菜单
Track and recall context across sessions. Auto-captures rolling summaries as skills complete. Invoke to review session history, search past work, or see what happened recently.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Multi-agent coordination via SeleneDB context bridge. Check active peers, read shared context, declare work intents, and avoid conflicts when multiple agents work simultaneously.
Commit discipline with graph linking. MERGE GitCommit nodes and link to active milestones, findings, and decisions. Verify before committing, never push unless asked.
Background annotation discipline. All skills auto-annotate graph nodes with rationale, observations, TODOs, and bookmarks as they work. Captures context that feeds into graph-docs documentation generation.
Language-specific standards plus graph-sourced conventions. Reads Convention nodes from SeleneDB, enforces project rules alongside built-in standards. Promotes recurring violations to new conventions.
Structured multi-perspective debate for evaluating decisions. 3-phase methodology: independent generation, adversarial exchange, synthesis. Use when a decision has multiple valid approaches or high stakes.
Systematic debugging: reproduce, hypothesize, isolate, verify. Includes git bisect for regressions, 5 Whys root cause, and hypothesis tracking. Use when encountering bugs, test failures, or unexpected behavior.
| name | session-tracker |
| description | Track and recall context across sessions. Auto-captures rolling summaries as skills complete. Invoke to review session history, search past work, or see what happened recently. |
| argument-hint | [history | search <query> | timeline] |
Maintain continuity across sessions so agents and users can pick up where they left off. When a session ends (user closes terminal, context expires), the context is lost — but the graph remembers. This skill ensures that context is captured incrementally and loaded automatically when a new session starts.
When NOT to use: Looking for a specific finding or decision (use SeleneDB search directly). Tracking a specific initiative (use milestone-tracking). The session just started and context was already loaded by the SessionStart hook.
Every SeleneDB-integrated skill writes a checkpoint when it completes (see selene-integration.md):
MATCH (s:Session) WHERE id(s) = $session_id
SET s.outcome = $outcome,
s.summary = $summary,
s.next_steps = $next_steps
This happens automatically — no user action needed. Each skill writes what it accomplished and what should happen next. Because checkpoints are rolling, context is always current even if the session ends abruptly.
When a new session starts, the SessionStart hook queries SeleneDB for recent session context and injects it alongside the skill routing table. This gives the agent immediate awareness of:
No user action needed. The agent starts informed.
At session start, after creating the Session node, register with the context bridge so other agents can discover this session:
register_agent(
agent_id: "claude-<project>-<instance>",
project: "<project>",
working_on: "<initial task or 'starting session'>",
capabilities: "<skill being invoked>"
)
Also query for active peers and recent shared context:
list_agents(project: "<project>")
get_shared_context(scope: "<project>", since_ms: <last 30 minutes>)
If other agents are active, briefly note their presence:
Peer awareness: [agent_id] is active on [project], working on [working_on].
During skills that run for extended periods (deep-review, research,
feature-design), call heartbeat approximately every 60 seconds with
updated working_on status. This prevents the reaper from marking the
session stale during legitimate work.
heartbeat(
agent_id: "<my agent id>",
working_on: "<current task description>",
files_touched: ["<files being modified>"]
)
When writing the final session summary (the rolling checkpoint), also deregister from the context bridge and share the session summary:
share_context(
author: "<my agent id>",
context_type: "decision",
scope: "<project>",
content: "<session summary with key outcomes and next steps>",
visibility: "project",
ttl_ms: 86400000
)
deregister_agent(agent_id: "<my agent id>")
This ensures clean shutdown: intents are released, the session is marked done, and the summary is available to future agents for 24h.
When invoked with history or without arguments:
MATCH (s:Session)
WHERE s.project = $project AND s.summary IS NOT NULL
OPTIONAL MATCH (s)-[:produced]->(n)
RETURN s.date, s.skill, s.scope, s.summary, s.next_steps,
s.outcome, count(n) AS artifacts_produced
ORDER BY s.date DESC
LIMIT 10
Present as a timeline:
Session history (last 10 sessions):
[date] [skill] — [outcome] [summary] Next: [next_steps] ([N] artifacts produced)
When invoked with search <query>:
CALL graph.hybridSearch('Session', $query, 10)
YIELD node_id, score
MATCH (s:Session) WHERE id(s) = node_id
WHERE s.project = $project
RETURN s.date, s.skill, s.summary, s.next_steps, score
ORDER BY score DESC
This searches across session summaries, scopes, and next_steps. Useful for "when did we last work on auth?" or "what sessions touched the query engine?"
When invoked with timeline:
Query sessions with their linked artifacts to show the full reasoning chain:
MATCH (s:Session)
WHERE s.project = $project
MATCH (s)-[:produced]->(n)
RETURN s.date, s.skill, s.summary, labels(n) AS artifact_type,
CASE
WHEN n:Decision THEN n.summary
WHEN n:Finding THEN n.summary
WHEN n:Insight THEN n.summary
WHEN n:Hypothesis THEN n.statement
WHEN n:Note THEN n.content
ELSE 'artifact'
END AS artifact_summary
ORDER BY s.date DESC, labels(n)
LIMIT 50
Present grouped by session:
[date] [skill]: [summary]
- Decision: [summary]
- Finding: [summary]
- Note: [content] Next: [next_steps]
For skills that need to understand what happened recently across all skills (not just their own):
MATCH (s:Session)
WHERE s.project = $project AND s.date >= date() - duration('P7D')
OPTIONAL MATCH (s)-[:produced]->(d:Decision)
OPTIONAL MATCH (s)-[:produced]->(f:Finding {severity: 'S1_critical'})
OPTIONAL MATCH (s)-[:produced]->(f2:Finding {severity: 'S2_high'})
RETURN s.date, s.skill, s.summary, s.next_steps,
collect(DISTINCT d.summary) AS decisions,
collect(DISTINCT f.summary) AS critical_findings,
collect(DISTINCT f2.summary) AS high_findings
ORDER BY s.date DESC
Summaries should be specific and actionable. "Did some work" is useless. "Researched query optimization, evaluated BTreeMap vs HashMap vs skip list, recommended BTreeMap for ordered range scans" gives the next session everything it needs.
Next steps should be the first thing the next agent reads. Write them as if briefing a colleague who just walked in. Include enough context to start working without re-reading the entire session.
Don't over-capture. The graph already stores decisions, findings, hypotheses, and notes. The session summary is a narrative thread that connects them — not a duplicate of every artifact.
Cross-session queries are the multiplier. A single skill's auto-recall shows "what debug found before." Cross-session context shows "debug found a race condition, deep-review flagged the same module, and there's a deferred item about concurrency." That's the full picture.
| Rationalization | Why It's Wrong |
|---|---|
| "The graph already captures everything" | Nodes capture what happened. Summaries capture the narrative — what it means and what's next. |
| "I'll remember what I was working on" | You might. The agent won't. Session summaries are for agent continuity, not human memory. |
| "Summary is too vague to be useful" | Then write a better one. Specific summaries are the highest-leverage thing a session can produce. |