fix(learn): read session transcript in forked execution context (#230)

* fix(learn): read session transcript in forked execution context

The learn skill runs with `context: fork`, which creates an isolated
subagent with no access to the parent conversation. Pass the
transcript_path from the stop hook input through to the skill
invocation, and add Step 0 to SKILL.md instructing the forked agent
to read the transcript file for analysis.

* fix(learn): disambiguate fallback file format in SKILL.md

Addresses CodeRabbit review finding: the fallback guidance that points
to .evolve/trajectories/ was ambiguous about format. Explicitly name
both file types the directory may contain (trajectory_*.json as a
single JSON object, claude-transcript_*.jsonl as raw JSONL) and how to
parse each.

* fix(learn): document how to extract transcript_path from stop hook message

Addresses CodeRabbit review finding: SKILL.md referenced <transcript_path>
as a placeholder but didn't tell the forked agent how to obtain the actual
value. Point to the literal marker "The session transcript is at: <path>"
produced by on_stop.py and describe how to extract and strip the path.

---------

Co-authored-by: Vatche Isahagian <visahak@users.noreply.github.com>

Author: vinodmut

platform-integrations/claude/plugins/evolve-lite/skills/learn/SKILL.md

@@ -25,9 +25,28 @@ This skill analyzes the current conversation to extract guidelines that **correc
 
 ## Workflow
 
+### Step 0: Load the Conversation
+
+This skill runs in a forked context with no access to the parent conversation. The stop hook message (produced by `on_stop.py`) contains the literal marker `The session transcript is at: <path>` — find that exact phrase, take everything after the colon, strip surrounding whitespace and quotes, and use the result as `transcript_path`. Then read it:
+
+```bash
+cat <transcript_path>
+```
+
+**You must read this file to analyze the conversation** — the forked context has no other access to it.
+
+The transcript is JSONL: each line is a separate JSON object. Focus on lines where `"type": "assistant"` or `"type": "human"` to reconstruct the conversation flow. Look for tool calls, errors in tool results, and user corrections.
+
+If no transcript path was provided, fall back to `.evolve/trajectories/`, which may contain either format:
+
+- **`trajectory_*.json`** — a single JSON object with `messages: [{role, content}, …]`. Prefer the most recent one; parse with `json.load`.
+- **`claude-transcript_*.jsonl`** — raw Claude JSONL (same format as the primary `transcript_path`). Parse line-by-line.
+
+If no transcript is available at all, output zero entities.
+
 ### Step 1: Analyze the Conversation
 
-Review the conversation and identify:
+Review the conversation (loaded from the transcript) and identify:
 
 - **Wasted steps**: Where did the agent go down a path that turned out to be unnecessary? What would have been the direct route?
 - **Errors hit**: What errors occurred? What knowledge would have prevented them?

platform-integrations/claude/plugins/evolve-lite/skills/learn/scripts/on_stop.py

@@ -14,11 +14,16 @@ def main():
     if input_data.get("stop_hook_active") is True:
         return
 
+    transcript_path = input_data.get("transcript_path", "")
+    reason = "Run the /evolve-lite:learn skill."
+    if transcript_path:
+        reason += f" The session transcript is at: {transcript_path}"
+
     print(
         json.dumps(
             {
                 "decision": "block",
-                "reason": "Run the /evolve-lite:learn skill to extract guidelines from this conversation.",
+                "reason": reason,
                 "suppressOutput": True,
                 "systemMessage": "Running the evolve-lite learn skill...",
             }