Skip to content
Back to skills

Session Continuity Management

ASecurity

Start of any session, especially after context compaction. Also triggered

  • 2 stars
  • 0 votes
  • 0 copies
  • 3 views
  • Added September 19, 2026
ai-agents

Security analysis

A100/100

Scanned September 19, 2026

npx -y skills add Stranglehold/Exocortex --skill session-continuity-management --agent claude-code

Installs into .claude/skills of the current project.

Are you the author of Session Continuity Management?

Add the live security badge to your README. It updates with every re-scan.

Security grade badge for Session Continuity Management
[![Security: A — Skills Directory](https://www.skillsdirectory.com/api/skills/stranglehold-session-continuity-management/badge)](https://www.skillsdirectory.com/skills/stranglehold-session-continuity-management)

More formats (shields.io, HTML) on the badges page. Keep it an A: scan every change in CI with Pro.

Download with Pro
SKILL.md
---
name: session-continuity-management
description: Start of any session, especially after context compaction. Also triggered
  when user references past work, says "where...
triggers:
- Start of any session, especially after context compaction. Also triggered when user
  references past work, says "where...
version: '1.0'
author: Exocortex
---

# Skill: Session Continuity Management

## Trigger
Start of any session, especially after context compaction. Also triggered when user references past work, says "where were we," "continue from," or asks about decisions made in previous sessions.

## Inputs Available
- **Compaction summary** — if present at top of context, contains structured summary of prior conversation
- **Transcripts** — full conversation logs at `/mnt/transcripts/`, named by datetime
- **Journal** — `/mnt/transcripts/journal.txt`, one-line summaries of every session
- **Memory** — `userMemories` tag in system prompt, background knowledge about the user
- **Past chats tools** — `conversation_search` and `recent_chats` for cross-session retrieval
- **Workflow state** — `/home/claude/.workflow_state.json` if a plan was active

## Procedure

### 1. Assess Context State
At session start, determine what you know:
- Is there a compaction summary? Read it — it contains the structured state of the previous conversation.
- Is the user continuing from a specific topic? Check if there's enough context to proceed or if you need to recover.
- Is there an active workflow plan? Check workflow state.

### 2. Recover Context (if needed)
If the user references something not in your current context:

**For specific topics:** Use `conversation_search` with substantive keywords (nouns, project names, technical terms). Run 2-3 searches with different keyword strategies — the same query expansion principle from MemR3.

**For recent work:** Use `recent_chats` with appropriate time filters. If the user says "yesterday," filter by yesterday's date range.

**For deep history:** Read the journal at `/mnt/transcripts/journal.txt` first — it's a one-line index of every session. Then read specific transcript files for detail. Use `view` with line ranges — don't try to read 400KB transcripts in one call.

**For project state:** Check `/mnt/user-data/outputs/` for the latest versions of specs, prompts, and README. These are the ground truth for current project state.

### 3. Establish Continuity
When referencing past work:
- State the context naturally, as if you remember it. Do not say "according to my records" or "I found in the transcript."
- If you're uncertain about a detail, say so directly rather than fabricating.
- If the user corrects you, update your understanding immediately. Use `memory_user_edits` if the correction is persistent.

### 4. Manage Active Workflows
If a workflow plan exists:
- Run `workflow.py status` to see current position
- Resume from where the plan left off
- If the plan is stale (from a previous session), confirm with the user before continuing

### 5. Handle Compaction Gracefully
When context is compacted mid-session:
- The compaction summary preserves key state. Read it carefully.
- Transcript of the full conversation is available at the path noted in the summary.
- Do NOT re-read the entire transcript unless you need a specific detail. The summary is designed to be sufficient for continuity.
- If you need a detail not in the summary, use targeted `view` with line ranges on the transcript.

## Key Paths
```
/mnt/transcripts/journal.txt          — Session index (read first)
/mnt/transcripts/*.txt                — Full transcripts (read targeted sections)
/mnt/user-data/outputs/               — Current project deliverables
/home/claude/.workflow_state.json      — Active workflow plan
```

## Quality Checks
- [ ] Never claim "I don't have access to previous conversations" without first trying the past chats tools
- [ ] Never read an entire 400KB+ transcript in one call — use journal index, then targeted reads
- [ ] Context references are natural, not meta-commentary about memory systems
- [ ] Active workflows are checked at session start
- [ ] Compaction summaries are treated as authoritative for session state

## Anti-Patterns
- **Reading everything.** The journal exists so you don't have to read every transcript. Index first, detail second.
- **Pretending to remember.** If you don't have the context and can't find it, say so. Don't fabricate continuity.
- **Ignoring the workflow tracker.** If a plan was active, it should be resumed or explicitly closed. Orphaned plans create confusion.
- **Meta-commentary about tools.** "Let me search my past conversations" is unnecessary. Just search and present the information naturally.
- **Re-reading compaction summaries aloud.** The user was there. They don't need a recap unless they ask for one.

Attribution

Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.

Comments

Loading comments…

Newsletter

Get the weekly list of safe new skills, and the ones we pulled.