Summarize Manual

Summarize Manual

Use this reference after you have inspected a bulky tool result and know what a future turn must retain, or when current Context guidance asks you to compact consumed results. It is the focused procedure; resident meta_guidance remains the source for current pressure timing.

Choose the operation

• context(action="summarize") is record-only. It replaces selected prior tool-result blocks in runtime history with your agent-authored summaries, marks them pending, and preserves the raw event for recovery. The active provider continuation may still carry the raw block.
• context(action="rebuild") is the one active application path, but a rare exception rather than routine compaction: full replay is costly, can disturb prompt-prefix cache reuse, and retains bulky history. It first re-reads and recomposes all canonical prompt sources, then records/applies supplied and pending summaries, then requests provider replay with the new prompt and history. Bare input={} is valid, including with no pending summaries.
• context(action="molt") is the whole-session boundary. Choose it when pressure, an explicit human reset request, or conversation confusion makes a fresh briefing worthwhile; do not use it as task-completion housekeeping.

The root summarize boolean is a separate generic result-presentation control. It is not the context action, is not an item-list input, and does not apply a recorded summary to provider context. The action/schema descriptions are authoritative for fields and requiredness.

A-priori root result summaries

Before a call, root summarize defaults to true for migrated families and requests a generated, noncanonical, lossy summary. Use it for anticipated long, bulky output; set summarize=false for expected short results or required exact text, such as a mandatory gate, name/path, field, source, diff or evidence. Give CLEAR retention targets in reasoning: status, key fields, counts, anchors and evidence, not extra interpretation or conclusions. Do not infer that an unprinted check was not performed. Narrow the read and preserve raw_locator (and producer ID) for recovery. If generation fails, it is fail-closed: no raw payload is returned, so rerun narrowly or with summarize=false. Check summary_effect.prev_chars, after_chars, and saved_chars; weak savings are a lesson to sharpen the retention specification or choose a narrower call.

Record a summary

Summarize only a completed result you have read. Pick targets from the current_tool_result_chars.top_results IDs returned by system(action="meta", input={}) (the per-result tail no longer lists them). In each item, tool_call_id is the producer call ID, not the visible _tool_call_id event reference; preserve any raw_locator or spill path. Preserve the conclusion, evidence or error, paths/URLs/IDs, validation status, risks, and next step—enough to decide whether the raw result must be reopened. Batch several already-digested results when useful:

{
  "action": "summarize",
  "input": {
    "items": [
      {"tool_call_id": "call_abc123", "summary": "Conclusion, evidence, anchors, risks, and next step."}
    ]
  },
  "reasoning": "Keep the evidence and recovery anchor needed by the next turn."
}

Recording is not durable-store authoring and does not rebuild. Do not summarize text that still needs exact inspection, quotation, patching, or comparison. If the result is still ambiguous, reopen it first.

Apply and recover

Prefer molt when history is substantial or noisy. Call one context(action="rebuild", input={}) only when recorded summaries or new canonical prompt sources must apply in this conversation and a molt is unsuitable; routine durable edits can wait for the next natural molt or reload. Reconstruction is deliberately delayed for cache efficiency; do not loop summarize/rebuild calls. If composition fails, the Context contract leaves summaries unapplied and reports the reconstruction error.

The rebuild receipt reports the provider round that requested the rebuild; post-rebuild context usage does not exist yet in that receipt. Observe it on the next provider round.

Keep the original tool_call_id in any summary. The normal fallback is a narrow JSONL search:

rg 'call_abc123' <workdir>/logs/events.jsonl

For a spilled result, preserve and reopen its tmp/tool-results/ path. Use system(action="manual", input={}, reasoning="load System routes") for settings, cache-miss budget ownership, and current automatic threshold details; this reference does not duplicate System-owned numbers or configuration procedures.

When the current pressure reminder is active, follow the resident cadence: prefer preparing a deliberate molt; at most one summarize pass (and a rebuild only under the rare exception above) may precede it, and never repeat either while context remains too high. The pressure reminder is a decision aid, not an automatic molt order. A deliberate molt still requires the Context manual's journal gate and concise successor handoff.