fix(complete): tolerate non-JSON enrich replies; never abort finalize

The backfill model sometimes answers with prose instead of the JSON
array — e.g. continuing the transcript's own dialogue ('Контекст в
норме... Что дальше?'). The parse error aborted the whole `complete`,
losing the retitle and close. Backfill is best-effort: skip an
unparseable chunk reply (warn), extract a JSON array even when wrapped
in prose, and re-assert 'output ONLY the JSON array, do not continue the
transcript' after the transcript. Retitle/close run regardless of what
enrich recovers.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Author: Shahinyanm

CHANGELOG.md

@@ -7,6 +7,18 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+## [0.22.2] - 2026-06-13
+
+### Fixed
+- **`complete` survives a non-JSON enrich reply.** When the backfill model
+  answered with prose instead of the requested JSON array — e.g. continuing the
+  transcript's own dialogue ("Контекст в норме… Что дальше?") — the parse error
+  aborted the whole `complete`, losing the retitle and close. Backfill is now
+  best-effort: an unparseable chunk reply is skipped (with a warning), the parser
+  extracts a JSON array even when wrapped in prose, and the prompt re-asserts
+  "output ONLY the JSON array, do not continue the transcript" after the
+  transcript. Retitle/close always run regardless of what enrich recovers.
+
 ## [0.22.1] - 2026-06-13
 
 ### Fixed

Cargo.lock

@@ -7,7 +7,7 @@ members = [
 ]
 
 [workspace.package]
-version = "0.22.1"
+version = "0.22.2"
 edition = "2021"
 rust-version = "1.88"
 license = "MIT"

Cargo.toml

@@ -23,7 +23,7 @@ default = ["embed"]
 embed = ["tj-core/embed"]
 
 [dependencies]
-tj-core = { package = "task-journal-core", version = "0.22.1", path = "../tj-core", default-features = false }
+tj-core = { package = "task-journal-core", version = "0.22.2", path = "../tj-core", default-features = false }
 anyhow = { workspace = true }
 clap = { workspace = true }
 tracing = { workspace = true }

crates/tj-cli/Cargo.toml

@@ -40,7 +40,16 @@ impl DreamBackend for LlmDreamBackend {
             };
             let prompt = crate::dream::prompt::build_prompt(&chunk_input);
             let text = self.llm.complete(&prompt, 1024)?;
-            out.extend(parse_backfill_json(&text)?);
+            // Backfill is best-effort: a model that replied with prose instead
+            // of the JSON array (e.g. continued the transcript dialogue) yields
+            // nothing for this chunk, but must NOT abort the whole finalize —
+            // the retitle/close still need to run.
+            match parse_backfill_json(&text) {
+                Ok(evs) => out.extend(evs),
+                Err(e) => {
+                    tracing::warn!(error = %e, "dream backfill: skipping unparseable chunk reply")
+                }
+            }
         }
         Ok(out)
     }
@@ -86,8 +95,12 @@ pub fn parse_backfill_json(text: &str) -> anyhow::Result<Vec<BackfillEvent>> {
         .trim_start_matches("```")
         .trim_end_matches("```")
         .trim();
-    serde_json::from_str(json_str)
-        .with_context(|| format!("dream JSON parse failed; got: {json_str}"))
+    // Tolerate a JSON array wrapped in prose by slicing to the outer brackets.
+    let slice = match (json_str.find('['), json_str.rfind(']')) {
+        (Some(a), Some(b)) if b > a => &json_str[a..=b],
+        _ => json_str,
+    };
+    serde_json::from_str(slice).with_context(|| format!("dream JSON parse failed; got: {json_str}"))
 }
 
 #[cfg(test)]
@@ -111,6 +124,43 @@ mod tests {
         assert!(parse_backfill_json("[]").unwrap().is_empty());
     }
 
+    #[test]
+    fn parse_extracts_array_wrapped_in_prose() {
+        let reply = "Here are the missed events:\n[{\"event_type\":\"finding\",\
+\"task_id\":\"tj-1\",\"text\":\"found\",\"timestamp\":\"2026-06-13T00:00:00Z\"}]\nHope that helps!";
+        let evs = parse_backfill_json(reply).unwrap();
+        assert_eq!(evs.len(), 1);
+    }
+
+    #[test]
+    fn parse_errors_on_pure_prose() {
+        // A conversational reply with no array at all must be an Err so the
+        // backfill loop can skip the chunk instead of inventing events.
+        assert!(parse_backfill_json("Контекст в норме. Что дальше?").is_err());
+    }
+
+    #[test]
+    fn backfill_skips_unparseable_chunk_reply() {
+        // Model replies with prose, not JSON → backfill yields nothing but does
+        // NOT error, so the surrounding finalize (retitle/close) still runs.
+        struct ChattyLlm;
+        impl LlmBackend for ChattyLlm {
+            fn complete(&self, _prompt: &str, _max: u32) -> anyhow::Result<String> {
+                Ok("Контекст в норме. 566.5k/1M использовано. Что дальше?".to_string())
+            }
+            fn name(&self) -> &'static str {
+                "chatty"
+            }
+        }
+        let b = LlmDreamBackend::new(Box::new(ChattyLlm));
+        let input = BackfillInput {
+            tasks: vec![],
+            transcript: "user: hi\nassistant: hello".into(),
+        };
+        let evs = b.backfill(&input).unwrap();
+        assert!(evs.is_empty());
+    }
+
     #[test]
     fn small_transcript_is_one_chunk() {
         let c = chunk_transcript("a\nb\nc\n", 100);

crates/tj-core/src/dream/llm_backend.rs

@@ -34,7 +34,10 @@ pub fn build_prompt(input: &BackfillInput) -> String {
          - Respond with ONLY a JSON array of objects: \
          {{\"event_type\",\"task_id\",\"text\",\"timestamp\"}}. Empty array if nothing missed.\n\n\
          # Candidate tasks and their existing events\n{tasks}\n\
-         # Transcript\n{transcript}\n",
+         # Transcript\n{transcript}\n\n\
+         Remember: output ONLY the JSON array of missed events described above. \
+         Do NOT reply to, summarise, or continue the transcript; if nothing was \
+         missed, output [].\n",
         types = ALLOWED_TYPES,
         tasks = tasks_block,
         transcript = input.transcript,

crates/tj-core/src/dream/prompt.rs

@@ -17,7 +17,7 @@ path = "src/main.rs"
 
 [dependencies]
 # Lean: the MCP server doesn't embed yet, so it skips the model2vec backend.
-tj-core = { package = "task-journal-core", version = "0.22.1", path = "../tj-core", default-features = false }
+tj-core = { package = "task-journal-core", version = "0.22.2", path = "../tj-core", default-features = false }
 anyhow = { workspace = true }
 tokio = { workspace = true }
 tracing = { workspace = true }

crates/tj-mcp/Cargo.toml

@@ -1,6 +1,6 @@
 {
   "name": "task-journal",
-  "version": "0.22.1",
+  "version": "0.22.2",
   "description": "Append-only journal of AI-coding task reasoning chains: hypotheses, decisions, rejections, evidence. Renders compact resume packs so an agent can pick up a 2-week-old task with full context.",
   "author": {
     "name": "Mher Shahinyan"