Context Manual

Context Manual

Use the schema for routine calls; open this manual for unfamiliar workflows or recovery risk. Context has four operations:

• summarize records compact replacements for selected old tool results; it does not rebuild provider context.
• rebuild recomposes all canonical prompt sources, applies summaries, then requests provider replay. Bare input={} is valid; durable file edits do not hot-load. It is strongly discouraged as routine compaction, tool-result cleanup, or a ceremony after every Pad/Knowledge edit: full replay is costly, can disturb prompt-prefix cache reuse, and still retains the bulky history.
• molt sheds conversation while retaining durable files and stores. It is an irreversible boundary, not housekeeping, but it is the preferred deliberate response to substantial or noisy history or sustained pressure when a fresh briefing is worth its cost.
• manual returns the installed context-manual without changing context.

Shortest useful path

When current meta_guidance says a consumed result needs a posteriori compaction, inspect it first, then use context(action="summarize", input={"items": [...]}); do not use this as routine cleanup. Prefer a deliberate molt over a manual rebuild when history is substantial or noisy. Routine durable edits can simply wait for the next natural molt or reload; they do not need an immediate rebuild. Rare exception: when new canonical prompt sources or pending summaries must apply in this conversation and a molt is unsuitable, make one targeted context(action="rebuild", input={})—never a loop. The summarize reference owns selection, recovery, and pressure guidance.

Before a deliberate molt:

1. Before every deliberate molt, reconcile and edit system/pad.md into current working state, not an appended timeline. Keep active work, authority, holds, unresolved choices, live IDs, evidence pointers and the next action inline. Losslessly save completed/superseded history to the appropriate Knowledge entry before trimming it from Pad; retain a retrieval pointer and verify the archive is readable. Do not erase unique evidence or treat an old status as current. Update other stores only when changed; no summarize or rebuild is required first.
2. Write the session-journal child and parent-index pointer. It must be knowledge/session-journal/<entry>/KNOWLEDGE.md, nonempty UTF-8, with valid name/description frontmatter and type: session-journal or session_journal: true. Use the journal scaffold if useful.
3. Pass its path as session_journal_path and write the shortest sufficient handoff: active state, verified results, pending authority, next steps, live IDs, artifact/evidence paths, blockers, and lessons. Use the optional handoff scaffold only when it lowers real risk.
4. Call context(action="molt", ...). The journal is validated before shedding; invalid input leaves molt count/history untouched. keep_tool_calls refuses unknown IDs; keep_last preserves a suffix and complete tool-call batches. The schema is authoritative.

Molt preserves durable files and source reconstruction: it archives/snapshots the old conversation, rebuilds the fresh prompt, and publishes a post-molt notification. It does not delete durable stores. Molt for sustained pressure, an explicit human reset, confusion, or a long-inactivity reminder as described below—not merely task completion. Tending stores need not create gratuitous edits or new skills.

An eligible system.auto_molt notification requires BOTH continuous IDLE/ASLEEP inactivity ≥ the positive duration (valid LINGTAI_AUTO_MOLT_IDLE_SECONDS > v2 settings/system.json auto_molt_idle_seconds > default 7200 s) and CURRENT context strictly >100000 tokens, not cumulative input or since-molt spending. It is a valid actionable molt reason, not task completion or a forced reset. At a safe idle boundary, refresh Pad and archive stale history to Knowledge as above, write a valid session journal, then molt once. Preserve unfinished authorized work and explicitly defer the molt rather than discard context. After post-molt reorientation, remain IDLE until human input or other real work; do not invent work from the reminder or handoff. One reminder is sent per continuous inactivity period; an ordinary ACTIVE notification wake clears the period and later inactivity starts another, still subject to both conditions. Unloaded/unknown context is ineligible; after molt the replacement context is checked, with no guarantee it is below 100000. The timer/latch are process-local and reset on restart; the runtime gate contract owns these semantics. Notification delivery may start an ordinary paid turn, but the kernel never forces discard or generates a journal.

Ordinary molt/rebuild retains notification delivery identity within this agent process: already delivered events do not automatically replay. Carry still-needed message content and routing IDs in your authorized handoff; new post-molt notices and new producer events still deliver. A new Agent/process restart starts fresh bookkeeping, not cross-crash exactly-once.

After a molt

Read applicable AGENTS.md and CLAUDE.md working instructions after fresh context, and applicable nested instructions when entering that project scope. Do not paste their bodies into private memory or create/overwrite those files merely for orientation. Pad and Knowledge remain private memory; legacy character is preserved once into Pad, not injected as a separate section.

Read the post-molt notification/summary_path, current prompt/Pad, journal index, and new producer messages. An inactivity-triggered molt with no remaining authorized work ends in IDLE waiting for human input or other real work; its own post-molt notice is reorientation, not new work. Treat the briefing as context, not a command to run blindly. Re-check IDs and evidence paths. After reorientation, explicitly choose continue, defer, or obsolete and record that decision in your session journal/pad. The post-molt notification is delivered once and is not re-attached automatically; there is no notification dismiss action, and delivery does not mean any human task is completed (notification(action="check") rereads it deliberately). Resume through the originating channel/tool; a saved briefing or provider replay is not proof of delivery.

The action summarize records Context history; the optional root summarize boolean only presents results and defaults to true. They are not the same input. Proactively set root summarize=false for verbatim manuals and mandatory gates, exact names/paths, field presence, source, diffs, or evidence; otherwise specify what to retain in reasoning. Use psyche(action="manual", input={}, reasoning="load durable-store routes") for store ownership; durable writes go through shell (verify the old text exists exactly once before replacing, and read the file back afterwards), and may wait for the next natural molt or reload; rebuild only under the rare exception above.