A fresh context.
A note from yourself.

Let the agent prepare its own return

Engel's proposal starts from a simple expectation: the agent already has memory, daily notes, and tools for reading and writing them. Before clearing context, it should use those facilities to record unfinished requests, decisions, locations, and whatever else it will need to continue. The new tool can also carry a short personal handoff: a pointer to notes, the task it was doing, or simply “hi from your past self.”

The September 19 proposal replaced the original summary-style tool handoff with a full voluntary reset. On September 20, Engel selected a separate, explicitly configured hard-condensation fallback. That fallback produces a summary; the voluntary tool path still does not.

Four chances to prepare

The new condenser owns a configurable ascending list of warning thresholds. These defaults mean percentage of the resolved effective main-model input budget, counting the projected history, fixed context and tool declarations. They do not mean output tokens or the separate summarizer model's window.

Warnings are visible to the agent and ask it to save notes before calling condense. The agent alone chooses when to call the tool. Crossing a warning threshold, exceeding a local token estimate, or reaching an event-count threshold does not automatically condense or block an otherwise valid model request in this mode. The agent can also call the tool early.

The existing CondensationRequest is an internal trigger and is excluded from the model's view. A warning therefore needs explicit model-visible rendering. The separately configured provider-error fallback below handles emergency recovery; warning thresholds alone do not invoke it.

Two user-facing options

Keep the existing condenser choice and add a separate optional hard_condenser. The first selects the agent-controlled reset and its warnings. The second supplies the LLM summarizer used only when the main provider rejects a request for exceeding its context window.

{
  "condenser": {
    "condenser_kind": "agent_reset",
    "warning_thresholds": [0.75, 0.80, 0.85, 0.90]
  },
  "hard_condenser": {
    "condenser_kind": "llm_summarizing",
    "llm_profile_ref": "my-condenser-profile"
  }
}

Illustrative future configuration, not accepted by current settings. Field names are provisional; these example warning values are fractions of the effective input budget. The profile reference is a placeholder for an existing saved profile, not a secret or an inline model definition.

Omitting hard_condenser or setting it to null disables automatic fallback. The agent can still use its reset tool, but a provider context-window error cannot trigger an unconfigured summarizer. If hard condensation is configured, validate and resolve its own profile, preserve that binding across restore, and expose the existing bounded retry controls as advanced settings. keep_first, event-count and proactive token triggers are not controls for this full-view fallback.

The initial change applies these two options to agent-controlled mode. Existing standalone summarizer configurations retain their current behavior. Exact validation and host/API presentation belong in the contract and server tasks; no condenser pipeline is part of the proposal.

The agent calls condense

condense({
  message_to_future_self: "Hi from before the reset. Read the daily note I just saved; the next task is recorded there."
})

The parameter is optional free-form text. Field spelling is provisional; its meaning is the agent's own message, preserved verbatim. The tool does not write daily memory on the agent's behalf, discover missing notes, or ask another model to select what matters. Memory writes should have succeeded before it is called.

The tool records a correlated condensation request for the configured condenser. The new reset implementation handles that request at a safe step boundary and records a durable Condensation. Existing LLMSummarizingCondenser remains a separate supported configuration with its ordinary summarization behavior. Whether this new tool is exposed for that configuration is a contract decision; it must never silently discard a supplied future-self message.

Exactly what the next context starts with

1. Fixed context

   System prompt, identity, configured memory context, runtime/channel information, skills, and tool definitions, composed as usual.

2. 1 · user

   The agent triggered context condensation.
   Exact synthetic notice. Its provenance is the environment; it is not a new message from the human.

3. 2 · assistant

   The agent's genuine condense tool call, including its optional message. Preserve its real call ID and required reasoning/provenance.

4. 3 · tool

   The matching result: confirmation of the agent-triggered reset, recovery instructions, and the message to its future self.

5. Then · new input

   Any genuine user input that arrived after the model request which authored the reset. It was not available when the agent prepared the reset and must survive, in order.

The optional message block is omitted when no message was supplied. This wording guides the agent; it does not automatically send a message to the user. There is no CondensationSummaryEvent in the reset view.

New tool payloads inside existing event kinds

The host already has conversation.condense(); the Python SDK and our pinned port do not expose an agent-callable condensation tool. Existing request events carry no notes or originating tool references, and the current view reduces requests to a flag. A thin wrapper around that API would not implement this proposal.

Yes, the tool needs its own Action and Observation schemas. The proposed names CondenseAction and CondenseObservation are payload types, following existing tools such as SwitchLLMAction and SwitchLLMObservation. In TypeScript these are validated schemas and inferred types. They travel inside the existing ActionEvent/ObservationEvent envelopes; separate top-level CondenseActionEvent/CondenseObservationEvent kinds are not required. The previous table omitted this distinction. Typed, versioned reset metadata is also needed; its exact carrier remains a contract decision. A reset-aware projection must express both the voluntary notice/tool pair and the emergency notice/summary.

TRANSPILE treatment: tool/schema additions may be additive extensions; changed request handling, view replay and prompt ordering affect parity-governed behavior. Register the appropriate stable policy in TRANSPILE_CONTRACT.md and the canonical manifest, with tests and source evidence. Review and merge the SDK first, then vendor that exact commit into the server and update its contract.

If the agent runs out of room first

Selected September 20: only after catching an actual, recognized context-window error returned by the main provider, invoke the explicitly configured emergency LLMSummarizingCondenser.hard_context_reset(). The agent decides when to condense; warnings help it prepare. An estimated hard ceiling, event count, ignored warning or judgment about note readiness does not trigger fallback. Even an estimate at or above 100% is not a substitute for a provider error. Unrelated authentication, rate-limit and transport errors retain their own error handling.

The recovery context starts with:

1. Fixed context

   The normal system, identity, memory snapshot, runtime information, skills and tools.

2. 1 · user

   The environment triggered emergency context condensation after the model provider returned a context-window error.
   Earlier active history was replaced by the summary below. Your notes may be incomplete. Take a deep breath, read the summary and your notes, and check unfinished user requests before continuing. The user may not know this happened.

3. 2 · summary

   The fallback’s CondensationSummaryEvent, rendered in its existing user role after the environment warning. It is a generated summary, not a message authored by the agent’s past self.

4. Then · new input

   Genuine pending user input protected by the request boundary, followed by arrivals during fallback processing. Preserve its original text/media and order.

There is no retained historical keep_first prefix and no fabricated condense Action/Observation pair on this automatic path. The new notice is separate from the summary and must be replayed exactly once with truthful environment provenance.

“All events” describes the selected input View, not a promise of lossless full-text transfer: the current summarizer uses event previews, and retry clipping can shorten them further. The existing preview-loss bead remains relevant.

Decisions to settle before implementation

The agent owns the content of its notes; the harness cannot prove they are sufficient. If notes are missing, unreadable, or incomplete, the recovery instruction should lead to inspection or an honest question to the user. Reset cannot repair a fixed prompt that already exceeds the budget.

Tracked work, in dependency order

smolpaws-te15 remains the proposal's parent. All seven implementation tasks are open. Publishing this page and filing the beads does not implement or enable the feature.

Required evidence includes the complete warning → note write → tool → reset → note retrieval → continued task path; concurrent text/media; crash/restart at every persistence boundary; repeated reset; missing notes; mixed calls; exact native provider message ordering; and warning rearming. Also cover ignored warnings → actual provider context-window error → emergency fallback → notice/summary → continuation. Prove that estimates at/above 100%, high event counts and unrelated provider errors do not trigger fallback or preempt the agent’s decision. Cover reset-only configuration, missing fallback credentials, exhausted retries, oversized fixed context and user arrivals during fallback. Live cost and actual continuation must be reported separately from deterministic tests.

Related work remains separate: smolpaws-d3fw (existing 1M/400k budget snapshots), smolpaws-29dr (summary preview loss), smolpaws-neo (history invariants), and smolpaws-eh4z (Fast-Jev evaluation). None silently changes this proposal.

Source anchors and scope

This is a dated design proposal based on Engel's September 19 instructions and September 20 fallback clarification, the prior bead, and inspected source. Authorship and note selection are motivations to test; a message-role change alone has not been demonstrated to solve continuity.

• Python host condensation operation and registered SDK tools: no agent-callable condense tool in this inspected revision.
• Pinned Python condensation events: the original request, persisted condensation, and user-role summary representation.
• TypeScript view replay, tool/batch integrity, and step serialization.
• SDK TRANSPILE contract and server contract: classifications and ownership.
• Product context snapshots and existing standard condensation.
• September 20 source check: Python full-view hard reset, HARD requirement routing. This inspection does not advance the transpilation pin.
• Typed Action/Observation payload example and existing event envelopes.
• SmolPaws Beads owns task status; this page explains the proposal.