orch/

parser.rs

//! Agent response parser — normalizes JSON output from different agents.
//!
//! Each agent (Claude, Codex, OpenCode) returns a different JSON shape.
//! This module parses them all into a common `AgentResponse` struct.
//! Replaces `scripts/parse_response.sh` + `jq` pipelines.

use anyhow::Context;
use once_cell::sync::Lazy;
use regex::Regex;
use serde::{Deserialize, Serialize};
use std::io::Read;

static TRAILING_COMMA_RE: Lazy<Regex> =
    Lazy::new(|| Regex::new(r",(\s*[}\]])").expect("trailing comma regex is valid"));

/// Normalized agent response.
#[derive(Debug, Clone, Default, Serialize, Deserialize)]
pub struct AgentResponse {
    pub status: String,
    #[serde(default)]
    pub summary: String,
    #[serde(default)]
    pub accomplished: Vec<String>,
    #[serde(default)]
    pub remaining: Vec<String>,
    #[serde(default, alias = "files_changed", alias = "files_modified")]
    pub files: Vec<String>,
    #[serde(default)]
    pub error: Option<String>,
    /// Input token count (if available from agent).
    #[serde(default)]
    pub input_tokens: Option<u64>,
    /// Output token count (if available from agent).
    #[serde(default)]
    pub output_tokens: Option<u64>,
    /// Key learnings from this attempt (for memory persistence across retries).
    #[serde(default)]
    pub learnings: Vec<String>,
    /// Subtask delegations — agent requests child tasks to be created.
    #[serde(default)]
    pub delegations: Vec<Delegation>,
}

/// A subtask delegation requested by an agent.
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct Delegation {
    pub title: String,
    pub body: String,
    #[serde(default)]
    pub labels: Vec<String>,
}

/// Parse an agent response from a file path (or stdin if "-").
pub fn parse_and_print(path: &str) -> anyhow::Result<()> {
    let content = if path == "-" {
        let mut buf = String::new();
        std::io::stdin()
            .read_to_string(&mut buf)
            .context("reading stdin")?;
        buf
    } else {
        std::fs::read_to_string(path).with_context(|| format!("reading {path}"))?
    };

    let response = parse(&content)?;
    println!("{}", serde_json::to_string_pretty(&response)?);
    Ok(())
}

/// Known canonical statuses - used to prefer better candidates.
/// NOTE: Does NOT include aliases like "ready_for_review" - those are
/// normalized to canonical forms in normalize_status().
fn status_is_known(status: &str) -> bool {
    matches!(
        status,
        "new"
            | "routed"
            | "in_progress"
            | "done"
            | "completed"
            | "ok"
            | "success"
            | "running"
            | "blocked"
            | "error"
            | "failed"
            | "in_review"
            | "reviewing"
            | "needs_review"
            | "pending_review"
    )
}

/// Parse raw agent output into a normalized response.
pub fn parse(raw: &str) -> anyhow::Result<AgentResponse> {
    let mut last_err: Option<anyhow::Error> = None;
    let mut saw_jsonish_candidate = false;
    // Pair: (parsed response, raw candidate string) so we can check the raw object
    // for substantive fields that map_generic_response handles but that direct
    // deserialization into AgentResponse ignores (e.g. "output", "message").
    let mut best_candidate: Option<(AgentResponse, String)> = None;

    for candidate in json_candidates(raw) {
        match parse_candidate(&candidate) {
            Ok(mut resp) => {
                resp = normalize_status(resp);
                // If status is known, return immediately (good candidate found).
                if status_is_known(&resp.status) {
                    return Ok(resp);
                }
                // Non-canonical status - remember it but keep looking for better.
                if best_candidate.is_none() {
                    best_candidate = Some((resp, candidate));
                }
            }
            Err(err) => {
                if err.to_string() != "invalid agent response candidate" {
                    saw_jsonish_candidate = true;
                    last_err = Some(err);
                }
            }
        }
    }

    // If we found a candidate with non-canonical status but no known-status
    // candidate was found, use the best one we have — but only if it has at
    // least one substantive orch control field. Domain JSON records (career
    // pipeline entries, trading alerts, etc.) have a `status` field that can
    // be deserialized into AgentResponse, but they carry no orch fields like
    // `summary`, `accomplished`, `files`, etc. Returning them would silently
    // promote a domain value (e.g. "inbound-not-submitted") as the task status.
    if let Some((resp, raw_candidate)) = best_candidate {
        if has_substantive_fields(&raw_candidate) {
            return Ok(resp);
        }
    }

    if saw_jsonish_candidate {
        if let Some(err) = last_err {
            return Err(err);
        }
    }

    if let Some(err) = last_err {
        return Err(err);
    }

    anyhow::bail!(
        "agent output is not valid JSON or a supported JSON wrapper: {}",
        raw.trim()
    )
}

/// Map known non-canonical agent statuses to their canonical equivalents.
///
/// Agents sometimes return statuses like `"ok"`, `"success"`, or `"completed"`
/// that are semantically equivalent to `"done"`. Normalizing them here ensures
/// that `classify_run_outcome` correctly classifies the run as success and
/// prevents opaque `"no error info available"` failures when an agent returns
/// a valid but non-canonical status.
///
/// Unknown statuses that read as completions normalize to done via the shared
/// heuristic in `status_heuristics`; the rest stay as-is for the audit trail.
fn normalize_status(mut resp: AgentResponse) -> AgentResponse {
    let normalized = resp.status.to_ascii_lowercase();
    if normalized.starts_with("pushed to ") || normalized.starts_with("pushed ") {
        resp.status = "done".to_string();
        return resp;
    }

    resp.status = match normalized.as_str() {
        // Canonical completion statuses.
        // Some agents return `complete` or `no_changes_needed` to indicate
        // a successful run with no further edits required — treat them as
        // `done` so the run is classified as successful.
        "done"
        | "completed"
        | "complete"
        | "skipped"
        | "skip"
        | "ok"
        | "success"
        | "noop"
        | "green"
        | "no_changes_needed"
        | "no_trades_no_positions"
        | "no_setups"
        | "changes_made"
        | "change_made"
        | "changes_pushed"
        | "pushed"
        | "acknowledged"
        | "flat"
        | "no_action"
        | "no_action_needed"
        | "missed"
        | "changes_addressed"
        | "changes_requested_addressed"
        | "review_addressed"
        | "already_finalized_in_attempt_1"
        | "alert"
        | "alerts"
        | "alerts_sent"
        | "not_configured"
        | "healthy"
        | "alerts_fired"
        | "verified"
        | "fixed" => "done".to_string(),
        // Canonical progress statuses.
        // `partial` is used by some models to indicate partial progress —
        // treat it as in_progress so the task remains open for follow-up.
        "in_progress" | "running" | "partial" | "waiting" => "in_progress".to_string(),
        "in_review" | "reviewing" => "in_review".to_string(),
        "needs_review" | "pending_review" | "ready_for_review" => "needs_review".to_string(),
        // Canonical error statuses.
        "blocked" | "error" | "failed" | "failure" => "blocked".to_string(),
        // Passthrough: known canonical statuses (already canonical, keep as-is).
        s if s == "new"
            || s == "routed"
            || s == "in_progress"
            || s == "done"
            || s == "blocked"
            || s == "in_review"
            || s == "needs_review" =>
        {
            s.to_string()
        }
        // Unknown status reading as completion, done if the response carries orch
        // fields — domain records without them stay non-canonical and get rejected (#3336)
        _ if !resp.summary.is_empty()
            || !resp.accomplished.is_empty()
            || !resp.files.is_empty()
            || resp.error.is_some() =>
        {
            if crate::status_heuristics::status_looks_like_descriptive_completion(&normalized) {
                "done".to_string()
            } else {
                resp.status.clone()
            }
        }
        _ => resp.status.clone(),
    };
    resp
}

/// Extract the first JSON code block from markdown.
fn extract_json_block(text: &str) -> Option<String> {
    let start = text.find("```json")?;
    let fence_end = start + "```json".len();
    let remainder = &text[fence_end..];

    // Find key positions after the opening fence.
    let newline_pos = remainder.find('\n');
    let json_start_pos = remainder.find(['{', '[']);
    let first_nl = newline_pos.unwrap_or(usize::MAX);

    // Decide fenced vs inline:
    // - If a JSON delimiter ('{' or '[') appears before the first newline, treat as inline.
    // - If there is no newline at all, treat as inline.
    // - Otherwise, treat as a standard fenced block.
    let is_inline = json_start_pos.is_some_and(|js| js < first_nl) || newline_pos.is_none();

    if is_inline {
        // Inline: closing fence may appear anywhere on the same logical line.
        // Using find("```") here is safe because the JSON start already appeared
        // before any newline, so the closing fence is on the same line too.
        let closing_fence_pos = remainder.find("```")?;
        let content_start = remainder
            .char_indices()
            .find(|(_, ch)| !ch.is_whitespace())
            .map_or(fence_end, |(idx, _)| fence_end + idx);
        let end = fence_end + closing_fence_pos;
        if content_start > end {
            return None;
        }
        Some(text[content_start..end].to_string())
    } else {
        // Fenced block: content begins after the first newline.
        // The closing fence must be anchored to a line boundary (\n```) to avoid
        // false positives from triple-backticks inside JSON string values.
        let nl = newline_pos?;
        let content_start = fence_end + nl + 1;
        let content = &text[content_start..];
        let closing_pos = find_closing_fence_at_line_boundary(content)?;
        Some(text[content_start..content_start + closing_pos].to_string())
    }
}

/// Find the position (within `content`) where a ``` closing fence starts, anchored to a
/// line boundary.  Returns the byte offset of the leading ``` so that `content[..pos]`
/// is everything before the closing fence (including the preceding newline).
fn find_closing_fence_at_line_boundary(content: &str) -> Option<usize> {
    // Edge case: content itself begins with the closing fence (no preceding content).
    if content.starts_with("```") {
        return Some(0);
    }
    // Scan for \n``` pattern — the closing fence must start at the beginning of a line.
    let mut search_from = 0;
    while let Some(nl_idx) = content[search_from..].find('\n') {
        let after_nl = search_from + nl_idx + 1;
        if content[after_nl..].starts_with("```") {
            // Return the position of the ``` (after the \n).
            // The content slice up to this point includes the \n, matching Markdown semantics.
            return Some(after_nl);
        }
        search_from = after_nl;
    }
    None
}

/// Returns true if the raw JSON candidate has at least one substantive orch control field.
///
/// Operates on the raw string (not the deserialized `AgentResponse`) so that fields like
/// `output` and `message` — which `map_generic_response` maps to `summary` but which are
/// not struct fields on `AgentResponse` and are silently ignored by direct deserialization —
/// are correctly treated as substantive. This prevents domain JSON records (career pipeline
/// entries, trading alerts, etc.) whose only orch-like field is `status` from being promoted
/// as the task status.
fn has_substantive_fields(raw: &str) -> bool {
    if let Ok(serde_json::Value::Object(obj)) = serde_json::from_str::<serde_json::Value>(raw) {
        return SUBSTANTIVE_RESPONSE_FIELDS
            .iter()
            .any(|key| obj.contains_key(*key))
            || obj.contains_key("error");
    }
    false
}

/// Fields that indicate a JSON blob is an AgentResponse (higher score = better match).
const AGENT_RESPONSE_FIELDS: &[&str] = &[
    "\"status\"",
    "\"summary\"",
    "\"accomplished\"",
    "\"remaining\"",
    "\"files\"",
    "\"learnings\"",
    "\"delegations\"",
    "\"error\"",
];

fn agent_response_score(blob: &str) -> usize {
    AGENT_RESPONSE_FIELDS
        .iter()
        .filter(|&&f| blob.contains(f))
        .count()
}

fn json_candidates(raw: &str) -> Vec<String> {
    let mut candidates = Vec::new();

    if let Some(block) = extract_json_block(raw) {
        candidates.push(block);
    }

    // Collect ALL balanced JSON blobs and sort by how much they look like an
    // AgentResponse, so the real response is tried before telemetry blobs.
    let mut all_json = extract_all_balanced_json(raw);
    all_json.sort_by_key(|b| std::cmp::Reverse(agent_response_score(b)));

    for candidate in all_json {
        if !candidates.iter().any(|existing| existing == &candidate) {
            candidates.push(candidate);
        }
    }

    candidates.push(raw.to_string());
    candidates
}

fn parse_candidate(candidate: &str) -> anyhow::Result<AgentResponse> {
    if let Ok(resp) = serde_json::from_str::<AgentResponse>(candidate) {
        return Ok(resp);
    }

    let repaired = repair_json_like(candidate);
    if repaired != candidate {
        if let Ok(resp) = serde_json::from_str::<AgentResponse>(&repaired) {
            return Ok(resp);
        }
    }

    if let Ok(val) = serde_json::from_str::<serde_json::Value>(candidate) {
        return map_generic_response(&val);
    }

    if repaired != candidate {
        if let Ok(val) = serde_json::from_str::<serde_json::Value>(&repaired) {
            return map_generic_response(&val);
        }
    }

    anyhow::bail!("invalid agent response candidate")
}

/// Scan `text` and return every top-level balanced JSON object or array found.
fn extract_all_balanced_json(text: &str) -> Vec<String> {
    let mut results = Vec::new();
    let mut pos = 0;

    while pos < text.len() {
        // Find the next `{` or `[` starting from `pos`.
        let Some(start_offset) = text[pos..]
            .char_indices()
            .find(|(_, ch)| matches!(ch, '{' | '['))
            .map(|(idx, _)| idx)
        else {
            break;
        };
        let start = pos + start_offset;

        let mut stack = Vec::new();
        let mut in_string = false;
        let mut escape = false;
        let mut end: Option<usize> = None;

        for (idx, ch) in text[start..].char_indices() {
            let abs = start + idx;
            if in_string {
                if escape {
                    escape = false;
                    continue;
                }
                match ch {
                    '\\' => escape = true,
                    '"' => in_string = false,
                    _ => {}
                }
                continue;
            }

            match ch {
                '"' => in_string = true,
                '{' | '[' => stack.push(ch),
                '}' => {
                    if stack.pop() != Some('{') {
                        break; // malformed — skip past this `{`
                    }
                    if stack.is_empty() {
                        end = Some(abs);
                        break;
                    }
                }
                ']' => {
                    if stack.pop() != Some('[') {
                        break; // malformed — skip past this `[`
                    }
                    if stack.is_empty() {
                        end = Some(abs);
                        break;
                    }
                }
                _ => {}
            }
        }

        if let Some(end_idx) = end {
            results.push(text[start..=end_idx].to_string());
            pos = end_idx + 1;
        } else {
            // No closing brace found — skip past the opening character.
            pos = start + 1;
        }
    }

    results
}

fn repair_json_like(text: &str) -> String {
    let mut repaired = text.trim().replace(['“', '”'], "\"");
    repaired = repaired.replace(['‘', '’'], "'");
    repaired = TRAILING_COMMA_RE.replace_all(&repaired, "$1").into_owned();

    let mut stack = Vec::new();
    let mut in_string = false;
    let mut escape = false;

    for ch in repaired.chars() {
        if in_string {
            if escape {
                escape = false;
                continue;
            }
            match ch {
                '\\' => escape = true,
                '"' => in_string = false,
                _ => {}
            }
            continue;
        }

        match ch {
            '"' => in_string = true,
            '{' => stack.push('}'),
            '[' => stack.push(']'),
            '}' | ']' => {
                if let Some(expected) = stack.last().copied() {
                    if expected == ch {
                        stack.pop();
                    }
                }
            }
            _ => {}
        }
    }

    while let Some(ch) = stack.pop() {
        repaired.push(ch);
    }

    repaired
}

/// Fields that indicate a substantive agent response (not just a status blob).
/// NOTE: We don't include "output" or "message" here because those are
/// handled via summary extraction in map_generic_response. This list is for
/// validating that the payload has meaningful agent response structure beyond
/// just a status field.
const SUBSTANTIVE_RESPONSE_FIELDS: &[&str] = &[
    "summary",
    "message",
    "accomplished",
    "remaining",
    "files",
    "files_changed",
    "files_modified",
    "learnings",
    "delegations",
    "output",
    "pr_url",
];

/// Infer a canonical completion status from known non-standard completion envelope shapes.
///
/// Claude agents sometimes return JSON without a `status`/`result` field, using
/// alternative completion indicators instead. This function maps those shapes to
/// a canonical status so they are not misclassified as `InvalidResponse`.
///
/// Recognized patterns (all map to `"done"`):
/// - `{"success": true, ...}` — explicit success boolean
/// - `{"type": "task_complete", ...}` — explicit task_complete type string
/// - `{"tests_passed": true, ...}` — tests-passed completion signal
fn infer_completion_status(obj: &serde_json::Map<String, serde_json::Value>) -> Option<String> {
    // success: true
    if obj.get("success").and_then(|v| v.as_bool()) == Some(true) {
        return Some("done".to_string());
    }
    // type: "task_complete"
    if obj.get("type").and_then(|v| v.as_str()) == Some("task_complete") {
        return Some("done".to_string());
    }
    // tests_passed: true
    if obj.get("tests_passed").and_then(|v| v.as_bool()) == Some(true) {
        return Some("done".to_string());
    }
    None
}

/// Map a generic JSON object to AgentResponse.
///
/// Requires stricter shape checking: the object must have:
/// 1. A status/result field (or a recognized completion envelope shape)
/// 2. AND at least one substantive field OR an error field
///
/// This prevents acceptance of non-response JSON blobs (telemetry, tool outputs)
/// that happen to contain a status field.
fn map_generic_response(val: &serde_json::Value) -> anyhow::Result<AgentResponse> {
    let obj = val.as_object().context("expected JSON object")?;

    // Check for status or result field first; fall back to known completion envelopes.
    let status = obj
        .get("status")
        .or_else(|| obj.get("result"))
        .and_then(|v| v.as_str())
        .map(String::from)
        .or_else(|| infer_completion_status(obj))
        .context("agent response is missing required 'status' or 'result' field")?;

    // Require at least one substantive field OR an error field.
    // This prevents accepting telemetry/event JSON blobs that happen to have
    // a status field but are not actual agent responses.
    let has_substantive_field = SUBSTANTIVE_RESPONSE_FIELDS
        .iter()
        .any(|key| obj.contains_key(*key));
    let has_error = obj.contains_key("error");

    if !has_substantive_field && !has_error {
        anyhow::bail!("JSON object lacks substantive response fields");
    }

    let summary = obj
        .get("summary")
        .or_else(|| obj.get("message"))
        .or_else(|| obj.get("output"))
        .and_then(|v| v.as_str())
        .unwrap_or("")
        .to_string();

    let accomplished = extract_string_array(obj.get("accomplished"));
    let remaining = extract_string_array(obj.get("remaining"));
    let files = {
        let f = extract_string_array(obj.get("files"));
        if !f.is_empty() {
            f
        } else {
            let fc = extract_string_array(obj.get("files_changed"));
            if !fc.is_empty() {
                fc
            } else {
                extract_string_array(obj.get("files_modified"))
            }
        }
    };
    let error = obj.get("error").and_then(|v| v.as_str()).map(String::from);
    let learnings = extract_string_array(obj.get("learnings"));

    // Extract token counts if available
    let input_tokens = extract_u64(obj.get("input_tokens"))
        .or_else(|| extract_u64(obj.get("tokens_input")))
        .or_else(|| extract_usage_tokens(obj.get("usage"), true));
    let output_tokens = extract_u64(obj.get("output_tokens"))
        .or_else(|| extract_u64(obj.get("tokens_output")))
        .or_else(|| extract_usage_tokens(obj.get("usage"), false));

    // Extract delegations. If the `delegations` field is present but malformed,
    // return an error instead of silently ignoring it so callers can surface
    // and debug malformed agent output.
    let delegations = if let Some(v) = obj.get("delegations") {
        serde_json::from_value::<Vec<Delegation>>(v.clone())
            .with_context(|| format!("invalid delegations field: {}", v))?
    } else {
        Vec::new()
    };

    Ok(AgentResponse {
        status,
        summary,
        accomplished,
        remaining,
        files,
        error,
        input_tokens,
        output_tokens,
        learnings,
        delegations,
    })
}

/// Extract a u64 from a JSON value.
fn extract_u64(val: Option<&serde_json::Value>) -> Option<u64> {
    val.and_then(|v| v.as_u64())
}

/// Extract tokens from a usage object (common in OpenAI-compatible APIs).
fn extract_usage_tokens(usage: Option<&serde_json::Value>, is_input: bool) -> Option<u64> {
    let obj = usage?.as_object()?;
    let key = if is_input {
        "input_tokens"
    } else {
        "output_tokens"
    };
    obj.get(key).and_then(|v| v.as_u64())
}

fn extract_string_array(val: Option<&serde_json::Value>) -> Vec<String> {
    val.and_then(|v| v.as_array())
        .map(|arr| {
            arr.iter()
                .filter_map(|v| v.as_str().map(String::from))
                .collect()
        })
        .unwrap_or_default()
}

#[cfg(test)]
mod tests {
    use super::*;

    #[test]
    fn parse_direct_json() {
        let input = r#"{"status":"done","summary":"Fixed bug","accomplished":["fixed it"],"remaining":[],"files":["src/main.rs"]}"#;
        let resp = parse(input).unwrap();
        assert_eq!(resp.status, "done");
        assert_eq!(resp.summary, "Fixed bug");
        assert_eq!(resp.accomplished, vec!["fixed it"]);
        assert_eq!(resp.files, vec!["src/main.rs"]);
        assert!(resp.error.is_none());
    }

    #[test]
    fn parse_json_in_markdown_block() {
        let input = r#"Here is the result:

```json
{"status":"in_progress","summary":"Working on it","accomplished":[],"remaining":["finish tests"],"files":[]}
```

Done.
"#;
        let resp = parse(input).unwrap();
        assert_eq!(resp.status, "in_progress");
        assert_eq!(resp.remaining, vec!["finish tests"]);
    }

    #[test]
    fn parse_generic_json_with_different_field_names() {
        let input = r#"{"result":"done","message":"All good","files":["a.rs","b.rs"]}"#;
        let resp = parse(input).unwrap();
        assert_eq!(resp.status, "done");
        assert_eq!(resp.summary, "All good");
        assert_eq!(resp.files, vec!["a.rs", "b.rs"]);
    }

    #[test]
    fn parse_with_error_field() {
        let input = r#"{"status":"blocked","summary":"Cannot proceed","accomplished":[],"remaining":[],"files":[],"error":"missing dependency"}"#;
        let resp = parse(input).unwrap();
        assert_eq!(resp.status, "blocked");
        assert_eq!(resp.error, Some("missing dependency".to_string()));
    }

    #[test]
    fn parse_fallback_raw_text() {
        let input = "This is just plain text output from the agent.";
        let err = parse(input).unwrap_err();
        assert!(err
            .to_string()
            .contains("agent output is not valid JSON or a supported JSON wrapper"));
    }

    #[test]
    fn parse_malformed_json_block_returns_error() {
        let input = r#"Here is the result:

```json
{"status":"done","summary":"missing brace"
```
"#;
        let resp = parse(input).unwrap();
        assert_eq!(resp.status, "done");
        assert_eq!(resp.summary, "missing brace");
    }

    #[test]
    fn parse_json_with_trailing_comma() {
        let input =
            r#"{"status":"done","summary":"Fixed","accomplished":[],"remaining":[],"files":[],}"#;
        let resp = parse(input).unwrap();
        assert_eq!(resp.status, "done");
        assert_eq!(resp.summary, "Fixed");
    }

    #[test]
    fn parse_json_embedded_in_commentary() {
        let input = r#"Done:
{"status":"in_progress","summary":"working","accomplished":[],"remaining":["tests"],"files":[]}
Thanks"#;
        let resp = parse(input).unwrap();
        assert_eq!(resp.status, "in_progress");
        assert_eq!(resp.remaining, vec!["tests"]);
    }

    #[test]
    fn extract_json_block_from_markdown() {
        let text = "prefix\n```json\n{\"key\":\"value\"}\n```\nsuffix";
        let block = extract_json_block(text).unwrap();
        assert_eq!(block, "{\"key\":\"value\"}\n");
    }

    #[test]
    fn extract_json_block_inline_fence() {
        let text = "prefix\n```json{\"key\":\"value\"}```\nsuffix";
        let block = extract_json_block(text).unwrap();
        assert_eq!(block, "{\"key\":\"value\"}");
    }

    #[test]
    fn extract_json_block_with_info_string() {
        let text = "prefix\n```json some-meta\n{\"key\":\"value\"}\n```\nsuffix";
        let block = extract_json_block(text).unwrap();
        assert_eq!(block, "{\"key\":\"value\"}\n");
    }

    #[test]
    fn extract_json_block_with_info_string_and_whitespace() {
        let text = "prefix\n```json   some-meta   \n{\"key\":\"value\"}\n```\nsuffix";
        let block = extract_json_block(text).unwrap();
        assert_eq!(block, "{\"key\":\"value\"}\n");
    }

    #[test]
    fn extract_json_block_inline_with_internal_newline() {
        // Inline JSON that contains a newline inside the JSON value
        let text = "prefix\n```json{\"key\":\"value\\nwith newline\"}```\nsuffix";
        let block = extract_json_block(text).unwrap();
        assert_eq!(block, r#"{"key":"value\nwith newline"}"#);
    }

    #[test]
    fn extract_json_block_inline_with_whitespace_newline() {
        // Inline JSON that contains a newline as whitespace between key and value
        let text = "prefix\n```json{\"key\":\n\"value\"}```\nsuffix";
        let block = extract_json_block(text).unwrap();
        assert_eq!(block, "{\"key\":\n\"value\"}");
    }

    #[test]
    fn extract_json_block_missing_closing_fence() {
        let text = "prefix\n```json{\"key\":\"value\"}\n";
        assert!(extract_json_block(text).is_none());
    }

    #[test]
    fn extract_json_block_array_inline() {
        let text = "prefix\n```json[1,2,3]```\nsuffix";
        let block = extract_json_block(text).unwrap();
        assert_eq!(block, "[1,2,3]");
    }

    #[test]
    fn extract_json_block_multiple_fences() {
        let text = "prefix\n```json{\"first\":1}``` middle\n```json{\"second\":2}``` suffix";
        let block = extract_json_block(text).unwrap();
        assert_eq!(block, "{\"first\":1}");
    }

    #[test]
    fn extract_json_block_missing_returns_none() {
        assert!(extract_json_block("no code block here").is_none());
    }

    #[test]
    fn extract_string_array_from_json() {
        let val: serde_json::Value = serde_json::json!(["a", "b", "c"]);
        let result = extract_string_array(Some(&val));
        assert_eq!(result, vec!["a", "b", "c"]);
    }

    #[test]
    fn extract_string_array_none() {
        let result = extract_string_array(None);
        assert!(result.is_empty());
    }

    #[test]
    fn parse_empty_object() {
        let input = "{}";
        let err = parse(input).unwrap_err();
        let msg = err.to_string();
        assert!(
            msg.contains("agent response is missing required")
                || msg.contains("JSON object lacks substantive"),
            "unexpected error: {msg}"
        );
    }

    #[test]
    fn parse_with_learnings() {
        let input = r#"{"status":"needs_review","summary":"Need to fix imports","accomplished":[],"remaining":["fix imports"],"files":["src/main.rs"],"learnings":["Use std::sync::Arc for shared state","Check imports before committing"]}"#;
        let resp = parse(input).unwrap();
        assert_eq!(resp.status, "needs_review");
        assert_eq!(resp.summary, "Need to fix imports");
        assert_eq!(resp.learnings.len(), 2);
        assert_eq!(resp.learnings[0], "Use std::sync::Arc for shared state");
        assert_eq!(resp.learnings[1], "Check imports before committing");
    }

    #[test]
    fn parse_learnings_empty_by_default() {
        let input = r#"{"status":"done","summary":"Completed","accomplished":["done"],"remaining":[],"files":[]}"#;
        let resp = parse(input).unwrap();
        assert!(resp.learnings.is_empty());
    }

    #[test]
    fn parse_with_delegations() {
        let input = r#"{
            "status": "blocked",
            "summary": "Task requires subtasks",
            "accomplished": ["Analyzed requirements"],
            "remaining": ["Waiting on subtasks"],
            "files_changed": [],
            "blockers": ["Waiting on delegated subtasks"],
            "reason": "Decomposed into subtasks",
            "delegations": [
                {"title": "Implement login API", "body": "Create POST /api/login", "labels": ["backend"]},
                {"title": "Add login form", "body": "Create React login component", "labels": ["frontend"]}
            ]
        }"#;
        let resp = parse(input).unwrap();
        assert_eq!(resp.status, "blocked");
        assert_eq!(resp.delegations.len(), 2);
        assert_eq!(resp.delegations[0].title, "Implement login API");
        assert_eq!(resp.delegations[0].body, "Create POST /api/login");
        assert_eq!(resp.delegations[0].labels, vec!["backend"]);
        assert_eq!(resp.delegations[1].title, "Add login form");
        assert_eq!(resp.delegations[1].labels, vec!["frontend"]);
    }

    #[test]
    fn parse_delegations_empty_by_default() {
        let input =
            r#"{"status":"done","summary":"Done","accomplished":[],"remaining":[],"files":[]}"#;
        let resp = parse(input).unwrap();
        assert!(resp.delegations.is_empty());
    }

    #[test]
    fn parse_delegations_without_labels() {
        let input = r#"{
            "status": "blocked",
            "summary": "Delegating",
            "delegations": [
                {"title": "Subtask", "body": "Do the thing"}
            ]
        }"#;
        let resp = parse(input).unwrap();
        assert_eq!(resp.delegations.len(), 1);
        assert_eq!(resp.delegations[0].title, "Subtask");
        assert!(resp.delegations[0].labels.is_empty());
    }

    #[test]
    fn parse_response_after_telemetry_json() {
        // Telemetry JSON appears before the actual response JSON.
        // The parser must skip the telemetry blob and find the real response.
        let input = r#"Processing...
{"type":"telemetry","message":"starting task","tokens":42}
Some output here.
{"status":"done","summary":"Fixed the bug","accomplished":["patched parser"],"remaining":[],"files":["src/parser.rs"]}
"#;
        let resp = parse(input).unwrap();
        assert_eq!(resp.status, "done");
        assert_eq!(resp.summary, "Fixed the bug");
        assert_eq!(resp.files, vec!["src/parser.rs"]);
    }

    // --- Edge-case tests for line-boundary fence extraction ---

    #[test]
    fn extract_json_block_backticks_inside_json_string() {
        // Triple-backticks appear inside a JSON string value.
        // The old code would stop at the first ``` (inside the string).
        // The new code must find the real closing fence at a line boundary.
        let text = "```json\n{\"code\":\"example: ```snippet``` end\"}\n```\n";
        let block = extract_json_block(text).unwrap();
        assert_eq!(block, "{\"code\":\"example: ```snippet``` end\"}\n");
    }

    #[test]
    fn extract_json_block_backticks_inside_json_string_parse_roundtrip() {
        // Full parse roundtrip: fenced block with backticks inside a JSON string value.
        let input = "```json\n{\"status\":\"done\",\"summary\":\"use ```code``` here\"}\n```\n";
        let resp = parse(input).unwrap();
        assert_eq!(resp.status, "done");
        assert_eq!(resp.summary, "use ```code``` here");
    }

    #[test]
    fn extract_json_block_inline_with_embedded_backtick_pairs() {
        // Inline JSON: embedded backticks that are NOT triple should not confuse the parser.
        let text = "```json{\"key\":\"a`b``c\"}```\n";
        let block = extract_json_block(text).unwrap();
        assert_eq!(block, "{\"key\":\"a`b``c\"}");
    }

    #[test]
    fn parse_mixed_telemetry_and_fenced_response() {
        // Telemetry JSON with backtick sequences followed by a fenced agent response.
        let input = r#"Processing...
{"type":"log","msg":"calling tool: ```bash```"}
```json
{"status":"done","summary":"fixed it","accomplished":["patched"],"remaining":[],"files":["src/x.rs"]}
```
"#;
        let resp = parse(input).unwrap();
        assert_eq!(resp.status, "done");
        assert_eq!(resp.summary, "fixed it");
        assert_eq!(resp.files, vec!["src/x.rs"]);
    }

    #[test]
    fn parse_response_from_ndjson_multiple_objects() {
        // NDJSON stream: only the last object has `status` — must select it.
        let input = r#"{"type":"text","text":"Working on it..."}
{"type":"tool_call","name":"bash","input":{"command":"cargo test"}}
{"type":"tool_result","output":"test passed"}
{"status":"needs_review","summary":"All tests pass","accomplished":["fixed bug"],"remaining":[],"files":["src/lib.rs"]}
"#;
        let resp = parse(input).unwrap();
        assert_eq!(resp.status, "needs_review");
        assert_eq!(resp.summary, "All tests pass");
        assert_eq!(resp.files, vec!["src/lib.rs"]);
    }

    // ── normalize_status tests ────────────────────────────────────────────────

    #[test]
    fn parse_normalizes_ok_alias_to_done() {
        let input =
            r#"{"status":"ok","summary":"all good","accomplished":[],"remaining":[],"files":[]}"#;
        let resp = parse(input).unwrap();
        assert_eq!(resp.status, "done");
    }

    #[test]
    fn parse_normalizes_success_alias_to_done() {
        let input = r#"{"status":"success","summary":"fixed it","accomplished":[],"remaining":[],"files":[]}"#;
        let resp = parse(input).unwrap();
        assert_eq!(resp.status, "done");
    }

    #[test]
    fn parse_normalizes_completed_alias_to_done() {
        let input = r#"{"status":"completed","summary":"done","accomplished":[],"remaining":[],"files":[]}"#;
        let resp = parse(input).unwrap();
        assert_eq!(resp.status, "done");
    }

    #[test]
    fn parse_normalizes_missed_alias_to_done_case_insensitive() {
        let input = r#"{"status":"MISSED","summary":"no setups found","accomplished":[],"remaining":[],"files":[]}"#;
        let resp = parse(input).unwrap();
        assert_eq!(resp.status, "done");
    }

    #[test]
    fn parse_normalizes_changes_addressed_alias_to_done() {
        let input = r#"{"status":"changes_addressed","summary":"feedback addressed","accomplished":[],"remaining":[],"files":[]}"#;
        let resp = parse(input).unwrap();
        assert_eq!(resp.status, "done");
    }

    #[test]
    fn parse_normalizes_changes_requested_addressed_alias_to_done() {
        let input = r#"{"status":"changes_requested_addressed","summary":"review feedback addressed","accomplished":[],"remaining":[],"files":[]}"#;
        let resp = parse(input).unwrap();
        assert_eq!(resp.status, "done");
    }

    #[test]
    fn parse_normalizes_review_addressed_alias_to_done() {
        let input = r#"{"status":"review_addressed","summary":"review feedback addressed","accomplished":[],"remaining":[],"files":[]}"#;
        let resp = parse(input).unwrap();
        assert_eq!(resp.status, "done");
    }

    #[test]
    fn parse_normalizes_already_finalized_in_attempt_1_alias_to_done() {
        let input = r#"{"status":"already_finalized_in_attempt_1","summary":"work already done in attempt 1","accomplished":[],"remaining":[],"files":[]}"#;
        let resp = parse(input).unwrap();
        assert_eq!(resp.status, "done");
    }

    #[test]
    fn parse_normalizes_noop_alias_to_done() {
        // noop / green / alerts_sent → done, matching #3273 alias map.
        let input = r#"{"status":"noop","summary":"nothing to do","accomplished":[],"remaining":[],"files":[]}"#;
        let resp = parse(input).unwrap();
        assert_eq!(resp.status, "done");
    }

    #[test]
    fn parse_normalizes_green_alias_to_done() {
        let input = r#"{"status":"green","summary":"all passing","accomplished":[],"remaining":[],"files":[]}"#;
        let resp = parse(input).unwrap();
        assert_eq!(resp.status, "done");
    }

    #[test]
    fn parse_normalizes_alerts_sent_alias_to_done() {
        let input = r#"{"status":"alerts_sent","summary":"notifications dispatched","accomplished":[],"remaining":[],"files":[]}"#;
        let resp = parse(input).unwrap();
        assert_eq!(resp.status, "done");
    }

    #[test]
    fn parse_normalizes_verified_alias_to_done() {
        let input = r#"{"status":"verified","summary":"result verified correct","accomplished":[],"remaining":[],"files":[]}"#;
        let resp = parse(input).unwrap();
        assert_eq!(resp.status, "done");
    }

    #[test]
    fn parse_normalizes_fixed_alias_to_done() {
        let input = r#"{"status":"fixed","summary":"bug has been fixed","accomplished":["fixed bug"],"remaining":[],"files":["src/parser.rs"]}"#;
        let resp = parse(input).unwrap();
        assert_eq!(resp.status, "done");
    }

    #[test]
    fn parse_normalizes_alerts_fired_alias_to_done() {
        let input = r#"{"status":"alerts_fired","summary":"alerts dispatched","accomplished":[],"remaining":[],"files":[]}"#;
        let resp = parse(input).unwrap();
        assert_eq!(resp.status, "done");
    }

    #[test]
    fn parse_normalizes_waiting_alias_to_in_progress() {
        // waiting → in_progress (distinct from running/partial in the alias map).
        let input = r#"{"status":"waiting","summary":"blocked on upstream","accomplished":[],"remaining":[],"files":[]}"#;
        let resp = parse(input).unwrap();
        assert_eq!(resp.status, "in_progress");
    }

    #[test]
    fn parse_normalizes_changes_pushed_alias_to_done() {
        let input = r#"{"status":"changes_pushed","summary":"changes pushed to branch","accomplished":[],"remaining":[],"files":[]}"#;
        let resp = parse(input).unwrap();
        assert_eq!(resp.status, "done");
    }

    #[test]
    fn parse_normalizes_bare_pushed_alias_to_done() {
        let input = r#"{"status":"pushed","summary":"changes pushed to branch","accomplished":[],"remaining":[],"files":[]}"#;
        let resp = parse(input).unwrap();
        assert_eq!(resp.status, "done");
    }

    #[test]
    fn parse_normalizes_dynamic_pushed_to_branch_status_to_done() {
        let input = r#"{"status":"pushed to internal-155195-weekly-review-patterns-progress-and-next","summary":"changes pushed to branch","accomplished":[],"remaining":[],"files":[]}"#;
        let resp = parse(input).unwrap();
        assert_eq!(resp.status, "done");
    }

    #[test]
    fn parse_normalizes_running_alias_to_in_progress() {
        let input = r#"{"status":"running","summary":"working","accomplished":[],"remaining":[],"files":[]}"#;
        let resp = parse(input).unwrap();
        assert_eq!(resp.status, "in_progress");
    }

    #[test]
    fn parse_normalizes_reviewing_alias_to_in_review() {
        let input = r#"{"status":"reviewing","summary":"reviewing","accomplished":[],"remaining":[],"files":[]}"#;
        let resp = parse(input).unwrap();
        assert_eq!(resp.status, "in_review");
    }

    #[test]
    fn parse_normalizes_pending_review_alias_to_needs_review() {
        let input = r#"{"status":"pending_review","summary":"ready","accomplished":[],"remaining":[],"files":[]}"#;
        let resp = parse(input).unwrap();
        assert_eq!(resp.status, "needs_review");
    }

    #[test]
    fn parse_normalizes_error_alias_to_blocked() {
        let input =
            r#"{"status":"error","summary":"failed","accomplished":[],"remaining":[],"files":[]}"#;
        let resp = parse(input).unwrap();
        assert_eq!(resp.status, "blocked");
    }

    #[test]
    fn parse_normalizes_failure_alias_to_blocked() {
        let input = r#"{"status":"failure","summary":"could not complete","accomplished":[],"remaining":[],"files":[]}"#;
        let resp = parse(input).unwrap();
        assert_eq!(resp.status, "blocked");
    }

    #[test]
    fn parse_keeps_canonical_statuses_unchanged() {
        for status in &[
            "new",
            "routed",
            "in_progress",
            "done",
            "blocked",
            "in_review",
            "needs_review",
        ] {
            let input = format!(
                r#"{{"status":"{status}","summary":"test","accomplished":[],"remaining":[],"files":[]}}"#
            );
            let resp = parse(&input).unwrap();
            assert_eq!(
                resp.status, *status,
                "canonical status '{status}' should be unchanged"
            );
        }
    }

    #[test]
    fn parse_keeps_non_canonical_status_as_is_for_audit_trail() {
        // Non-canonical statuses that are not in the alias map are kept as-is
        // so operators can identify which statuses need to be added to normalize_status.
        let input = r#"{"status":"fix_deployed","summary":"deployed","accomplished":[],"remaining":[],"files":[]}"#;
        let resp = parse(input).unwrap();
        // Kept as-is (not normalized away) so build_run_audit can report it explicitly.
        assert_eq!(resp.status, "fix_deployed");
    }

    #[test]
    fn parse_normalizes_heuristic_completion_alias_to_done() {
        // #3627: bare "addressed" and compound "duplicate_skipped" are not in the
        // literal alias map but clearly read as completion via the shared heuristic.
        for status in &["addressed", "duplicate_skipped"] {
            let input = format!(
                r#"{{"status":"{status}","summary":"done","accomplished":[],"remaining":[],"files":[]}}"#
            );
            let resp = parse(&input).unwrap();
            assert_eq!(
                resp.status, "done",
                "status '{status}' should normalize to done"
            );
        }
    }

    #[test]
    fn parse_fails_when_summary_present_but_status_missing() {
        // Payload contains a supported field but omits status/result — should NOT
        // be treated as done. Regression test for gh-issue-2835.
        let input = r#"{"summary":"still working","accomplished":[],"remaining":[],"files":[]}"#;
        let result = parse(input);
        assert!(
            result.is_err(),
            "missing status should error, not default to done"
        );
    }

    #[test]
    fn parse_fails_when_files_present_but_status_missing() {
        // Payload contains files/summary but no status — should not default to done.
        let input = r#"{"summary":"partial output","files":["src/foo.rs"]}"#;
        let result = parse(input);
        assert!(
            result.is_err(),
            "missing status should error, not default to done"
        );
    }

    #[test]
    fn parse_fails_when_only_message_present_without_status() {
        // Payload contains only a message field, no status — should not default to done.
        let input = r#"{"message":"working on it"}"#;
        let result = parse(input);
        assert!(
            result.is_err(),
            "missing status should error, not default to done"
        );
    }

    // ── Regression tests for gh-issue-2880 ───────────────────────────────

    #[test]
    fn parse_regression_2880_ready_for_review_normalized() {
        // ready_for_review should be normalized to needs_review.
        let input = r#"{"status":"ready_for_review","summary":"Ready for review","accomplished":["done"],"remaining":[],"files":["src/main.rs"]}"#;
        let resp = parse(input).unwrap();
        assert_eq!(resp.status, "needs_review");
    }

    #[test]
    fn parse_regression_2880_status_with_output_accepted() {
        // status + output SHOULD be accepted (output is a substantive field).
        // This was already supported via summary extraction in map_generic_response.
        let input = r#"{"status":"done","output":"All tests passed"}"#;
        let resp = parse(input).unwrap();
        assert_eq!(resp.status, "done");
    }

    #[test]
    fn parse_regression_2880_mixed_ndjson_chooses_best_candidate() {
        // In a mixed NDJSON stream:
        // - First blob: file path as status (like the bug)
        // - Later blob: proper response with done status
        // The parser should prefer the one with known status.
        let input = r#"{"status":"src/channels/discord_ws.rs","summary":"","accomplished":[],"remaining":[],"files":[]}
{"status":"done","summary":"Fixed websocket timeout","accomplished":["fixed"],"remaining":[],"files":["src/channels/discord_ws.rs"]}"#;
        let resp = parse(input).unwrap();
        // The parser now prefers known statuses, so it should pick the "done" one.
        assert_eq!(resp.status, "done");
        assert_eq!(resp.summary, "Fixed websocket timeout");
    }

    #[test]
    fn parse_regression_2880_normalize_aliases() {
        // Verify all status aliases are normalized correctly.
        for (input, expected) in [
            ("ok", "done"),
            ("success", "done"),
            ("completed", "done"),
            ("complete", "done"),
            ("skipped", "done"),
            ("skip", "done"),
            ("no_changes_needed", "done"),
            ("no_trades_no_positions", "done"),
            ("changes_made", "done"),
            ("changes_pushed", "done"),
            ("changes_addressed", "done"),
            ("changes_requested_addressed", "done"),
            ("fixed", "done"),
            ("running", "in_progress"),
            ("partial", "in_progress"),
            ("reviewing", "in_review"),
            ("pending_review", "needs_review"),
            ("ready_for_review", "needs_review"),
            ("error", "blocked"),
            ("failed", "blocked"),
        ] {
            let input = format!(
                r#"{{"status":"{input}","summary":"test","accomplished":[],"remaining":[],"files":[]}}"#
            );
            let resp = parse(&input).unwrap();
            assert_eq!(
                resp.status, expected,
                "status '{input}' should normalize to '{expected}'"
            );
        }
    }

    #[test]
    fn parse_regression_2880_telemetry_skipped() {
        // Telemetry objects without proper response structure should be ignored.
        let input = r#"{"type":"telemetry","event":"tool_call"}
{"status":"done","summary":"Work complete","accomplished":["done"],"remaining":[],"files":["src/main.rs"]}"#;
        let resp = parse(input).unwrap();
        assert_eq!(resp.status, "done");
        assert_eq!(resp.summary, "Work complete");
    }

    // ── Regression tests for gh-issue-2979 ───────────────────────────────

    #[test]
    fn parse_regression_2979_success_true_envelope() {
        // Claude returns prose + fenced JSON with `success: true` instead of `status`.
        // Reproduces internal:148284 attempt 3.
        let input = r#"The fix removes all actionable entry/SL/TP/R:R parameters from the bearish-mode trade section. The three candidates are now a plain watchlist table.

```json
{"success": true, "summary": "Fixed bearish/longs contradiction: replaced actionable long setups with a non-executable watchlist table."}
```"#;
        let resp = parse(input).unwrap();
        assert_eq!(resp.status, "done");
        assert_eq!(
            resp.summary,
            "Fixed bearish/longs contradiction: replaced actionable long setups with a non-executable watchlist table."
        );
    }

    #[test]
    fn parse_regression_2979_type_task_complete_envelope() {
        // Claude returns prose + fenced JSON with `type: "task_complete"` instead of `status`.
        // Reproduces internal:147845 attempt 7.
        let input = r#"The daily trending doc is complete at md/research/2026-04-22-daily-trending.md.

```json
{
  "type": "task_complete",
  "summary": "Fetched global trending topics and saved to md/research/2026-04-22-daily-trending.md.",
  "files_changed": ["md/research/2026-04-22-daily-trending.md"]
}
```"#;
        let resp = parse(input).unwrap();
        assert_eq!(resp.status, "done");
        assert!(resp.summary.contains("Fetched global trending topics"));
        assert_eq!(resp.files, vec!["md/research/2026-04-22-daily-trending.md"]);
    }

    #[test]
    fn parse_regression_2979_tests_passed_envelope() {
        // Claude returns prose + fenced JSON with `tests_passed: true` and `files_modified`.
        // Reproduces task 2955 attempt 9.
        let input = r#"PR created: gabrielkoerich/orch#2969

The fix from attempt #8 was already committed. All quality gates pass.

```json
{
  "summary": "Retry watermark persistence on store failure.",
  "files_modified": ["src/engine/review_poll.rs"],
  "pr_url": "https://github.com/gabrielkoerich/orch/pull/2969",
  "tests_passed": true,
  "quality_gates": ["cargo fmt", "cargo clippy", "cargo nextest run"]
}
```"#;
        let resp = parse(input).unwrap();
        assert_eq!(resp.status, "done");
        assert!(resp.summary.contains("Retry watermark persistence"));
        assert_eq!(resp.files, vec!["src/engine/review_poll.rs"]);
    }

    #[test]
    fn parse_regression_2979_success_false_does_not_infer_done() {
        // success: false should NOT infer done — only success: true triggers inference.
        let input = r#"```json
{"success": false, "summary": "Could not complete the task."}
```"#;
        let result = parse(input);
        // success: false has no `status` and `success` is not a valid status value,
        // and infer_completion_status only fires for success=true, so this must error.
        assert!(
            result.is_err(),
            "success: false should not be treated as done"
        );
    }

    #[test]
    fn parse_regression_2979_type_other_does_not_infer_done() {
        // type: "something_else" should NOT infer done.
        let input = r#"{"type": "tool_call", "summary": "calling bash"}"#;
        let result = parse(input);
        assert!(
            result.is_err(),
            "type other than task_complete should not be treated as done"
        );
    }

    #[test]
    fn parse_regression_2979_files_modified_alias() {
        // files_modified should be picked up as the files list.
        let input = r#"{"status": "done", "summary": "done", "files_modified": ["src/foo.rs", "src/bar.rs"]}"#;
        let resp = parse(input).unwrap();
        assert_eq!(resp.files, vec!["src/foo.rs", "src/bar.rs"]);
    }

    // ── Regression tests for gh-issue-3336 ───────────────────────────────

    #[test]
    fn parse_regression_3336_jsonl_domain_status_field_rejected() {
        // Career pipeline JSONL output. Neither record has an orch canonical status
        // and neither has any substantive orch fields — must be rejected, not returned
        // with domain status "inbound-not-submitted" as the task status.
        let input = r#"{"date":"2026-06-15","role":"Senior Solana Engineer","company":"M0","status":"inbound-not-submitted","source":"referral"}
{"date":"2026-06-15","role":"Rust Engineer (Solana)","company":"Exo Tech","status":"applied-follow-up-overdue","source":"linkedin"}"#;
        let err = parse(input).unwrap_err();
        let msg = err.to_string();
        assert!(
            msg.contains("agent output is not valid JSON or a supported JSON wrapper")
                || msg.contains("JSON object lacks substantive"),
            "unexpected error: {msg}"
        );
    }

    #[test]
    fn parse_regression_3336_single_domain_record_rejected() {
        // A single domain JSON record with a non-canonical status and no orch fields.
        let input = r#"{"status":"inbound-not-submitted","date":"2026-06-15","role":"Engineer","company":"Acme"}"#;
        let err = parse(input).unwrap_err();
        let msg = err.to_string();
        assert!(
            msg.contains("agent output is not valid JSON or a supported JSON wrapper")
                || msg.contains("JSON object lacks substantive"),
            "unexpected error: {msg}"
        );
    }

    #[test]
    fn parse_regression_3336_noncanonical_status_with_summary_still_accepted() {
        // A non-canonical orch status WITH a summary is a real agent response and must
        // still be accepted (not broken by the substantive-field guard).
        let input = r#"{"status":"fix_deployed","summary":"deployed to staging","accomplished":[],"remaining":[],"files":[]}"#;
        let resp = parse(input).unwrap();
        assert_eq!(resp.status, "fix_deployed");
        assert_eq!(resp.summary, "deployed to staging");
    }

    #[test]
    fn parse_regression_3336_jsonl_domain_plus_orch_response_prefers_orch() {
        // JSONL output where domain records precede an actual orch response with a
        // canonical status. The parser must pick the orch response.
        let input = r#"{"date":"2026-06-15","role":"Senior Solana Engineer","company":"M0","status":"inbound-not-submitted","source":"referral"}
{"date":"2026-06-15","role":"Rust Engineer (Solana)","company":"Exo Tech","status":"applied-follow-up-overdue","source":"linkedin"}
{"status":"done","summary":"Career radar written to md/career/2026-06-15-radar.md","accomplished":["Wrote career radar"],"remaining":[],"files":["md/career/2026-06-15-radar.md"]}"#;
        let resp = parse(input).unwrap();
        assert_eq!(resp.status, "done");
        assert_eq!(
            resp.summary,
            "Career radar written to md/career/2026-06-15-radar.md"
        );
        assert_eq!(resp.files, vec!["md/career/2026-06-15-radar.md"]);
    }

    #[test]
    fn parse_regression_3336_noncanonical_with_output_accepted() {
        // A non-canonical status with `output` (handled by map_generic_response but not a
        // struct field on AgentResponse) must still be accepted — not rejected with a parse
        // error. Previously `has_substantive_fields` checked the deserialized struct and
        // missed `output`, causing this to fail with "JSON object lacks substantive".
        let input = r#"{"status":"mystery_status","output":"All tests passed"}"#;
        let resp = parse(input).unwrap();
        assert_eq!(resp.status, "mystery_status");
    }

    #[test]
    fn parse_regression_3336_noncanonical_with_message_accepted() {
        // Same as above for `message` — also substantive but absent from the AgentResponse
        // struct, so direct deserialization leaves summary empty.
        let input = r#"{"status":"mystery_status","message":"All tests passed"}"#;
        let resp = parse(input).unwrap();
        assert_eq!(resp.status, "mystery_status");
    }
}