Agent-Skills.md

Agent Skills: Recap: multimodel-safe session continuation

Token-savvy session continuation. Rebuilds working context from transcript + soul memories in ~1500 tokens instead of replaying full history. Use when starting a new session to continue previous work.

UncategorizedID: genomewalker/cc-soul/recap

Repository

genomewalker/cc-soul
genomewalker
33

Install this agent skill to your local

pnpm dlx add-skill https://github.com/genomewalker/cc-soul/tree/HEAD/skills/recap

Skill Files

Browse the full folder contents for recap.

Download Skill

Loading file tree…

skills/recap/SKILL.md

Skill Metadata

Name
recap
Description
"Rebuild working context from transcripts and chitta memories in approximately 1500 tokens. Use when starting a new session to continue previous work."

Recap: multimodel-safe session continuation

Reconstruct a bounded working context from an exact Claude or Codex transcript. Claude and Codex share one Chitta registry and one thread ledger. Never choose a transcript from global file recency and never resume a live session.

1. Resolve the shared implementation

Locate chitta-mcp/resume_selector.py in this plugin root. Prefer $CHITTA_PLUGIN_DIR (falling back to $CC_SOUL_PLUGIN_DIR), then the source root containing this skill, then the newest installed ~/.claude/plugins/cache/genomewalker-chitta/chitta/* (or the pre-rename genomewalker-cc-soul/cc-soul/*), then the Python environment from the chitta-mcp executable (read its shebang and import resume_selector to locate the module directory). If the selector is unavailable, stop and report that chitta must be updated; do not fall back to "newest transcript" heuristics.

Set the invoking client to codex when a CODEX_* session variable is present, otherwise claude when a CLAUDE_* session variable is present. The selector also identifies the current session from its registered ancestor PID.

2. Inventory before selecting

Run:

python3 "$_MCP_DIR/resume_selector.py" --project-dir "$PWD" --client "$CLIENT"

Arguments map as follows:

  • /recap <session_id>: add --session-id <session_id>.
  • /recap <thread id or title>: add --thread <value>.
  • /recap last N: do not select; show the first N candidates.

Always surface the live_sessions inventory before proceeding. Show client, model, short session ID, project, thread, and PID. This inventory includes both Claude and Codex, so parallel work in the same project is visible.

Handle status strictly:

  • registry_unavailable: Chitta did not answer the liveness query. Stop and retry later; never interpret a timeout as an empty live-session list.
  • locked: report the owning live session/client/thread. Do not read, recap, or resume it.
  • ambiguous: show the candidate session/client/project/thread choices and ask the user to choose. Never break a tie using mtime alone.
  • none: report that no registered or project-exact historical transcript was found.
  • selected: continue with that exact session and transcript only.

3. Atomically claim the thread

Before reading or presenting continuation context, rerun the same command with --claim. This closes the race between selection and takeover. Continue only when claim.claimed is true. If the claim reports live_owner, show that owner and stop. Never use --force unless the user explicitly asks to take over a known live thread.

The selector may create a durable thread for a historical transcript that predates thread tracking. It records the current session as the new owner; do not write a global .current_thread_id file.

4. Read bounded exact context

Use the selected session_id with Chitta's exact transcript reader, or parse the selected transcript_path if MCP is unavailable. Read at most:

  • First 20 and last 30 user turns, 180 characters each.
  • Last 20 assistant turns, 240 characters each.

Support both formats:

  • Claude: top-level type=user|assistant, content in message.content.
  • Codex: top-level type=response_item, with payload.type=message, payload.role, and payload.content.

Ignore environment blocks, system/developer reminders, command markup, task notifications, interruption markers, and duplicate messages. User corrections and explicit scope changes have highest priority.

Query Chitta smart context using the recovered task as the query, then inspect git status --short and git log --oneline -5 in the selected project.

5. Produce the recap

Keep the result near 1500 tokens:

## Session Recap: <short session id> (<client/model>)

### What was being done
<primary task and scope>

### Key decisions
<important choices and user corrections>

### Current state
- Project/branch
- Uncommitted files
- Other live Claude/Codex sessions in this project

### Pending work
<unfinished work and next action>

### Key context
<critical errors, constraints, and architectural facts>

Do not replay full tool output. Prefer decisions and pending work over a diary of actions. Mention files only when they are still active.