1;
 41        } else {
 42            repair_failed += 1;
 43        }
 44    }
 45
 46    let needed_repair = total - no_repair_needed;
 47
 48    eprintln!();
 49    eprintln!("Repair summary ({total} examples):");
 50    eprintln!(
 51        "  {no_repair_needed}/{total} didn't need repair (confidence > {confidence_threshold})"
 52    );
 53    if needed_repair > 0 {
 54        eprintln!("  {needed_repair}/{total} needed repair:");
 55        if repaired > 0 {
 56            eprintln!("    {repaired} repaired successfully");
 57        }
 58        if repair_failed > 0 {
 59            eprintln!("    {repair_failed} failed to repair");
 60        }
 61    }
 62}
 63
 64/// Arguments for the repair command.
 65#[derive(Debug, Clone, clap::Args)]
 66pub struct RepairArgs {
 67    /// Use synchronous API instead of batch
 68    #[clap(long)]
 69    pub no_batch: bool,
 70
 71    /// Confidence threshold: repair predictions with confidence <= this value (1-5)
 72    #[clap(long, default_value = "2")]
 73    pub confidence_threshold: u8,
 74
 75    /// Which LLM provider to use (anthropic or openai)
 76    #[clap(long, default_value = "anthropic")]
 77    pub backend: BatchProvider,
 78}
 79
 80fn model_for_backend(backend: BatchProvider) -> &'static str {
 81    match backend {
 82        BatchProvider::Anthropic => "claude-sonnet-4-6",
 83        BatchProvider::Openai => "gpt-5.2",
 84    }
 85}
 86
 87/// Build the quality feedback string from QA results.
 88fn build_qa_feedback(example: &Example) -> Option<String> {
 89    let qa = example.qa.first()?.as_ref()?;
 90
 91    let qa_reasoning = qa.reasoning.as_deref().unwrap_or("No reasoning provided");
 92    let reverts_edits = qa
 93        .reverts_edits
 94        .map_or("unknown", |v| if v { "yes" } else { "no" });
 95    let confidence = qa
 96        .confidence
 97        .map_or("unknown".to_string(), |v| v.to_string());
 98
 99    Some(format!(
100        "- **Reverts user edits**: {reverts_edits}\n\
101         - **Confidence score**: {confidence}/5\n\
102         - **Reasoning**: {qa_reasoning}"
103    ))
104}
105
106/// Build the quality feedback string from computed scores when QA is unavailable.
107fn build_score_feedback(example: &Example) -> Option<String> {
108    let score = example.score.first()?;
109
110    let mut issues = Vec::new();
111
112    if score.reversal_ratio > 0.9 {
113        issues.push(format!(
114            "Automated analysis detected a high reversal ratio ({:.2}), which suggests this \
115             prediction may be reverting changes the user intentionally made. Double-check that \
116             the prediction doesn't undo the user's recent edits. If the prediction is actually \
117             fine and the edits are intentional completions rather than reversals, keep it as-is. \
118             If it truly reverts the user's changes, generate an improved prediction that \
119             continues the user's intent instead.",
120            score.reversal_ratio
121        ));
122    }
123
124    if score.wrong_editable_region == Some(true) {
125        issues.push(
126            "Automated analysis detected that the prediction may be modifying code outside \
127             the expected editable region, or producing changes misaligned with the editable \
128             region boundaries. Make sure the prediction only modifies code within the editable \
129             region and is properly aligned."
130                .to_string(),
131        );
132    }
133
134    if issues.is_empty() {
135        return None;
136    }
137
138    let mut feedback = String::from(
139        "No human quality assessment is available, but automated scoring flagged potential issues:\n\n",
140    );
141    for issue in &issues {
142        feedback.push_str(&format!("- {issue}\n"));
143    }
144    feedback.push_str(
145        "\nRemember: if the previous prediction was actually correct, output `KEEP_PREVIOUS`. \
146         If no edits should be made at all and you are unsure how to improve it, output `NO_EDITS`.",
147    );
148
149    Some(feedback)
150}
151
152/// Build the repair message (Turn 3) for a multi-turn conversation.
153///
154/// This message is sent after the original teacher prompt (Turn 1) and
155/// teacher response (Turn 2) to request an improved prediction.
156pub fn build_repair_message(example: &Example) -> Result<String> {
157    let prediction = example
158        .predictions
159        .first()
160        .context("no predictions available")?;
161    let actual_patch = prediction
162        .actual_patch
163        .as_ref()
164        .context("no actual_patch available (run predict first)")?;
165
166    let quality_feedback = build_qa_feedback(example)
167        .or_else(|| build_score_feedback(example))
168        .context("no quality feedback available (need either QA results or computed scores)")?;
169
170    let actual_patch_word_diff = unified_to_word_diff(actual_patch);
171
172    let token_counts = count_patch_token_changes(actual_patch);
173    let mut token_change_info = format!(
174        "\n## Token Change Statistics\n\n\
175         - **Deleted tokens**: {}\n\
176         - **Inserted tokens**: {}",
177        token_counts.deleted_tokens, token_counts.inserted_tokens,
178    );
179    if token_counts.deleted_tokens > 100 || token_counts.inserted_tokens > 100 {
180        token_change_info.push_str(
181            "\n\n> **Note:** The token change count is high. \
182             Consider producing a more scoped edit that targets only the lines \
183             that truly need to change, rather than rewriting large sections.",
184        );
185    }
186
187    let prompt_template = crate::prompt_assets::get_prompt("repair.md");
188    Ok(prompt_template
189        .replace("{actual_patch_word_diff}", &actual_patch_word_diff)
190        .replace("{quality_feedback}", &quality_feedback)
191        .replace("{token_change_info}", &token_change_info))
192}
193
194/// Check if an example needs repair based on QA feedback or computed scores.
195pub fn needs_repair(example: &Example, confidence_threshold: u8) -> bool {
196    // Check QA-based signals first.
197    if let Some(qa) = example.qa.first().and_then(|q| q.as_ref()) {
198        if qa.reverts_edits == Some(true) {
199            return true;
200        }
201
202        if let Some(confidence) = qa.confidence {
203            if confidence <= confidence_threshold {
204                return true;
205            }
206        }
207
208        return false;
209    }
210
211    // When QA is unavailable, fall back to computed score signals.
212    if let Some(score) = example.score.first() {
213        if score.reversal_ratio > 0.9 {
214            return true;
215        }
216
217        if score.wrong_editable_region == Some(true) {
218            return true;
219        }
220    }
221
222    false
223}
224
225/// Parse repair model output into a patch and optional cursor.
226///
227/// Handles the `KEEP_PREVIOUS` sentinel by copying the teacher's prediction,
228/// and delegates normal output to `TeacherPrompt::parse`.
229pub fn parse(example: &Example, actual_output: &str) -> Result<(String, Option<ActualCursor>)> {
230    if let Some(last_codeblock) = extract_last_codeblock(actual_output) {
231        if last_codeblock.trim() == KEEP_PREVIOUS {
232            let original = example
233                .predictions
234                .first()
235                .context("no original prediction to keep")?;
236            let patch = original.actual_patch.clone().unwrap_or_default();
237            let cursor = original.actual_cursor.clone();
238            return Ok((patch, cursor));
239        }
240    }
241
242    TeacherPrompt::parse(example, actual_output)
243}
244
245/// Check if an example already has a successful repair prediction.
246fn has_successful_repair(example: &Example) -> bool {
247    example
248        .predictions
249        .iter()
250        .any(|p| p.provider == PredictionProvider::Repair && p.actual_patch.is_some())
251}
252
253static ANTHROPIC_CLIENT_BATCH: OnceLock<AnthropicClient> = OnceLock::new();
254static ANTHROPIC_CLIENT_PLAIN: OnceLock<AnthropicClient> = OnceLock::new();
255static OPENAI_CLIENT_BATCH: OnceLock<OpenAiClient> = OnceLock::new();
256static OPENAI_CLIENT_PLAIN: OnceLock<OpenAiClient> = OnceLock::new();
257
258/// Run repair for a single example.
259///
260/// This sends a multi-turn conversation to the LLM:
261/// - Turn 1 (User): Original teacher prompt
262/// - Turn 2 (Assistant): Original teacher response
263/// - Turn 3 (User): Repair critique and instructions
264/// - Turn 4 (Assistant): Improved prediction (the response we parse)
265pub async fn run_repair(
266    example: &mut Example,
267    args: &RepairArgs,
268    example_progress: &ExampleProgress,
269) -> Result<()> {
270    if has_successful_repair(example) {
271        return Ok(());
272    }
273
274    if !needs_repair(example, args.confidence_threshold) {
275        return Ok(());
276    }
277
278    run_parse_output(example).context("Failed to execute run_parse_output")?;
279
280    if example.prompt_inputs.is_none() {
281        anyhow::bail!("prompt_inputs missing (run context retrieval first)");
282    }
283
284    if example.predictions.is_empty() {
285        anyhow::bail!("no predictions available (run predict first)");
286    }
287
288    let teacher_prompt = example
289        .prompt
290        .as_ref()
291        .context("prompt missing (run format_prompt first)")?;
292
293    let teacher_response = &example.predictions[0].actual_output;
294    if teacher_response.is_empty() {
295        anyhow::bail!("teacher response is empty (run predict first)");
296    }
297
298    let step_progress = example_progress.start(Step::Repair);
299
300    let model = model_for_backend(args.backend);
301    let repair_message = build_repair_message(example).context("Failed to build repair message")?;
302
303    step_progress.set_substatus("generating");
304
305    let response = match args.backend {
306        BatchProvider::Anthropic => {
307            let client = if args.no_batch {
308                ANTHROPIC_CLIENT_PLAIN.get_or_init(|| {
309                    AnthropicClient::plain().expect("Failed to create Anthropic client")
310                })
311            } else {
312                ANTHROPIC_CLIENT_BATCH.get_or_init(|| {
313                    AnthropicClient::batch(&LLM_CACHE_DB)
314                        .expect("Failed to create Anthropic client")
315                })
316            };
317
318            let messages = vec![
319                // Turn 1: Original teacher prompt
320                anthropic::Message {
321                    role: anthropic::Role::User,
322                    content: vec![anthropic::RequestContent::Text {
323                        text: teacher_prompt.input.clone(),
324                        cache_control: None,
325                    }],
326                },
327                // Turn 2: Original teacher response
328                anthropic::Message {
329                    role: anthropic::Role::Assistant,
330                    content: vec![anthropic::RequestContent::Text {
331                        text: teacher_response.clone(),
332                        cache_control: None,
333                    }],
334                },
335                // Turn 3: Repair critique and instructions
336                anthropic::Message {
337                    role: anthropic::Role::User,
338                    content: vec![anthropic::RequestContent::Text {
339                        text: repair_message,
340                        cache_control: None,
341                    }],
342                },
343            ];
344
345            let Some(response) = client.generate(model, 16384, messages, None, false).await? else {
346                return Ok(());
347            };
348
349            response
350                .content
351                .iter()
352                .filter_map(|c| match c {
353                    anthropic::ResponseContent::Text { text } => Some(text.as_str()),
354                    _ => None,
355                })
356                .collect::<Vec<_>>()
357                .join("")
358        }
359        BatchProvider::Openai => {
360            let client = if args.no_batch {
361                OPENAI_CLIENT_PLAIN
362                    .get_or_init(|| OpenAiClient::plain().expect("Failed to create OpenAI client"))
363            } else {
364                OPENAI_CLIENT_BATCH.get_or_init(|| {
365                    OpenAiClient::batch(&LLM_CACHE_DB).expect("Failed to create OpenAI client")
366                })
367            };
368
369            let messages = vec![
370                // Turn 1: Original teacher prompt
371                open_ai::RequestMessage::User {
372                    content: open_ai::MessageContent::Plain(teacher_prompt.input.clone()),
373                },
374                // Turn 2: Original teacher response
375                open_ai::RequestMessage::Assistant {
376                    content: Some(open_ai::MessageContent::Plain(teacher_response.clone())),
377                    tool_calls: vec![],
378                },
379                // Turn 3: Repair critique and instructions
380                open_ai::RequestMessage::User {
381                    content: open_ai::MessageContent::Plain(repair_message),
382                },
383            ];
384
385            let Some(response) = client.generate(model, 16384, messages, None, false).await? else {
386                return Ok(());
387            };
388
389            response
390                .choices
391                .into_iter()
392                .filter_map(|choice| match choice.message {
393                    open_ai::RequestMessage::Assistant { content, .. } => {
394                        content.map(|c| match c {
395                            open_ai::MessageContent::Plain(text) => text,
396                            open_ai::MessageContent::Multipart(parts) => parts
397                                .into_iter()
398                                .filter_map(|p| match p {
399                                    open_ai::MessagePart::Text { text } => Some(text),
400                                    _ => None,
401                                })
402                                .collect::<Vec<_>>()
403                                .join(""),
404                        })
405                    }
406                    _ => None,
407                })
408                .collect::<Vec<_>>()
409                .join("")
410        }
411    };
412
413    let parse_result = parse(example, &response);
414    let err = parse_result
415        .as_ref()
416        .err()
417        .map(|e| format!("Failed to parse repair response: {}", e));
418
419    let (actual_patch, actual_cursor) = parse_result.ok().unzip();
420    let actual_cursor = actual_cursor.flatten();
421
422    example.predictions.push(ExamplePrediction {
423        actual_patch,
424        actual_output: response,
425        actual_cursor,
426        error: err,
427        provider: PredictionProvider::Repair,
428    });
429
430    Ok(())
431}
432
433/// Sync batches for repair (upload pending requests, download finished results).
434pub async fn sync_batches(args: &RepairArgs) -> Result<()> {
435    if args.no_batch {
436        return Ok(());
437    }
438
439    match args.backend {
440        BatchProvider::Anthropic => {
441            let client = ANTHROPIC_CLIENT_BATCH.get_or_init(|| {
442                AnthropicClient::batch(&LLM_CACHE_DB).expect("Failed to create Anthropic client")
443            });
444            client.sync_batches().await?;
445        }
446        BatchProvider::Openai => {
447            let client = OPENAI_CLIENT_BATCH.get_or_init(|| {
448                OpenAiClient::batch(&LLM_CACHE_DB).expect("Failed to create OpenAI client")
449            });
450            client.sync_batches().await?;
451        }
452    }
453
454    Ok(())
455}