RFC 0111: Context Economy — Transcript Token Budget & Declared Summarization

Summary

A v1-conformant host MAY feed an unbounded transcript to each orchestrator turn: multi-agent-execution.md §"Stateful agent-loop lifecycle" defines transcriptWindow as an event-count bound and states "Absent ⇒ unbounded on the wire." This RFC adds an opt-in contextBudget object to the multi-agent execution-model capability that lets a host advertise (a) a per-turn token budget for the orchestrator transcript and (b) a declared summarization contract for turns evicted beyond that budget — making the per-iteration context contract observable and conformance-testable. It does not change any existing default: a host that advertises neither field behaves exactly as today.

Motivation

The protocol is byte-lean on the wire (content-free events, reference-by-id payloads, updates delta streams as the default) but says nothing normative about the prompt a host assembles and sends to the model each iteration. A third-party token-efficiency red-team (2026-06-15) identified this as the dominant cost: a conformant host re-sending the full transcript each turn grows per-turn cost O(turns) and cumulative run cost O(turns²). Two facts make this a spec concern rather than a pure implementation choice:

1. transcriptWindow already exists (RFC 0061, version >= 5) but bounds by event count, which does not bound tokens — a single large tool result blows the budget regardless of count. 2. The spec is silent on what a host does with evicted turns. Per COMPATIBILITY.md §4, a new normative requirement on previously-undefined behavior is additive — so the spec is the right place to define a budgeted, summarization-aware context contract that hosts can advertise and clients can rely on.

This RFC makes the contract declarable without forcing it on anyone: the "kill the unbounded default" framing from the red-team is deliberately not adopted, because flipping absent ⇒ unbounded to absent ⇒ bounded is a behavior change requiring the safety-fix process. Opt-in advertisement is the additive-safe path.

Proposal

Capability surface

Add an optional contextBudget object under multiAgent.executionModel (the same object that carries transcriptWindow). It is version >= 5 (RFC 0061 stateful loop) gated.

   "executionModel": {
     "type": "object",
     "properties": {
       "transcriptWindow": { "type": "integer", "minimum": 1 },
+      "contextBudget": {
+        "type": "object",
+        "additionalProperties": false,
+        "properties": {
+          "transcriptTokenBudget": {
+            "type": "integer", "minimum": 1,
+            "description": "RFC 0111. Max tokens of transcript the host feeds any single orchestrator turn. Advertise-and-honor; complements (does not replace) transcriptWindow. Absent ⇒ no token bound."
+          },
+          "tokenCounter": {
+            "type": "string",
+            "enum": ["o200k_base", "cl100k_base", "chars", "host-defined"],
+            "description": "RFC 0111. The unit transcriptTokenBudget is denominated in, so the bound is interpretable across hosts. REQUIRED when transcriptTokenBudget is present. Same enum as RFC 0113 memory.injectionBudget.tokenCounter — transcript + memory budgets denominate consistently."
+          },
+          "summarization": {
+            "type": "object",
+            "additionalProperties": false,
+            "properties": {
+              "supported": { "type": "boolean" },
+              "strategy": { "type": "string", "enum": ["sliding-window", "recursive", "map-reduce"] },
+              "keepLastTurns": { "type": "integer", "minimum": 0,
+                "description": "Turns kept verbatim at the head of the window; older turns MAY be replaced by a summary." }
+            },
+            "required": ["supported"]
+          }
+        }
+      }
     }
   }

Normative requirements

• Scope. contextBudget governs the RFC 0061 per-iteration orchestrator-loop transcript — multi-agent-execution.md §"Per-iteration state inputs" input 3, the SAME transcript transcriptWindow bounds — NOT a general chat-conversation history (RFC 0005) or any other prompt. A host whose multi-agent orchestrator loop does not run real model turns (e.g. a mock supervisor with no live inference) MUST NOT advertise contextBudget, exactly as it MUST NOT dishonestly advertise transcriptWindow. Budget-only advertisement (transcriptTokenBudget + tokenCounter without summarization.supported) is valid for a host that HAS a real orchestrator loop but does not summarize.
• A host advertising contextBudget.transcriptTokenBudget MUST NOT feed more than that many tokens of transcript to any single orchestrator turn, measured in the unit named by contextBudget.tokenCounter.
• A host advertising transcriptTokenBudget MUST also advertise tokenCounter; the budget is otherwise uninterpretable.
• transcriptTokenBudget and transcriptWindow MAY both be advertised; when both are present the host MUST honor whichever bound is tighter for a given turn.
• A host advertising contextBudget.summarization.supported: true MUST keep the most recent keepLastTurns turns verbatim and MAY replace older in-window turns with a host-produced summary. The active (most recent) turn MUST NOT be summarized.
• Replay determinism (normative). Per multi-agent-execution.md §"Per-iteration state inputs", the transcript is a deterministic, replay-reproducible iteration input — today a pure projection of the event-log tail. A host-produced summary is nondeterministic host output that breaks that purity, so it is governed exactly like a nondeterministic LLM envelope under RFC 0041 (version >= 5 implies the version >= 4 replay-determinism contract). Therefore:

- Each substitution MUST be recorded as a context.summarized event (payload below) whose summaryRef points at the persisted summary. - On a POST /v1/runs/{runId}:fork mode: replay, when a summarization point falls at sequence ≥ fromSeq, the host MUST reuse the recorded summary (resolved via summaryRef) and MUST NOT re-summarize. A host that re-summarizes and produces a different model-facing transcript is non-conformant — the direct analogue of RFC 0041 §"Envelope-refusal recovery in replay". - summaryRef is an artifactId (a run-produced artifact read via GET /v1/runs/{runId}/artifacts/{artifactId}), keeping the event content-free. By direct analogy to the MAE-3 memory carry-forward (multi-agent-execution.md §"Replay carry-forward"), the summary artifact MUST be persisted and readable as-of the iteration's event-log index; a host that cannot serve it at the requested fromSeq MUST refuse the fork (mirroring replay_memory_snapshot_unavailable). - Summarization that is not recorded this way breaks replay and MUST NOT be advertised under this capability.

Event-payload addition

schemas/run-event-payloads.schema.json gains context.summarized:

{
  "type": "context.summarized",
  "payload": {
    "iteration": 12,
    "replacedTurns": ["evt_0a1", "evt_0a2"],
    "summaryRef": "art_sum_7c",
    "tokenCounter": "o200k_base",
    "tokensBefore": 4120,
    "tokensAfter": 280
  }
}

summaryRef is a reference-by-id to the summary artifact (the summary text is not inlined on the wire — content-free event discipline). replacedTurns lists the event ids the summary stands in for, so a replay engine can reconstruct the exact transcript.

Conformance seam

The assembled transcript is host-internal and never crosses the wire, so conformance cannot observe it from discovery alone. This RFC adds one seam to host-sample-test-seams.md:

GET /v1/host/sample/agent/transcript-window?runId=…&iteration=N
→ { tokenCounter, tokenCount, eventIds: [...], summarizedRanges: [...] }

The seam returns the host's own accounting of what it fed the model that iteration. Conformance: drive enough turns that an unbounded transcript would exceed the budget, then assert (a) tokenCount ≤ transcriptTokenBudget; (b) the harness independently reads the events named in eventIds from the run event-log and re-computes their token sum in the advertised tokenCounter unit, confirming the host's reported tokenCount is internally consistent and ≤ budget; (c) eventIds are the recent tail (no older event included while a newer eligible one is dropped); and (d) every summarizedRanges entry has a matching context.summarized event.

Honest non-vacuity limit (documented, not a gap to close): the model-facing prompt is genuinely host-internal, so conformance verifies the host's declared transcript accounting is internally consistent and within budget — it cannot black-box-prove the host doesn't additionally feed off-seam content to the model. This is the inherent ceiling of governing host-internal assembly via a test seam; the capability is advertise-and-attest, and the seam proves the accounting is consistent + bounded, not that the prompt is nothing more. A host that lies in the seam fails the cross-check in (b)/(c); a host that lies past the seam is outside what any wire/seam witness can see.

Examples

Positive — host advertises a 4 000-token budget with sliding-window summarization keeping the last 3 turns verbatim:

{ "multiAgent": { "executionModel": { "version": 5,
  "contextBudget": { "transcriptTokenBudget": 4000, "tokenCounter": "o200k_base",
    "summarization": { "supported": true, "strategy": "sliding-window", "keepLastTurns": 3 } } } } }

Negative — transcriptTokenBudget without tokenCounter (uninterpretable budget) → fails schema (tokenCounter required when budget present, enforced by an if/then clause) and fails conformance.

Compatibility

Additive. Backward-compat clauses:

• contextBudget is a new optional object; a host that omits it behaves exactly as today (transcriptWindow semantics, including absent ⇒ unbounded, are untouched).
• context.summarized is a new event type per COMPATIBILITY.md §2.1; consumers that do not recognize it ignore it. Server-emitted event schemas are open (RFC 0094), so adding it does not break schema-validating clients.
• No existing field changes type, optionality, or meaning. No default flips. The "bound the transcript by default" option was explicitly rejected as a §4 behavior change (see Alternatives).

Conformance

• Existing: multi-agent-* scenarios cover the execution loop and transcriptWindow count semantics; none assert a token bound.
• New: context-budget-transcript-bound.test.ts, gated on capabilityFamily(d,'multiAgent')?.executionModel?.contextBudget?.transcriptTokenBudget being present (soft-skip on absence). Drives a multi-turn run, reads the host-sample transcript seam each iteration, asserts tokenCount ≤ budget and that keepLastTurns recent turns appear verbatim. A second scenario context-summarization-replay.test.ts (gated on summarization.supported) replays the run and asserts the reconstructed transcript matches, exercising the context.summarized → summaryRef path.

Alternatives considered

1. Flip the default (absent ⇒ bounded). The red-team's headline. Rejected: per COMPATIBILITY.md §4 this changes observed behavior on hosts that never advertised a window → safety-fix process, not additive. Not worth a v1.x break for an efficiency optimization hosts can already perform. 2. Leave it host-internal (do nothing). Hosts can already token-budget their own prompt assembly with zero conformance impact (Tier-A in the red-team). Rejected as the whole answer because it leaves the contract unobservable: a client cannot tell whether a host bounds context, so cross-host portability of "this agent stays cheap on long runs" is impossible. The capability advertisement is the value-add. 3. Reuse RFC 0084 budget-policy maxTokens. Rejected: maxTokens is a run-wide spend ceiling that fails or interrupts the run on exhaustion; it is not a per-turn context-shaping budget and has no summarization semantics.

Unresolved questions

1. Should tokenCounter permit provider-native counters (e.g., a named Anthropic/OpenAI tokenizer) or stay at a small portable enum? The enum keeps conformance deterministic but may diverge from a host's real provider count. 2. Is keepLastTurns the right knob, or should the verbatim floor be expressed in tokens (keepLastTokens) to compose with transcriptTokenBudget? 3. Does context.summarized belong in the normative core event set, or behind the same version >= 5 gate as the rest of the stateful loop? 4. Should the transcript seam be normative here or deferred to a host-sample-only test seam (as drafted)?

Implementation notes (non-normative)

• Tier-A host patches that need no spec change and SHOULD ship independently of this RFC: provider context-caching (byte-identical, front-loaded system prompt), per-task tool sub-allowlisting, gzipped SSE bus, stripping optional event meta. This RFC governs only the declarable transcript contract; the rest is host craft.
• Sequence after the schema lands: reference host (sqlite) implements the seam + a sliding-window summarizer first; it is the simplest strategy and exercises the context.summarized replay path.
• Reconcile with openwop-app ADR 0148 A1 (host-internal token-budgeted transcript assembly + transcriptWindow advert): A1 is the Tier-A producer of exactly the budgeted transcript that contextBudget advertises and the seam witnesses — the host's A1 assembly path emits the context.summarized event + the seam reads its accounting. One model, not two.
• tokenCounter shares the RFC 0113 enum (incl. chars), so a host budgets transcript (0111) and memory (0113) in one consistent unit.

Acceptance criteria

• [x] Spec text merged (multi-agent-execution.md §contextBudget + host-sample-test-seams.md seam) — landed with the Active surface 2026-06-26 (CHANGELOG [Unreleased]).
• [x] capabilities.schema.json + run-event-payloads.schema.json updated — contextBudget (capabilities) + context.summarized (run-event-payloads) both on main.
• [x] context-budget-transcript-bound.test.ts + context-summarization-replay.test.ts in @openwop/openwop-conformance — both present; capability-gated + driver-based, so they exercise non-vacuously only against a host advertising contextBudget on a real orchestrator loop.
• [x] CHANGELOG entry under the next minor — [Unreleased] "RFCs 0111 / 0114 / 0116 — token-economy surfaces (Active, 2026-06-26)".
• [x] A reference host advertises contextBudget and passes both scenarios non-vacuously, or the RFC explicitly defers reference-host implementation — this RFC takes the deferral: no current steward reference host runs a real orchestrator-loop model turn to witness against (the mock supervisor MUST NOT advertise contextBudget per §"Scope"). Both conformance scenarios soft-skip on non-advertising hosts; the non-vacuous witness comes from a host that runs real orchestrator turns (the Active → Accepted graduation gate).

References

• spec/v1/multi-agent-execution.md §"Stateful agent-loop lifecycle" (transcriptWindow, line 238).
• spec/v1/budget-policy.md (RFC 0084) — run-wide spend, contrasted in Alternatives.
• COMPATIBILITY.md §4 (behavior-only changes) — the §4 table is why the default cannot flip.
• RFC 0061 (stateful agent-loop lifecycle), RFC 0041 (replay determinism), RFC 0094 (schema closure).
• Prior art: Anthropic prompt caching; LangGraph checkpoint/summarization; sliding-window context management.
• Sibling RFCs in this set: 0112 (compact tools), 0113 (memory budget), 0114 (A2UI deltas), 0115 (transport economy), 0116 (prompt-prefix cache).