@aragon-agent/cli 0.6.1 → 0.6.3

package/dist/compaction/compactor.js

@@ -0,0 +1,689 @@
+/**
+ * `Compactor` — the `ContextManager` implementation (context-auto-compaction
+ * §3.4 / §3.6 / §3.7 / §3.8).
+ *
+ * ASCII ONLY: `src/compaction/**` is inside the glyph scanner's scope.
+ *
+ * THIS IS THE POLICY HALF OF THE FEATURE. The engine owns exactly one thing -
+ * structural validity - and every other decision is here: when to fire, what to
+ * keep, which model summarizes, what the prompt asks for, what happens when the
+ * call fails, and when to stop spending.
+ */
+import { planCompaction, } from '@aragon-agent/core';
+import { getLogger } from '../logging/logger.js';
+import { resolveFastTier } from '../fast/resolve.js';
+import { COMPACTION_LIMITS } from './limits.js';
+import { attemptTailRelief, reliefSummary, tailBudgetTokens } from './tail-budget.js';
+import { buildDigest, isPriorSummaryBlock } from './digest.js';
+import { buildAnchor, buildCompactedBlock, countProtectedPrefix, readGeneration, truncationBody, } from './summary-prompt.js';
+import { accumulateUsage, errorText, runSummarizeCall } from './summarize-call.js';
+import { computeEstimateOffset, computePressure, shouldCompactAt } from './pressure.js';
+export class Compactor {
+    deps;
+    log = getLogger().child('compaction');
+    // --- Per-run state (reset on `agent_start`) ---
+    compactionsThisRun = 0;
+    lastCompactionTurn = -Infinity;
+    // --- Per-session state ---
+    compactions = 0;
+    tokensReclaimed = 0;
+    sessionUsage = { inputTokens: 0, outputTokens: 0 };
+    pricingUnknown = false;
+    inFlight = false;
+    consecutiveNoProgress = 0;
+    selfDisabled = false;
+    selfDisabledReason;
+    pendingManual = null;
+    generation = 0;
+    /** Calibration for every ESTIMATED figure (D-23). Recomputed on every turn_end. */
+    estimateOffset;
+    /** The last pressure computed, so the status bar and `/compact status` agree. */
+    lastPressure = null;
+    /**
+     * The history LENGTH the measurement in `lastUsage` covered (hardening W1).
+     *
+     * `undefined` DISABLES THE DELTA and reproduces round 1's arithmetic exactly.
+     * Every host-side mutation that is not an append must clear it - see
+     * `invalidateMeasurement`.
+     */
+    measuredPrefixLength;
+    /**
+     * The messages the last splice dropped, held ONLY until the wiring collects
+     * them for the archive (W4 / RV-3).
+     *
+     * NOT ON `CompactionRecord`: that record is dispatched into the reducer and
+     * retained in view state for the card's lifetime, so putting a megabyte of
+     * history on it would park the dropped conversation in React state - precisely
+     * the memory this compaction just reclaimed.
+     *
+     * SET ONLY WHEN ARCHIVING IS ON, so with the key off the slice is never
+     * retained and W4 costs nothing.
+     */
+    lastDropped = null;
+    // --- Abort plumbing for the summarization call (see `summarize-call.ts`) ---
+    callAbort = null;
+    callAbortReason = null;
+    /** Guard 2's ceiling for this instance (hardening §3.4.2 mechanism B / DH-16). */
+    maxPerRun;
+    constructor(deps, opts) {
+        this.deps = deps;
+        this.maxPerRun = opts?.maxPerRun ?? COMPACTION_LIMITS.maxPerRun;
+    }
+    // =========================================================================
+    // Lifetime
+    // =========================================================================
+    /** `agent_start` — the per-run cap resets, the session totals do not. */
+    onRunStart() {
+        this.compactionsThisRun = 0;
+        this.lastCompactionTurn = -Infinity;
+    }
+    /**
+     * `turn_end` — recompute the calibration offset (D-23).
+     *
+     * RE-MEASURED EVERY TURN, not once, so it tracks a toolset that changed
+     * mid-session (`/tools`, a skill that registers one).
+     */
+    onTurnEnd(usage, messages, systemPrompt) {
+        this.estimateOffset = computeEstimateOffset(usage, messages, systemPrompt);
+        // THE SAME ARRAY `computeEstimateOffset` JUST MEASURED, so the two numbers
+        // describe the same request by construction (hardening §3.2.3).
+        this.measuredPrefixLength = messages.length;
+    }
+    /**
+     * Any host-side history mutation that is NOT an append (hardening §3.2.3).
+     *
+     * `/clear`, `/resume`, and the splice itself: after any of them a recorded
+     * index means nothing, and the delta would be measured against the wrong
+     * array. `estimateAppendedTokens` bounds-checks as well, so a MISSED call here
+     * costs accuracy - the number degrades to round 1's - and never correctness.
+     */
+    invalidateMeasurement() {
+        this.measuredPrefixLength = undefined;
+    }
+    /**
+     * The dropped messages, released as they are returned (W4 / RV-3).
+     *
+     * `take` RATHER THAN `get`: the reference is dropped on read, so a wiring that
+     * declines to archive (key off, or `applied: false`) cannot pin it and a second
+     * read cannot double-write. `settlePending` calls this on EVERY verdict and
+     * discards on the ones that do not archive, which is what guarantees the array
+     * cannot outlive one compaction regardless of which branch was taken.
+     */
+    takeLastDropped() {
+        const dropped = this.lastDropped;
+        this.lastDropped = null;
+        return dropped;
+    }
+    /** Cancel an in-flight summarization (Esc, `dispose`). */
+    abort() {
+        if (this.callAbort) {
+            this.callAbortReason = 'cancelled';
+            this.callAbort.abort();
+        }
+    }
+    // =========================================================================
+    // The manual queue (§4.4 / D-17)
+    // =========================================================================
+    /**
+     * Ask for a compaction at the next turn boundary.
+     *
+     * THE BOUNDARY IS THE ONLY MOMENT THE ENGINE IS SINGLE-THREADED WITH RESPECT
+     * TO THE HISTORY, by construction. Replacing `messages` under a live loop is a
+     * race; a flag `shouldCompact` honours there is not.
+     */
+    queueManual(instructions) {
+        this.pendingManual = instructions ? { instructions } : {};
+    }
+    hasPendingManual() {
+        return this.pendingManual !== null;
+    }
+    clearPendingManual() {
+        this.pendingManual = null;
+    }
+    // =========================================================================
+    // ContextManager — the sync gate
+    // =========================================================================
+    /**
+     * SYNCHRONOUS AND CHEAP. Two deliberate effects: the pressure cache, and
+     * `deps.onPressure` (hardening DH-15).
+     *
+     * THE MEASUREMENT IS TAKEN ONCE, ABOVE THE GUARDS, AND REPORTED
+     * UNCONDITIONALLY. Below them it is skipped on five paths — the per-run cap,
+     * the cooldown, manual, overflow and guard 4 — and TWO OF THOSE NEVER CLEAR
+     * WITHIN A RUN. A gauge fed from the measurement would therefore go dark
+     * exactly when compaction is exhausted and the window is still filling, which
+     * is the one state where the corrected number is the user's only warning. It
+     * would silently revert to the measured base — the very figure W1 exists
+     * because it under-reports.
+     *
+     * `!config.enabled` STAYS THE FIRST STATEMENT, so a session that turned the
+     * feature off pays nothing at all and AC-1's byte-identity claim is untouched.
+     */
+    shouldCompact(probe) {
+        const config = this.deps.getConfig().compaction;
+        if (!config.enabled)
+            return false;
+        // THE HISTORY IS PASSED IN, NEVER OMITTED (§3.4.2). Without it the estimate
+        // branch of `computePressure` measures an EMPTY array and reports 0 %, so a
+        // resumed 180 k-token session — where `lastUsage` is `undefined` by
+        // definition and this branch is the only one available — would never trigger.
+        // The measured branch reads `messages` only to estimate what was appended
+        // after the measurement (hardening §3.2.2); the slice is one turn wide.
+        const pressure = this.measure({
+            ...(probe.lastUsage ? { lastUsage: probe.lastUsage } : {}),
+            messages: this.deps.getMessages(),
+            systemPrompt: this.deps.getSystemPrompt(),
+        });
+        this.deps.onPressure?.(pressure);
+        const manual = this.pendingManual !== null;
+        // Guard 2 — the per-run cap. BOTH TRIGGERS, HARD. It is the bound that makes
+        // every other guard's failure mode "spends a little" rather than "spends
+        // forever". Read from the INSTANCE, not the module: a child manager gets a
+        // tighter one (W3 / DH-16).
+        if (this.compactionsThisRun >= this.maxPerRun)
+            return false;
+        // Guard 1 — cooldown, in `turnIndex` units. A compaction that just ran needs
+        // a real turn before its effect is measurable, because `lastUsage` is still
+        // the pre-compaction number until the next `turn_end`.
+        //
+        // BYPASSED FOR OVERFLOW, and that is not a courtesy: the cooldown cannot be
+        // allowed to block the only path that makes an already-refused request
+        // sendable. A manual request is a user instruction and bypasses it too.
+        const cooled = probe.turnIndex - this.lastCompactionTurn >= COMPACTION_LIMITS.minTurnsBetween;
+        if (probe.trigger !== 'overflow' && !manual && !cooled)
+            return false;
+        // An explicit `/compact` is a user instruction: it does not consult the
+        // threshold, and guard 4 does not suppress it.
+        if (manual)
+            return true;
+        // The reactive path is answering a request the provider has ALREADY REFUSED,
+        // so it never consults occupancy — and guard 4 must not be able to switch it
+        // off, which is the entire answer to R-5 (D-24 / P1-13).
+        if (probe.trigger === 'overflow')
+            return true;
+        // Guard 4 — self-disable suppresses the PRESSURE trigger only.
+        if (this.selfDisabled)
+            return false;
+        return shouldCompactAt(pressure, config.threshold, {
+            maxOutputTokens: this.effectiveMaxOutputTokens(),
+        });
+    }
+    // =========================================================================
+    // ContextManager — the work
+    // =========================================================================
+    async compact(ctx) {
+        const manual = this.pendingManual;
+        const instructions = manual?.instructions;
+        this.pendingManual = null;
+        const uiTrigger = manual ? 'manual' : ctx.trigger;
+        const config = this.deps.getConfig().compaction;
+        const summarizer = this.resolveSummarizer();
+        this.compactionsThisRun += 1;
+        this.lastCompactionTurn = ctx.turnIndex;
+        this.compactions += 1;
+        this.inFlight = true;
+        const index = this.compactions;
+        this.deps.emit({
+            type: 'compaction_start',
+            index,
+            trigger: uiTrigger,
+            model: summarizer?.ref.modelId ?? '',
+        });
+        const started = Date.now();
+        try {
+            return await this.runCompaction({
+                ctx,
+                index,
+                uiTrigger,
+                config,
+                summarizer,
+                ...(instructions ? { instructions } : {}),
+                started,
+            });
+        }
+        finally {
+            this.inFlight = false;
+            this.callAbort = null;
+            this.callAbortReason = null;
+        }
+    }
+    // =========================================================================
+    // Internals
+    // =========================================================================
+    async runCompaction(args) {
+        const { ctx, config, summarizer } = args;
+        const before = ctx.messages;
+        const protectedPrefix = countProtectedPrefix(before);
+        // THE ONE LEVER THE GUARDS DO NOT HAVE (D-24). A tail that alone exceeds the
+        // window is the single failure mode none of the four guards can fix — they
+        // can only stop spending on it. An overflow attempt halves `keepRecentTurns`
+        // FOR ITSELF, in memory, never persisted: it is applied exactly where the
+        // provider has PROVEN the request impossible.
+        const keepRecentTurns = ctx.trigger === 'overflow'
+            ? Math.max(1, Math.floor(config.keepRecentTurns / 2))
+            : config.keepRecentTurns;
+        const plan = planCompaction(before, { keepRecentTurns, protectedPrefix });
+        if (!plan) {
+            // THE CASE THAT KILLS A RUN TODAY (hardening §3.3.3 / W2). "Nothing to
+            // drop" and "the tail is too big" are the SAME outcome in round 1, and both
+            // are terminal: every rung of the ladder operates on the head, and there is
+            // no head left. Relief is the one lever that still applies, so the
+            // `plan === null` branch tries it before conceding.
+            const reliefOnly = this.relieveInPlace(before, protectedPrefix, ctx.systemPrompt);
+            if (reliefOnly) {
+                this.checkProgress(ctx, args.uiTrigger, reliefOnly.messages);
+                return this.finish({ ...args, tailRelief: reliefSummary(reliefOnly) }, {
+                    action: 'replace',
+                    // `truncated` ACROSS THE PORT (DH-5 / DH-20). The engine's question
+                    // is only "was this history reduced without a summary", and the
+                    // answer is yes. `finish` maps this reason to the CLI's own
+                    // `'relieved'`, which is what stops the card claiming a summary that
+                    // never ran.
+                    mode: 'truncated',
+                    messages: reliefOnly.messages,
+                    reason: 'tail_relief_only',
+                }, { droppedMessages: 0 }, keepRecentTurns);
+            }
+            this.noteNoProgress(args.uiTrigger, 'nothing_to_drop');
+            return this.finish(args, { action: 'keep', reason: 'nothing_to_drop' }, null, keepRecentTurns);
+        }
+        const anchor = buildAnchor(before);
+        if (!anchor) {
+            // A history with no `user` message at all cannot have a goal anchor, and
+            // D-6 makes the anchor non-negotiable. Refusing is the honest outcome.
+            return this.finish(args, { action: 'keep', reason: 'no_anchor' }, null, keepRecentTurns);
+        }
+        const head = before.slice(protectedPrefix, plan.cutIndex);
+        const priorBlock = before.slice(0, protectedPrefix).find((m) => isPriorSummaryBlock(m));
+        // The prior block is carried into the DIGEST rather than the tail, so the
+        // summarizer merges it forward (D-22).
+        const digestHead = priorBlock ? [priorBlock, ...head] : head;
+        const digest = buildDigest({
+            head: digestHead,
+            ...(summarizer
+                ? { summarizerWindow: this.deps.getModelInfoFor(summarizer.ref).contextWindow }
+                : {}),
+        });
+        const ladder = await this.runLadder({
+            ctx,
+            summarizer,
+            digest: digest.text,
+            hasPriorSummary: priorBlock !== undefined,
+            ...(args.instructions ? { instructions: args.instructions } : {}),
+        });
+        if (ladder.kind === 'aborted') {
+            return this.finish(args, { action: 'keep', reason: 'aborted' }, null, keepRecentTurns);
+        }
+        if (ladder.kind === 'failed' && config.onFailure === 'stop') {
+            // Rung 4. The run proceeds and will probably hit `context_overflow`, which
+            // is what the user asked for by setting this.
+            this.deps.notify('error', `Context compaction failed (${ladder.reason}) and compaction.onFailure is "stop". ` +
+                'The next request may exceed the context window. Try /compact again, or /clear, or /reset.');
+            return this.finish(args, { action: 'keep', reason: `summarize_failed: ${ladder.reason}` }, null, keepRecentTurns);
+        }
+        const truncated = ladder.kind === 'failed';
+        const mode = truncated ? 'truncated' : 'summarized';
+        const generation = readGeneration(before) + 1;
+        const block = buildCompactedBlock({
+            summary: truncated ? truncationBody(plan.droppedMessages) : ladder.text,
+            replaced: plan.droppedMessages,
+            turns: countUserTurns(head),
+            anchor: anchor.shape,
+            generation,
+            truncated,
+        });
+        const spliced = [anchor.message, block, ...before.slice(plan.cutIndex)];
+        // THE TAIL'S OWN SIZE IS ONLY KNOWABLE AFTER THE HEAD IS GONE. Projecting it
+        // before the splice would be projecting against a number that is about to
+        // change by 80 %. `from: 2` because the anchor and the block are never
+        // eligible: the anchor is the user's own goal statement (D-6 makes it
+        // non-negotiable) and the block is the summary this compaction just paid for.
+        const relief = this.relieveInPlace(spliced, 2, ctx.systemPrompt);
+        const messages = relief ? relief.messages : spliced;
+        // ARCHIVED BEFORE THE VERDICT, HELD UNTIL THE WIRING COLLECTS IT (W4 / RV-3).
+        // `head` is the only place the dropped messages exist as an array, and this
+        // function has returned long before `settlePending` runs. Retained ONLY when
+        // the key is on, so `compaction.archive: false` never holds a reference.
+        if (config.archive)
+            this.lastDropped = [...head];
+        if (truncated) {
+            // Rung 3 is a VISIBLE, BOUNDED DEGRADATION rather than silent data loss:
+            // it says so on the card, on the JSON event stream, and inside the block
+            // the model itself reads.
+            this.deps.notify('warn', `Context compaction could not summarize (${ladder.reason}); ` +
+                `${plan.droppedMessages} earlier messages were dropped without a summary.`);
+        }
+        this.generation = generation;
+        // GUARD 3 SEES THE RELIEVED HISTORY, so relief counts as progress for free -
+        // `checkProgress` measures the final array, which is the one being adopted.
+        // A test pins it (AC-H7), because "for free" is the kind of property a
+        // refactor breaks silently.
+        this.checkProgress(ctx, args.uiTrigger, messages);
+        return this.finish({ ...args, ...(relief ? { tailRelief: reliefSummary(relief) } : {}) }, {
+            action: 'replace',
+            mode,
+            messages,
+            ...(truncated ? {} : { summary: ladder.text }),
+            ...(truncated ? { reason: `truncated: ${ladder.reason}` } : {}),
+        }, { droppedMessages: plan.droppedMessages, summary: truncated ? undefined : ladder.text }, keepRecentTurns);
+    }
+    /** Tail relief against this session's window (hardening §3.3.3 / W2). */
+    relieveInPlace(messages, from, systemPrompt) {
+        return attemptTailRelief({
+            messages,
+            from,
+            systemPrompt,
+            budget: tailBudgetTokens({
+                contextWindow: this.modelInfo().contextWindow,
+                maxOutputTokens: this.effectiveMaxOutputTokens(),
+            }),
+            log: (event, fields) => this.log.info(event, fields),
+        });
+    }
+    /**
+     * The failure ladder (§3.7).
+     *
+     * Rung 1 is the fast tier (or the main model); rung 2 is ONE retry on the MAIN
+     * model when rung 1 used the fast tier, because "a fast model that cannot
+     * summarize" is the most likely single failure. Rungs 3 and 4 are the caller's,
+     * because they depend on `onFailure`.
+     */
+    async runLadder(args) {
+        const { summarizer } = args;
+        if (!summarizer)
+            return { kind: 'failed', reason: 'no_summarizer_model' };
+        const first = await this.callOnce(summarizer.ref, args);
+        if (first.kind === 'ok' || first.kind === 'aborted')
+            return first;
+        // Rung 2 — retry, on the MAIN model when rung 1 was the fast tier.
+        const mainRef = this.mainRef();
+        const retryRef = summarizer.fromFastTier && mainRef.modelId !== summarizer.ref.modelId ? mainRef : summarizer.ref;
+        this.log.warn('summarize_retry', { reason: first.reason, model: retryRef.modelId });
+        const second = await this.callOnce(retryRef, args);
+        if (second.kind === 'ok' || second.kind === 'aborted')
+            return second;
+        return { kind: 'failed', reason: second.reason };
+    }
+    async callOnce(ref, args) {
+        const apiKey = this.deps.getApiKey(ref.providerId);
+        if (!apiKey)
+            return { kind: 'failed', reason: 'no_api_key' };
+        if (args.ctx.signal.aborted)
+            return { kind: 'aborted' };
+        // The controller is owned HERE, not in `summarize-call.ts`, so the caller can
+        // stamp WHICH of "timeout" and "cancelled" happened before aborting — the
+        // distinction the retry layer flattens away (see that file's header).
+        const controller = new AbortController();
+        this.callAbort = controller;
+        this.callAbortReason = null;
+        const onRunAbort = () => {
+            this.callAbortReason = 'cancelled';
+            controller.abort();
+        };
+        args.ctx.signal.addEventListener('abort', onRunAbort, { once: true });
+        try {
+            const outcome = await runSummarizeCall({
+                ref,
+                apiKey,
+                digest: args.digest,
+                hasPriorSummary: args.hasPriorSummary,
+                ...(args.instructions ? { instructions: args.instructions } : {}),
+                signal: controller.signal,
+                now: Date.now(),
+                complete: (providerId, request) => this.deps.complete(providerId, request),
+                onTimeout: () => {
+                    this.callAbortReason = 'timeout';
+                    controller.abort();
+                },
+            });
+            if (!outcome.ok) {
+                if (this.callAbortReason === 'cancelled' || args.ctx.signal.aborted)
+                    return { kind: 'aborted' };
+                const reason = this.callAbortReason === 'timeout' ? 'timeout' : errorText(outcome.error);
+                return { kind: 'failed', reason };
+            }
+            this.noteUsage(ref, outcome.usage);
+            if (outcome.text.trim().length === 0)
+                return { kind: 'failed', reason: 'empty_summary' };
+            return { kind: 'ok', text: outcome.text.trim() };
+        }
+        finally {
+            args.ctx.signal.removeEventListener('abort', onRunAbort);
+        }
+    }
+    /**
+     * Guard 3 — the progress requirement (§3.8).
+     *
+     * ADVISORY ON OVERFLOW, NEVER A VETO (D-24 / P1-13). The arithmetic that makes
+     * this guard sensible for the proactive trigger INVERTS on the reactive one: a
+     * pressure compaction that reclaims little has spent money to buy two turns and
+     * should stop, while an overflow compaction that reclaims little is the
+     * difference between a continuing run and a dead one, because the provider has
+     * already refused the request.
+     *
+     * The projection is ESTIMATED — no usage exists for a history that has not been
+     * sent yet — and is therefore CALIBRATED with `estimateOffset` (D-23).
+     *
+     * IT TAKES THE *UI* TRIGGER, not `ctx.trigger`. A queued `/compact` reaches the
+     * loop as an ordinary `pressure` checkpoint (that is what the queue IS), so
+     * charging its result to guard 4's counter would let a user who types
+     * `/compact` twice on a short conversation self-disable the proactive trigger
+     * for the rest of the session — and be told, wrongly, that their recent turns
+     * exceed the threshold. `uiTrigger` is the only value that knows the difference.
+     */
+    checkProgress(ctx, uiTrigger, after) {
+        const config = this.deps.getConfig().compaction;
+        const info = this.modelInfo();
+        const projected = computePressure({
+            messages: after,
+            systemPrompt: ctx.systemPrompt,
+            contextWindow: info.contextWindow,
+            windowKnown: true,
+            ...(this.estimateOffset !== undefined ? { estimateOffset: this.estimateOffset } : {}),
+        });
+        const target = config.threshold - COMPACTION_LIMITS.minReclaimRatio;
+        const madeProgress = projected.ratio < target;
+        this.log.info('compaction_progress', {
+            trigger: uiTrigger,
+            projectedRatio: Number(projected.ratio.toFixed(3)),
+            target: Number(target.toFixed(3)),
+            madeProgress,
+        });
+        if (madeProgress) {
+            this.consecutiveNoProgress = 0;
+            return;
+        }
+        this.noteNoProgress(uiTrigger, 'insufficient_reclaim');
+    }
+    /**
+     * Guard 4 — stuck detection (§3.8).
+     *
+     * PRESSURE ONLY, AND `pressure` HERE MEANS THE *UI* TRIGGER. A session that
+     * self-disables still survives an overflow rather than dying at the next
+     * request, and if the tail really is too large the overflow attempt is the one
+     * that can do something about it (D-24). A `manual` compaction is excluded for
+     * a different reason: this guard exists to stop UNREQUESTED spend, and
+     * `/compact` on a conversation with nothing worth dropping is a user who asked
+     * — twice — not a runaway loop.
+     */
+    noteNoProgress(trigger, reason) {
+        if (trigger !== 'pressure')
+            return;
+        this.consecutiveNoProgress += 1;
+        if (this.consecutiveNoProgress < COMPACTION_LIMITS.stuckLimit || this.selfDisabled)
+            return;
+        this.selfDisabled = true;
+        this.selfDisabledReason = reason;
+        this.log.warn('compaction_self_disabled', { reason });
+        // THE WORDING NAMES THE ACTUAL CAUSE AND THE ACTUAL REMEDIES. A notice that
+        // says only "compaction is off" teaches the user that the feature is broken;
+        // this one tells them which number to change.
+        this.deps.notify('warn', 'Auto-compaction is off for this session: the most recent ' +
+            `${this.deps.getConfig().compaction.keepRecentTurns} turns alone exceed the threshold. ` +
+            'Lower compaction.keepRecentTurns, or use /clear or /reset.');
+    }
+    finish(args, outcome, applied, keepRecentTurnsUsed) {
+        const before = args.ctx.messages.length;
+        const after = outcome.action === 'replace' ? outcome.messages.length : before;
+        const tokensBefore = this.lastPressure?.occupied ?? 0;
+        const record = {
+            index: args.index,
+            trigger: args.uiTrigger,
+            // `'relieved'` IS A CLI-ONLY WORD (RV-5 / DH-20). The port reports
+            // `'truncated'` for a relief-only compaction, which is accurate for the
+            // engine's narrower question; the card's `summarized N messages with
+            // <model>` line would otherwise render "summarized 0 messages with
+            // claude-haiku-4-5" for a compaction in which no model was called.
+            mode: reliefOnly(outcome) ? 'relieved' : outcome.action === 'replace' ? outcome.mode : 'none',
+            // OPTIMISTIC AND CORRECTED BY THE WIRING (§3.2.1). Only the ENGINE knows
+            // whether the splice passed the structural gate, and it says so on
+            // `compaction_end`; the wiring rewrites this before the card settles.
+            applied: outcome.action === 'replace',
+            ...(outcome.action === 'keep' ? { reason: outcome.reason } : {}),
+            ...(outcome.action === 'replace' && outcome.reason ? { reason: outcome.reason } : {}),
+            messagesBefore: before,
+            messagesAfter: after,
+            tokensBefore,
+            tokensAfter: 0, // filled by the wiring from the engine's own estimator
+            ...(applied?.summary ? { summary: applied.summary } : {}),
+            // NO MODEL ON A RELIEF-ONLY RECORD, because no model was called. The card
+            // names the summarizer from this field, and naming one here would be the
+            // same claim `mode: 'relieved'` exists to stop making.
+            model: reliefOnly(outcome) ? '' : args.summarizer?.ref.modelId ?? '',
+            durationMs: Date.now() - args.started,
+            ...(args.ctx.trigger === 'overflow' ? { keepRecentTurnsUsed } : {}),
+            ...(args.tailRelief ? { tailRelief: args.tailRelief } : {}),
+        };
+        this.log.info('compaction', {
+            trigger: args.uiTrigger,
+            before,
+            after,
+            droppedMessages: applied?.droppedMessages ?? 0,
+            mode: record.mode,
+            model: record.model,
+            durationMs: record.durationMs,
+        });
+        this.deps.emit({ type: 'compaction_end', record });
+        return outcome;
+    }
+    // =========================================================================
+    // Reporting helpers
+    // =========================================================================
+    /**
+     * The pressure the trigger and the gauge both read.
+     *
+     * CACHED on every evaluation, so `/compact status` reports the same number the
+     * last decision used rather than recomputing one that has since moved.
+     */
+    measure(probe) {
+        const info = this.modelInfo();
+        const pressure = computePressure({
+            ...(probe.lastUsage ? { lastUsage: probe.lastUsage } : {}),
+            messages: probe.messages ?? [],
+            systemPrompt: probe.systemPrompt ?? '',
+            contextWindow: info.contextWindow,
+            windowKnown: this.deps.isPricedModel({
+                providerId: this.mainRef().providerId,
+                modelId: this.mainRef().modelId,
+            }),
+            ...(this.estimateOffset !== undefined ? { estimateOffset: this.estimateOffset } : {}),
+            ...(this.measuredPrefixLength !== undefined
+                ? { measuredPrefixLength: this.measuredPrefixLength }
+                : {}),
+        });
+        this.lastPressure = pressure;
+        return pressure;
+    }
+    lastMeasured() {
+        return this.lastPressure;
+    }
+    /** Written by the wiring once the engine has reported its verdict. */
+    recordApplied(tokensBefore, tokensAfter) {
+        const reclaimed = Math.max(0, tokensBefore - tokensAfter);
+        this.tokensReclaimed += reclaimed;
+    }
+    sessionTotals() {
+        return {
+            compactions: this.compactions,
+            tokensReclaimed: this.tokensReclaimed,
+            usage: this.sessionUsage,
+            pricingUnknown: this.pricingUnknown,
+            inFlight: this.inFlight,
+            selfDisabled: this.selfDisabled,
+            ...(this.selfDisabledReason ? { selfDisabledReason: this.selfDisabledReason } : {}),
+            generation: this.generation,
+        };
+    }
+    /** `/compact on` after a self-disable — the user overriding the guard. */
+    clearSelfDisable() {
+        this.selfDisabled = false;
+        this.selfDisabledReason = undefined;
+        this.consecutiveNoProgress = 0;
+    }
+    isSelfDisabled() {
+        return this.selfDisabled;
+    }
+    /**
+     * Which model summarizes (§3.6.4).
+     *
+     * SAY THE DEFAULT OUT LOUD (P2-8): FOR ALMOST EVERYONE THIS IS THE MAIN MODEL.
+     * `resolveFastTier` returns `{ok: false, reason: 'disabled'}` on
+     * `fast.enabled !== true` alone, and `fast.enabled` defaults to `false` — so
+     * `useFastTier: true` describes a PREFERENCE, not the common path, and the
+     * out-of-the-box first compaction is a ~30 k-token call on the session's own
+     * model. Three things make that the right trade rather than a hidden cost: it
+     * happens at most `maxPerRun` times per run, the alternative is losing the
+     * entire run, and the spend is on the card with the summarizer's own price.
+     */
+    resolveSummarizer() {
+        const config = this.deps.getConfig();
+        if (config.compaction.useFastTier) {
+            const tier = resolveFastTier(config, this.deps.hasKey);
+            if (tier.ok)
+                return { ref: tier.ref, fromFastTier: true };
+        }
+        const ref = this.mainRef();
+        if (!this.deps.hasKey(ref.providerId))
+            return null;
+        return { ref, fromFastTier: false };
+    }
+    mainRef() {
+        const config = this.deps.getConfig();
+        return {
+            providerId: config.provider,
+            modelId: config.model,
+            ...(config.baseUrl ? { baseUrl: config.baseUrl } : {}),
+        };
+    }
+    modelInfo() {
+        return this.deps.getModelInfoFor(this.mainRef());
+    }
+    effectiveMaxOutputTokens() {
+        const config = this.deps.getConfig();
+        // `undefined` means AUTO; the model's own ceiling is the honest bound then.
+        return config.maxTokens ?? this.modelInfo().maxOutputTokens;
+    }
+    noteUsage(ref, usage) {
+        this.sessionUsage = accumulateUsage(this.sessionUsage, usage);
+        // UNKNOWN PRICING IS NOT ZERO PRICING. It is sticky: once a summarization has
+        // run on an unpriced model the session total is no longer a complete figure,
+        // and a later priced call does not make it one again.
+        if (!this.deps.isPricedModel(ref))
+            this.pricingUnknown = true;
+        this.deps.emit({ type: 'usage', usage });
+    }
+}
+/** A "turn" is a `user`-role message; `tool_result` is a separate role. */
+function countUserTurns(messages) {
+    return messages.reduce((acc, m) => (m.role === 'user' ? acc + 1 : acc), 0);
+}
+/**
+ * The relief-only outcome, recognized by the one reason string that produces it.
+ *
+ * A REASON RATHER THAN A NEW `mode` ON THE PORT (DH-5): `CompactionOutcome.mode`
+ * is a public contract with two members, and widening it to carry a distinction
+ * only the CLI's card acts on would be a breaking change for a reporting nicety.
+ */
+function reliefOnly(outcome) {
+    return outcome.action === 'replace' && outcome.reason === 'tail_relief_only';
+}