Extract Chat Session

Extract Chat Session

What This Skill Does

Turn one or more chat session IDs into a continuation packet that is immediately usable for implementation work in the current Codex thread.

Outputs include:

• recovered goals and constraints
• what is already completed
• open loops and pending steps
• workspace and artifact paths
• the next executable action set for the current thread

When To Use

Use this skill when the request includes one or more of the following patterns:

• continue a prior Codex or Copilot chat by id
• continue a prior Cursor agent/composer chat by requestId or composerId
• recover memory or key points from old sessions
• resume long-running research/engineering tasks
• merge context from multiple sessions before continuing execution

Procedure

1. Normalize the request.

• Collect target session IDs.
• Capture the current objective (what to do next, not just what to summarize).
• If objective is missing, infer from the newest session messages.

2. Harvest session evidence.

• For Codex sessions, inspect local artifacts under ~/.codex/.
• Prefer rollout files in ~/.codex/sessions/**/rollout-<...>.jsonl.
• Use session_index.jsonl and history.jsonl as supporting context.
• For Cursor sessions, inspect ~/Library/Application Support/Cursor/User/globalStorage/state.vscdb.
• Prefer cursorDiskKV keys of the form composerData:* and bubbleId:<composerId>:<bubbleId>.
• Accept either:
  • a Cursor composerId directly, or
  • a Cursor requestId that appears inside a bubble payload
• Recover workspace, title, recent user/assistant messages, and any attached local artifact paths from bubble JSON plus decoded conversationState payloads when available.
• Use extractor script to build a concise packet.

3. Build a continuation packet per session.

• Required fields:
  • provider
  • session_id
  • intent_summary
  • workspace_paths
  • completed_work
  • open_loops
  • known_artifacts
  • risks_or_unknowns
  • recommended_next_actions
• Cursor packets should also expose:
  • cursor_composer_id
  • matched_request_ids
  • chat_title
• Keep claims evidence-backed; do not invent missing facts.

4. Merge packets (when multiple sessions are provided).

• Deduplicate repeated outcomes.
• Resolve conflicts by recency and evidence strength.
• Produce one merged execution lane with explicit assumptions.

5. Continue execution in current thread.

• Default behavior for Cursor is to continue in the current Codex thread after extraction; do not assume the user wants to jump back into Cursor UI.
• If the user asked for delivery, route with $delivery-conductor and execute.
• If remote compute is needed (for example 179 server), route with $conductor-remote-long-task.
• Keep verification explicit before completion claims.

6. Save mode (on user request).

• If user asks to save, archive, memo, or write notes, invoke $compile-tech-notes.
• Produce a local markdown artifact with:
  • scope and source session IDs
  • key decisions and evidence
  • completed work and pending tasks
  • execution checklist for next resume
• Use save mode contract.

Output Contract

Always return:

• session continuation packet(s)
• one merged next-step execution plan
• explicit assumptions and missing evidence

When save mode is requested, additionally return:

• local markdown path
• compact note abstract and action checklist

Quality Checks

Before closeout, verify:

• each key claim maps to session evidence
• pending tasks are concrete and executable
• workspace paths are valid
• Cursor requestId or composerId resolves to exactly one recoverable chat lane, or ambiguity is stated explicitly
• no unresolved blocker is hidden inside summary prose

References

• save mode contract
• extractor script