session-health

!git log --oneline -5

Session Health Check

You are running a session health check — a quick diagnostic of the current conversation to determine if you should continue, summarize, checkpoint, or start fresh. Context: $ARGUMENTS

When to Use

• When responses feel shorter, vaguer, or less accurate than earlier
• When you've been working for a long time and want to gut-check quality
• Before starting a major new piece of work in an existing session
• When the user asks "are you still tracking?" or similar

Diagnostic Process

1. Assess Context Load

Review what you've done this session:

• Files read: Roughly how many distinct files have you read?
• Files modified: How many files have you written or edited?
• Agents dispatched: How many subagent tasks have you launched?
• Topic breadth: How many distinct areas of the codebase have you touched?

Load Level	Files Read	Agents	Breadth	Signal
Light	<20	0-2	1-2 areas	Healthy, continue
Moderate	20-60	3-5	3-4 areas	Watch quality
Heavy	60-120	6-9	5+ areas	Check quality signals; consider fresh session
Overloaded	120+	10+	Many areas	Quality degraded — fresh session recommended

Note: Raw file counts are a lagging indicator. Quality signals (step 3) are the primary degradation signal — a session with 30 files but heavy hedging is more degraded than one with 80 files and sharp answers.

2. Check Scope Coherence

Ask yourself:

• Original goal: What did the user first ask for this session?
• Current work: What are you doing right now?
• Drift: Has scope crept from the original goal? How far?
• Batch progress: If processing a numbered list or collection, how many items are done vs remaining? If >50% done, recommend checkpointing results to a file before continuing.

Scope drift isn't always bad — discovered work is natural. But if you've drifted far from the original intent, flag it.

3. Check Quality Indicators

Quality signals are the primary degradation signal. A session can be Heavy by file count but still sharp; a session can be Moderate by file count but already degrading. Check these first and let them override the load level assessment:

Signal	Healthy	Degrading
Hedging language	Direct, confident answers	"I think", "it might", "probably", "as I recall" — increasing over time
Repetition	Novel, additive responses	Re-explaining things already stated earlier in the session
Tool call quality	Targeted searches, specific file reads	Broad grep patterns, re-reading files already read, "let me check again"
Tool call failure rate	Rare failures or none	Repeated tool failures, fallbacks to guessing
Answer specificity	Cites file paths, line numbers, exact values	Vague references, omitting specifics that were available

Rate each signal: Good / Watch / Degrading. If two or more signals are Degrading, elevate the overall session health to Fresh regardless of load level.

4. Check Session State

git status                     # Any uncommitted work?

Check your task tracker for in-progress and available work.

5. Check Team Health (conditional)

If .claude/team.yaml exists, read it and check each member's learnings:

ls .claude/tackline/memory/agents/*/learnings.md 2>/dev/null

For each member in the team manifest, check:

• Learnings file exists? If missing, flag as "unconfigured"
• Line count: >120 lines = bloated (flag for /curate)
• Last modified date: >14 days since last update = stale

Report team state in the output:

**Team**: [team name] — [N] members
- [member]: [line count] learnings, last updated [date] — [healthy / stale / bloated]

If no team is configured, skip this section silently (not an error — many projects don't use teams).

6. Recommendation

Based on the above, recommend ONE of:

Action	When	How
Continue	Light-moderate load, on-topic, quality fine	Keep working
Partial summarize	Heavy load but good progress, quality still sharp	Selectively compress completed work threads while preserving in-progress work at full context. For each completed thread (a task or subtask where the output is final): write a 3-5 bullet summary capturing goal, approach, outcome, and key decisions, then release the detailed exchange from working context. Leave all in-progress threads uncompressed. This is a model-side operation — do not ask the user to summarize.
Checkpoint & Continue	Moderate load, mid-batch work, or want a save point	Commit current work, write intermediate results to .claude/tackline/memory/scratch/<slug>.md if mid-batch, sync task state, then continue
Fresh session	Overloaded, scope has drifted, or 2+ quality signals degrading	Run /handoff to capture context, commit all work, sync task state, push, then start a new session
Break into subagents	Remaining work is parallelizable	Dispatch independent tasks to subagents to avoid further context fill

Output Format

## Session Health: [Healthy / Watch / Summarize / Fresh]

**Load**: [Light/Moderate/Heavy/Overloaded] — ~X files read, ~Y agents, Z areas
**Scope**: [On track / Minor drift / Significant drift] — started with [X], now doing [Y]
**Quality**: [Good / Showing wear / Degrading]
**Uncommitted work**: [Yes/No] — [details if yes]
**Team**: [team name — N members, M stale | No team configured]

### Recommendation
[Action] — [1-2 sentence rationale]

### If continuing, suggested next steps:

**Sharpening requirement**: Each step must be a concrete command or specific file reference, not a suggestion.

- **Before**: "Consider refactoring the authentication module"
- **After**: "Run /review on src/auth/ — 3 files changed in last 5 commits, fix rate 40%"

1. [runnable command or specific file path]
2. [runnable command or specific file path]

After producing the health report, emit a pipe-format summary so /advise or /retro can consume health signals:

## Session health signals

**Source**: /session-health
**Input**: [current session, $ARGUMENTS if provided]
**Pipeline**: (none — working from direct input)

### Items (4)

1. **Context load** — [Light | Moderate | Heavy | Overloaded]
   - files-read: [~N]
   - agents-dispatched: [~N]
   - topic-breadth: [N areas]
   - confidence: CONFIRMED

2. **Quality signals** — [Good | Showing wear | Degrading]
   - hedging: [Good | Watch | Degrading]
   - repetition: [Good | Watch | Degrading]
   - tool-call-quality: [Good | Watch | Degrading]
   - answer-specificity: [Good | Watch | Degrading]
   - confidence: CONFIRMED

3. **Scope coherence** — [On track | Minor drift | Significant drift]
   - original-goal: [one-line summary]
   - current-work: [one-line summary]
   - batch-progress: [N/M items complete | not applicable]

4. **Recommendation** — [Continue | Partial summarize | Checkpoint & Continue | Fresh session | Break into subagents]
   - rationale: [one sentence]
   - next-action: [specific command or file reference]
   - confidence: CONFIRMED

### Summary

[One paragraph: load level, quality assessment, and recommended action with a concrete reason.]

Important Notes

• This is a self-diagnostic — be honest, not optimistic
• Quality signals override file counts. Two or more degrading quality signals means Fresh, regardless of load level
• Committing and syncing before any session change is mandatory
• When in doubt, checkpoint (commit + sync) before deciding — it's cheap insurance
• A fresh session with good task context is often more productive than a degraded long session
• Before starting a fresh session, run /handoff to capture in-progress tasks, decisions, and open questions in a structured handoff document

See Also

• /handoff — structured session transition; captures context before clearing or starting fresh
• /status — lightweight status card; use before /session-health for a quick pulse
• /retro — end-of-session retrospective; run after recommending "Fresh session" to log learnings before closing