Interface: CompactionRecord

Defined in: src/core/agent/window/types.ts:298

What one OVER-BUDGET visit to summarizeOldest (what .compaction() configures) put in the ledger.

Properties

droppedObservations?

readonly optional droppedObservations?: readonly string[]

Defined in: src/core/agent/window/types.ts:207

The tools whose RESULTS left the window on this visit, in first-appearance order (9.57.0). Present only when at least one did.

This is the record's half of the sentence the drop notice says on the wire, and it is filed even when no notice was authored at all — a removal further into the window tells the model nothing, and then this is the only place the fact exists. Uncapped and unfiltered, because the record is not the wire: the notice prints at most four names and only plain identifiers, this prints every name exactly as the tool declared it.

The failure it exists for was measured: a model whose whats_here result had been evicted assembled a plausible id from an entity name it remembered, was refused, and spent actions on it — and nothing in the run said the evidence had gone.

Inherited from

WindowRecord.droppedObservations

droppedStandings?

readonly optional droppedStandings?: readonly object[]

Defined in: src/core/agent/window/types.ts:246

The standing of every tool result that LEFT the window on this visit, by id, as the model had declared it (9.102.0) — beside droppedObservations, which names the tools. Present only on an ARMED agent (.findings() configured) and only when at least one tool result left; an unarmed agent never carries this key.

standing absent is UNDECLARED — the model said nothing about that result — never 'open' and never a verdict the library inferred. It is the model's claim, and it is filed by the STAGE so a consumer-written strategy's record carries it too; a reader joining an evicted turn's hash on the receipt to the ledger fold at that stop gets the same answer.

Inherited from

WindowRecord.droppedStandings

iteration

readonly iteration: number

Defined in: src/core/agent/window/types.ts:181

ReAct iteration this visit belongs to.

Inherited from

WindowRecord.iteration

ledgerFacts?

readonly optional ledgerFacts?: WindowObservations

Defined in: src/core/agent/window/types.ts:232

What the ledger-fact pin did on this visit (9.102.0). Present only when it did something: held a turn, turned one away at the ceiling, or stood down — and only on an ARMED agent (.findings() configured), so an unarmed agent's records are the exact shape they were before.

The same WindowObservations shape as observations, a second block rather than a merged one: limit there is keepLastToolResults, here it is keepLedgerFacts, and a reader adding up what each pin cost needs them apart. The stand-down reads BOTH blocks: when the two previous visits removed nothing and named only pins ('last-tool-result' and/or 'ledger-fact'), the fact pins release for this visit and this block says so ({ pinned: [], yielded: 0, limit, standDown: true }).

Inherited from

WindowRecord.ledgerFacts

measuredTokens

readonly measuredTokens: number

Defined in: src/core/agent/window/types.ts:300

Adapter-reported input tokens of the last call — what tripped the check.

observations?

readonly optional observations?: WindowObservations

Defined in: src/core/agent/window/types.ts:217

What the last-tool-result pin did on this visit (9.57.0). Present only when it did something: held a turn, turned one away at the ceiling, or stood down.

See WindowObservations. An agent with keepLastToolResults: false — or one whose window held no pinnable tool result — never carries this key, so its records are the exact shape they were before 9.57.0.

Inherited from

WindowRecord.observations

overBudget

readonly overBudget: boolean

Defined in: src/core/agent/window/types.ts:304

True when the measurement was over budget (a fold was attempted).

refusals

readonly refusals: readonly WindowRefusal[]

Defined in: src/core/agent/window/types.ts:190

Every turn that refused to leave, named.

Inherited from

WindowRecord.refusals

removedMessageCount

readonly removedMessageCount: number

Defined in: src/core/agent/window/types.ts:185

How many messages left the window.

Inherited from

WindowRecord.removedMessageCount

removedStageIds

readonly removedStageIds: readonly string[]

Defined in: src/core/agent/window/types.ts:183

runtimeStageIds of the stages that appended the messages that left.

Inherited from

WindowRecord.removedStageIds

strategy

readonly strategy: string

Defined in: src/core/agent/window/types.ts:179

WindowStrategy.name of the strategy that decided — 'summarize-oldest', 'sliding-window', 'token-budget', or your own. Narrow on it.

Inherited from

WindowRecord.strategy

summarizerSkipped?

readonly optional summarizerSkipped?: boolean

Defined in: src/core/agent/window/types.ts:319

Present and true when the summarizer was deliberately NOT called this iteration (8.14.0): this exact span had already come back 'replacement-not-smaller', and the same span through the same summarizer gives the same answer. No call, so no summarizerTokens and no cost tick.

It is recorded rather than left out because a decision not to spend is still a decision. A record with this flag and one without are different facts, and a reader adding up an agent's fold attempts needs to see both.