Session Management Deep Dive

Session Management & Compaction (Deep Dive)

This document explains how CoderClaw manages sessions end-to-end:

• Session routing (how inbound messages map to a sessionKey)
• Session store (sessions.json) and what it tracks
• Transcript persistence (*.jsonl) and its structure
• Transcript hygiene (provider-specific fixups before runs)
• Context limits (context window vs tracked tokens)
• Compaction (manual + auto-compaction) and where to hook pre-compaction work
• Silent housekeeping (e.g. memory writes that shouldn’t produce user-visible output)

If you want a higher-level overview first, start with:

• /concepts/session
• /concepts/compaction
• /concepts/session-pruning
• /reference/transcript-hygiene

---

Source of truth: the Gateway

CoderClaw is designed around a single Gateway process that owns session state.

• UIs (macOS app, web Control UI, TUI) should query the Gateway for session lists and token counts.
• In remote mode, session files are on the remote host; “checking your local Mac files” won’t reflect what the Gateway is using.

---

Two persistence layers

CoderClaw persists sessions in two layers:

1. Session store (sessions.json)

   • Key/value map: sessionKey -> SessionEntry
   • Small, mutable, safe to edit (or delete entries)
   • Tracks session metadata (current session id, last activity, toggles, token counters, etc.)

2. Transcript (<sessionId>.jsonl)

   • Append-only transcript with tree structure (entries have id + parentId)
   • Stores the actual conversation + tool calls + compaction summaries
   • Used to rebuild the model context for future turns

---

On-disk locations

Per agent, on the Gateway host:

• Store: ~/.coderclaw/agents/<agentId>/sessions/sessions.json
• Transcripts: ~/.coderclaw/agents/<agentId>/sessions/<sessionId>.jsonl
  • Telegram topic sessions: .../<sessionId>-topic-<threadId>.jsonl

CoderClaw resolves these via src/config/sessions.ts.

---

Session keys (sessionKey)

A sessionKey identifies which conversation bucket you’re in (routing + isolation).

Common patterns:

• Main/direct chat (per agent): agent:<agentId>:<mainKey> (default main)
• Group: agent:<agentId>:<channel>:group:<id>
• Room/channel (Discord/Slack): agent:<agentId>:<channel>:channel:<id> or ...:room:<id>
• Cron: cron:<job.id>
• Webhook: hook:<uuid> (unless overridden)

The canonical rules are documented at /concepts/session.

---

Session ids (sessionId)

Each sessionKey points at a current sessionId (the transcript file that continues the conversation).

Rules of thumb:

• Reset (/new, /reset) creates a new sessionId for that sessionKey.
• Daily reset (default 4:00 AM local time on the gateway host) creates a new sessionId on the next message after the reset boundary.
• Idle expiry (session.reset.idleMinutes or legacy session.idleMinutes) creates a new sessionId when a message arrives after the idle window. When daily + idle are both configured, whichever expires first wins.

Implementation detail: the decision happens in initSessionState() in src/auto-reply/reply/session.ts.

---

Session store schema (sessions.json)

The store’s value type is SessionEntry in src/config/sessions.ts.

Key fields (not exhaustive):

• sessionId: current transcript id (filename is derived from this unless sessionFile is set)
• updatedAt: last activity timestamp
• sessionFile: optional explicit transcript path override
• chatType: direct | group | room (helps UIs and send policy)
• provider, subject, room, space, displayName: metadata for group/channel labeling
• Toggles:
  • thinkingLevel, verboseLevel, reasoningLevel, elevatedLevel
  • sendPolicy (per-session override)
• Model selection:
  • providerOverride, modelOverride, authProfileOverride
• Token counters (best-effort / provider-dependent):
  • inputTokens, outputTokens, totalTokens, contextTokens
• compactionCount: how often auto-compaction completed for this session key
• memoryFlushAt: timestamp for the last pre-compaction memory flush
• memoryFlushCompactionCount: compaction count when the last flush ran

The store is safe to edit, but the Gateway is the authority: it may rewrite or rehydrate entries as sessions run.

---

Transcript structure (*.jsonl)

Transcripts are managed by @mariozechner/pi-coding-agent’s SessionManager.

The file is JSONL:

• First line: session header (type: "session", includes id, cwd, timestamp, optional parentSession)
• Then: session entries with id + parentId (tree)

Notable entry types:

• message: user/assistant/toolResult messages
• custom_message: extension-injected messages that do enter model context (can be hidden from UI)
• custom: extension state that does not enter model context
• compaction: persisted compaction summary with firstKeptEntryId and tokensBefore
• branch_summary: persisted summary when navigating a tree branch

CoderClaw intentionally does not “fix up” transcripts; the Gateway uses SessionManager to read/write them.

---

Context windows vs tracked tokens

Two different concepts matter:

1. Model context window: hard cap per model (tokens visible to the model)
2. Session store counters: rolling stats written into sessions.json (used for /status and dashboards)

If you’re tuning limits:

• The context window comes from the model catalog (and can be overridden via config).
• contextTokens in the store is a runtime estimate/reporting value; don’t treat it as a strict guarantee.

For more, see /token-use.

---

Compaction: what it is

Compaction summarizes older conversation into a persisted compaction entry in the transcript and keeps recent messages intact.

After compaction, future turns see:

• The compaction summary
• Messages after firstKeptEntryId

Compaction is persistent (unlike session pruning). See /concepts/session-pruning.

---

When auto-compaction happens (Pi runtime)

In the embedded Pi agent, auto-compaction triggers in two cases:

1. Overflow recovery: the model returns a context overflow error → compact → retry.
2. Threshold maintenance: after a successful turn, when:

contextTokens > contextWindow - reserveTokens

Where:

• contextWindow is the model’s context window
• reserveTokens is headroom reserved for prompts + the next model output

These are Pi runtime semantics (CoderClaw consumes the events, but Pi decides when to compact).

---

Compaction settings (reserveTokens, keepRecentTokens)

Pi’s compaction settings live in Pi settings:

{
  compaction: {
    enabled: true,
    reserveTokens: 16384,
    keepRecentTokens: 20000,
  },
}

CoderClaw also enforces a safety floor for embedded runs:

• If compaction.reserveTokens < reserveTokensFloor, CoderClaw bumps it.
• Default floor is 20000 tokens.
• Set agents.defaults.compaction.reserveTokensFloor: 0 to disable the floor.
• If it’s already higher, CoderClaw leaves it alone.

Why: leave enough headroom for multi-turn “housekeeping” (like memory writes) before compaction becomes unavoidable.

Implementation: ensurePiCompactionReserveTokens() in src/agents/pi-settings.ts (called from src/agents/pi-embedded-runner.ts).

---

User-visible surfaces

You can observe compaction and session state via:

• /status (in any chat session)
• coderclaw status (CLI)
• coderclaw sessions / sessions --json
• Verbose mode: 🧹 Auto-compaction complete + compaction count

---

Silent housekeeping (NO_REPLY)

CoderClaw supports “silent” turns for background tasks where the user should not see intermediate output.

Convention:

• The assistant starts its output with NO_REPLY to indicate “do not deliver a reply to the user”.
• CoderClaw strips/suppresses this in the delivery layer.

As of 2026.1.10, CoderClaw also suppresses draft/typing streaming when a partial chunk begins with NO_REPLY, so silent operations don’t leak partial output mid-turn.

---

Pre-compaction “memory flush” (implemented)

Goal: before auto-compaction happens, run a silent agentic turn that writes durable state to disk (e.g. memory/YYYY-MM-DD.md in the agent workspace) so compaction can’t erase critical context.

CoderClaw uses the pre-threshold flush approach:

1. Monitor session context usage.
2. When it crosses a “soft threshold” (below Pi’s compaction threshold), run a silent “write memory now” directive to the agent.
3. Use NO_REPLY so the user sees nothing.

Config (agents.defaults.compaction.memoryFlush):

• enabled (default: true)
• softThresholdTokens (default: 4000)
• prompt (user message for the flush turn)
• systemPrompt (extra system prompt appended for the flush turn)

Notes:

• The default prompt/system prompt include a NO_REPLY hint to suppress delivery.
• The flush runs once per compaction cycle (tracked in sessions.json).
• The flush runs only for embedded Pi sessions (CLI backends skip it).
• The flush is skipped when the session workspace is read-only (workspaceAccess: "ro" or "none").
• See Memory for the workspace file layout and write patterns.

Pi also exposes a session_before_compact hook in the extension API, but CoderClaw’s flush logic lives on the Gateway side today.

---

Troubleshooting checklist

• Session key wrong? Start with /concepts/session and confirm the sessionKey in /status.
• Store vs transcript mismatch? Confirm the Gateway host and the store path from coderclaw status.
• Compaction spam? Check:
  • model context window (too small)
  • compaction settings (reserveTokens too high for the model window can cause earlier compaction)
  • tool-result bloat: enable/tune session pruning
• Silent turns leaking? Confirm the reply starts with NO_REPLY (exact token) and you’re on a build that includes the streaming suppression fix.