@aragon-agent/cli 0.6.1 → 0.6.3

package/dist/compaction/digest.d.ts

@@ -0,0 +1,62 @@
+/**
+ * The head of the history, rendered as a bounded plain-text transcript
+ * (context-auto-compaction §3.6.3).
+ *
+ * ASCII ONLY: `src/compaction/**` is inside the glyph scanner's scope.
+ *
+ * PURE. Two reasons this is a digest rather than raw messages, and the second
+ * one is the one that bites:
+ *
+ *  - Tool results are the bulk. Four 60 KB file reads are 240 KB of body that
+ *    says almost nothing about what was decided.
+ *  - THE SUMMARIZER MAY HAVE A SMALLER CONTEXT WINDOW THAN THE MODEL BEING
+ *    COMPACTED. The whole point of using a cheap model is that it is cheap, and
+ *    cheap models are exactly the ones with 32 k windows. Sending 180 k of
+ *    history to summarize would fail with the same error we are recovering from
+ *    (R-4).
+ */
+import { type Message } from '@aragon-agent/core';
+/**
+ * Is this `user` message a previous `<compacted_context>` block?
+ *
+ * Recognized on the OPENING TAG rather than a full parse: the block carries
+ * attributes (`version`, `replaced`, `turns`, `anchor`, `generation`) that this
+ * module has no business knowing about, and a stricter match would fail on the
+ * first attribute anybody adds.
+ */
+export declare function isPriorSummaryBlock(message: Message): boolean;
+export interface DigestInput {
+    /** The head — `messages[protectedPrefix..cutIndex)`. */
+    head: readonly Message[];
+    /**
+     * The summarizer's own context window, or `undefined` when it is unknown.
+     *
+     * NOT the lead model's. This is the whole reason a digest exists rather than
+     * raw messages (R-4).
+     */
+    summarizerWindow?: number;
+}
+export interface DigestResult {
+    text: string;
+    /** Entries the budget forced out. `> 0` means the marker is present. */
+    omitted: number;
+    /** The budget actually used, after the window derivation. */
+    budgetChars: number;
+}
+/**
+ * The character budget for one digest.
+ *
+ * `min(digestMaxChars, (summarizerWindow - summarizerReserveTokens) * CHARS_PER_TOKEN)`.
+ * `CHARS_PER_TOKEN` comes from core (P2-3) rather than being re-spelled here: two
+ * chars-per-token constants that must agree and nothing checking that they do is
+ * how a digest budget and an occupancy estimate drift apart.
+ */
+export declare function digestBudgetChars(summarizerWindow?: number): number;
+/**
+ * Render the head into a bounded transcript.
+ *
+ * OLDEST ENTRIES ARE DROPPED FIRST, except the pinned `PRIOR SUMMARY`, which is
+ * evicted last of all. A single omission marker is prepended so the summary can
+ * say that it is incomplete instead of pretending.
+ */
+export declare function buildDigest(input: DigestInput): DigestResult;

package/dist/compaction/digest.js

@@ -0,0 +1,164 @@
+/**
+ * The head of the history, rendered as a bounded plain-text transcript
+ * (context-auto-compaction §3.6.3).
+ *
+ * ASCII ONLY: `src/compaction/**` is inside the glyph scanner's scope.
+ *
+ * PURE. Two reasons this is a digest rather than raw messages, and the second
+ * one is the one that bites:
+ *
+ *  - Tool results are the bulk. Four 60 KB file reads are 240 KB of body that
+ *    says almost nothing about what was decided.
+ *  - THE SUMMARIZER MAY HAVE A SMALLER CONTEXT WINDOW THAN THE MODEL BEING
+ *    COMPACTED. The whole point of using a cheap model is that it is cheap, and
+ *    cheap models are exactly the ones with 32 k windows. Sending 180 k of
+ *    history to summarize would fail with the same error we are recovering from
+ *    (R-4).
+ */
+import { CHARS_PER_TOKEN } from '@aragon-agent/core';
+import { BLOCK_TAG_CLOSE, BLOCK_TAG_OPEN, COMPACTION_LIMITS } from './limits.js';
+/** Prepended when the budget forced entries out, so the summary can say so. */
+const OMISSION_MARKER = (n) => `[... ${n} earlier ${n === 1 ? 'message' : 'messages'} omitted from this digest ...]`;
+function clip(text, max) {
+    if (text.length <= max)
+        return text;
+    return `${text.slice(0, max)}...`;
+}
+/** Flatten `string | ContentPart[]` into text, naming images rather than dropping them. */
+function partsToText(content) {
+    if (typeof content === 'string')
+        return content;
+    if (!Array.isArray(content))
+        return '';
+    return content
+        .map((part) => (typeof part?.text === 'string' ? part.text : '[image]'))
+        .join('\n');
+}
+/**
+ * Is this `user` message a previous `<compacted_context>` block?
+ *
+ * Recognized on the OPENING TAG rather than a full parse: the block carries
+ * attributes (`version`, `replaced`, `turns`, `anchor`, `generation`) that this
+ * module has no business knowing about, and a stricter match would fail on the
+ * first attribute anybody adds.
+ */
+export function isPriorSummaryBlock(message) {
+    if (message.role !== 'user')
+        return false;
+    const text = partsToText(message.content).trimStart();
+    return text.startsWith(BLOCK_TAG_OPEN);
+}
+/** The block's body with its framing removed, for re-rendering as PRIOR SUMMARY. */
+function stripBlockTags(text) {
+    const trimmed = text.trim();
+    const open = trimmed.indexOf('>');
+    const close = trimmed.lastIndexOf(BLOCK_TAG_CLOSE);
+    if (open < 0)
+        return trimmed;
+    const body = close > open ? trimmed.slice(open + 1, close) : trimmed.slice(open + 1);
+    return body.trim();
+}
+function renderMessage(message) {
+    if (message.role === 'user') {
+        if (isPriorSummaryBlock(message)) {
+            // CARRIED WHOLE, up to `summaryMaxChars` (D-22 / P1-9).
+            //
+            // A `<compacted_context>` block IS a `user` message (D-5, deliberately),
+            // so a per-kind table that clipped it to `userChars` (2 000) would discard
+            // up to two thirds of the previous summary - which §5.4 permits to be
+            // 6 000 long - BEFORE summarizing it, and then report `summarized`. Every
+            // compaction after the first would decay the record silently.
+            const body = stripBlockTags(partsToText(message.content));
+            return [{ text: `PRIOR SUMMARY: ${clip(body, COMPACTION_LIMITS.summaryMaxChars)}`, pinned: true }];
+        }
+        return [
+            { text: `USER: ${clip(partsToText(message.content), COMPACTION_LIMITS.userChars)}`, pinned: false },
+        ];
+    }
+    if (message.role === 'tool_result') {
+        const status = message.isError ? 'error' : 'ok';
+        const body = clip(partsToText(message.content), COMPACTION_LIMITS.toolResultChars);
+        return [{ text: `RESULT[${status}] ${body}`, pinned: false }];
+    }
+    // Assistant: one entry per block, so a turn's text and its calls can be
+    // evicted independently.
+    const out = [];
+    for (const block of message.content) {
+        if (block.type === 'text') {
+            if (block.text.trim().length === 0)
+                continue;
+            out.push({ text: `ASSISTANT: ${clip(block.text, COMPACTION_LIMITS.assistantChars)}`, pinned: false });
+        }
+        else if (block.type === 'tool_call') {
+            const args = clip(JSON.stringify(block.args ?? {}), COMPACTION_LIMITS.toolArgChars);
+            out.push({ text: `CALL ${block.toolName}(${args})`, pinned: false });
+        }
+        // `thinking` blocks are DROPPED ENTIRELY (D-8): they are the model's private
+        // scratch, they are the single largest contributor at `thinkingLevel: high`,
+        // and nothing after the turn depends on them.
+    }
+    return out;
+}
+/**
+ * The character budget for one digest.
+ *
+ * `min(digestMaxChars, (summarizerWindow - summarizerReserveTokens) * CHARS_PER_TOKEN)`.
+ * `CHARS_PER_TOKEN` comes from core (P2-3) rather than being re-spelled here: two
+ * chars-per-token constants that must agree and nothing checking that they do is
+ * how a digest budget and an occupancy estimate drift apart.
+ */
+export function digestBudgetChars(summarizerWindow) {
+    const ceiling = COMPACTION_LIMITS.digestMaxChars;
+    if (!summarizerWindow || summarizerWindow <= 0)
+        return ceiling;
+    const usableTokens = summarizerWindow - COMPACTION_LIMITS.summarizerReserveTokens;
+    if (usableTokens <= 0)
+        return ceiling;
+    return Math.min(ceiling, usableTokens * CHARS_PER_TOKEN);
+}
+/**
+ * Render the head into a bounded transcript.
+ *
+ * OLDEST ENTRIES ARE DROPPED FIRST, except the pinned `PRIOR SUMMARY`, which is
+ * evicted last of all. A single omission marker is prepended so the summary can
+ * say that it is incomplete instead of pretending.
+ */
+export function buildDigest(input) {
+    const budgetChars = digestBudgetChars(input.summarizerWindow);
+    const entries = input.head.flatMap(renderMessage);
+    const sep = '\n';
+    let total = entries.reduce((acc, e) => acc + e.text.length + sep.length, 0);
+    if (total <= budgetChars) {
+        return { text: entries.map((e) => e.text).join(sep), omitted: 0, budgetChars };
+    }
+    // Evict oldest-first among the unpinned. The marker itself costs characters,
+    // so it is charged up front against a worst-case width.
+    const kept = [...entries];
+    let omitted = 0;
+    let reserve = OMISSION_MARKER(entries.length).length + sep.length;
+    for (let i = 0; i < kept.length && total + reserve > budgetChars; i += 1) {
+        const entry = kept[i];
+        if (entry.pinned)
+            continue;
+        total -= entry.text.length + sep.length;
+        kept[i] = { text: '', pinned: false };
+        omitted += 1;
+    }
+    // Still over budget with only pinned entries left: clip the pinned text rather
+    // than dropping it, so the densest record survives in some form.
+    if (total + reserve > budgetChars) {
+        const room = Math.max(0, budgetChars - reserve);
+        const joined = kept.filter((e) => e.text.length > 0).map((e) => e.text).join(sep);
+        return {
+            text: `${OMISSION_MARKER(omitted)}${sep}${joined.slice(0, room)}`,
+            omitted,
+            budgetChars,
+        };
+    }
+    const body = kept.filter((e) => e.text.length > 0).map((e) => e.text).join(sep);
+    return {
+        text: omitted > 0 ? `${OMISSION_MARKER(omitted)}${sep}${body}` : body,
+        omitted,
+        budgetChars,
+    };
+}

package/dist/compaction/limits.d.ts

@@ -0,0 +1,165 @@
+/**
+ * COMPACTION_LIMITS — the single authority on every bound context compaction
+ * enforces (context-auto-compaction §5.4).
+ *
+ * ASCII ONLY: `src/compaction/**` is inside the glyph scanner's scope
+ * (`glyphs.test.ts::inScope`), so no literal in this tree may hold a non-ASCII
+ * byte. That regex is a HARDCODED DIRECTORY LIST, which is why adding this tree
+ * and adding `compaction` to that list are the SAME change (C-7) - a scanner
+ * that silently stops scanning is worse than no scanner. `fast/limits.ts`
+ * records the identical trap one feature earlier.
+ *
+ * TWO KINDS OF NUMBER, AND THEY ARE NOT INTERCHANGEABLE - the distinction
+ * `TEAM_LIMITS`, `TODO_LIMITS` and `FAST_LIMITS` all state:
+ *
+ *  - The entries below are STRUCTURAL. They describe what a digest, a
+ *    summarization call and a card can physically carry. A user has no business
+ *    tuning them.
+ *  - The six `compaction.*` keys in `config/schema.ts` are POLICY: when to fire,
+ *    how much to keep, which model, what to do on failure. Those are clamped,
+ *    persisted and user-facing.
+ */
+export declare const COMPACTION_LIMITS: {
+    /** Ceiling on the rendered digest, before the summarizer-window check. ~30k tokens. */
+    readonly digestMaxChars: 120000;
+    /** Output + system prompt + framing, kept out of the digest's share of the window. */
+    readonly summarizerReserveTokens: 8000;
+    readonly userChars: 2000;
+    readonly assistantChars: 2000;
+    readonly toolArgChars: 200;
+    readonly toolResultChars: 600;
+    /**
+     * Clip bound on the TAGGED text anchor.
+     *
+     * A CLIP, NOT A SHAPE SWITCH (D-21 / P1-9). v1 of the design kept the anchor
+     * verbatim-standalone past this length and merged it into the block below it,
+     * which made the anchor rule LENGTH-DEPENDENT and therefore non-idempotent: on
+     * the second compaction the "first user message" is a summary. A 4 000-character
+     * first message clipped with a visible marker keeps the rule "one leading
+     * anchor, always recognizable"; a 40 000-character first message is not a goal
+     * statement, and truncating it with a marker is the honest outcome.
+     */
+    readonly anchorChars: 4000;
+    /** What the summarizer prompt asks for, and what the splice clips to. */
+    readonly summaryMaxChars: 6000;
+    readonly summaryOutputTokens: 2048;
+    /**
+     * Wall-clock budget for ONE summarization call.
+     *
+     * Longer than the fast reviewer's 20 s because the input is ~30x larger and the
+     * output ~4x larger, and because the cost of a timeout here is a DEGRADED RUN
+     * rather than a missing critique.
+     *
+     * COUPLED TO THE ENGINE'S `COMPACTION_HARD_TIMEOUT_MS` (120 s). The worst legal
+     * host budget is two calls at this value back to back plus the digest render,
+     * so `2 * callTimeoutMs` must stay strictly under the engine's ceiling. Raising
+     * this without raising that makes the §3.7 ladder unreachable and turns every
+     * slow-but-correct summarization into `manager_timeout`.
+     */
+    readonly callTimeoutMs: 45000;
+    /** Turns between compactions. Never two in a row: 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`. */
+    readonly minTurnsBetween: 2;
+    /** Compactions STARTED per run, both triggers, hard. */
+    readonly maxPerRun: 5;
+    /** A splice must project occupancy below `threshold - this`, or it bought two turns. */
+    readonly minReclaimRatio: 0.15;
+    /** Consecutive no-progress PRESSURE compactions before the proactive trigger self-disables. */
+    readonly stuckLimit: 2;
+    /**
+     * Clip bound on an oversized `tool_result` body inside the RETAINED tail.
+     *
+     * 2 000 rather than `toolResultChars`' 600, and the two are not the same
+     * question. That one bounds a message in the DIGEST, which the summarizer
+     * reads once and which exists to be lossy. This one bounds a message the model
+     * keeps WORKING FROM: it is the last two-to-four turns, and the run is about to
+     * continue on them. Relief only ever runs when the alternative is a request the
+     * provider cannot accept.
+     */
+    readonly tailToolResultChars: 2000;
+    /**
+     * Room reserved for the `<compacted_context>` block in the tail-budget
+     * projection.
+     *
+     * THE BLOCK IS WRITTEN AFTER THE PROJECTION, which is why this exists at all:
+     * relief measures a spliced history whose summary body is already in it, but
+     * the `plan === null` path measures one where nothing has been written yet, and
+     * a budget with no allowance would let the block push the result back over.
+     */
+    readonly blockAllowanceTokens: 2000;
+    /**
+     * Files kept for ONE `runId`, newest first.
+     *
+     * PER RUN, NOT PER DIRECTORY (RV-6). `<home>/compaction/` is shared by every
+     * `aragon` on the machine — `file-sink.ts:94-95` already treats concurrent
+     * processes as normal — so a global count-based prune would delete a live
+     * session's archives out from under it.
+     */
+    readonly archiveMaxFiles: 20;
+    /**
+     * The cross-run sweep: 7 days.
+     *
+     * AGE RATHER THAN COUNT, for the reason above. Age can only ever delete files
+     * whose run finished a week ago; a count can delete the newest file another
+     * process wrote a second ago.
+     */
+    readonly archiveMaxAgeMs: 604800000;
+    /**
+     * Per-file ceiling before dropped message bodies are clipped from the OLDEST
+     * end. The metadata and the summary are never clipped: a truncated archive
+     * that still says what happened beats none.
+     */
+    readonly archiveMaxBytes: 8000000;
+    /** A child's `keepRecentTurns`, delivered through its config view. */
+    readonly childKeepRecentTurns: 2;
+    /**
+     * A child's guard-2 ceiling, delivered through `Compactor`'s instance bound.
+     *
+     * HERE RATHER THAN AS A `compaction.*` KEY, and the placement is the whole
+     * argument: `maxPerRun` is STRUCTURAL (see this file's header), there is no
+     * config path to it by design, and a child needs a tighter one. Both numbers
+     * therefore stay in the one file that owns structural bounds.
+     */
+    readonly childMaxPerRun: 2;
+    /** Rows the collapsed transcript card renders; Ctrl+O expands. */
+    readonly cardTextRows: 6;
+    /**
+     * Below this many columns the status chip drops.
+     *
+     * 100, matching all four existing peers - `retry-view.ts`, `fast/limits.ts`,
+     * `team/limits.ts`, `todo/limits.ts` (P2-2). A status-bar degradation ladder
+     * that is nearly-but-not-quite uniform is a reading hazard bought for nothing.
+     * The guaranteed reporting surface is `/compact status`, which is what makes
+     * dropping the chip acceptable at all.
+     */
+    readonly statusCompactCols: 100;
+};
+/**
+ * Bumped whenever the wording of the `<compacted_context>` block or the
+ * summarizer system prompt changes, so a behaviour report can be tied to a block
+ * revision with one grep - exactly what `FAST_BLOCK_VERSION` and
+ * `TEAM_BLOCK_VERSION` are for.
+ */
+export declare const COMPACTION_BLOCK_VERSION = "v1-2026-08";
+/** Opening tag of the verbatim goal anchor (§3.6.5). */
+export declare const ANCHOR_TAG_OPEN = "<original_task>";
+export declare const ANCHOR_TAG_CLOSE = "</original_task>";
+/** Opening tag of the spliced summary block. Recognized on the NEXT compaction. */
+export declare const BLOCK_TAG_OPEN = "<compacted_context";
+export declare const BLOCK_TAG_CLOSE = "</compacted_context>";
+/** Marker appended to an anchor clipped at `anchorChars`. ASCII (see the header). */
+export declare const ANCHOR_TRUNCATION_MARK = " [truncated]";
+/**
+ * What tail relief writes where it removed text (hardening §3.3.2 rule 4).
+ *
+ * IN THE HISTORY THE MODEL READS, at the exact place the data went missing. A
+ * model that sees this knows the output is partial and can re-run the tool; a
+ * model handed a silently truncated file cannot tell the difference between "the
+ * file ends here" and "the file was cut here".
+ *
+ * THE HOST OWNS IT, not core: `engine/compaction.ts` has no glyph vocabulary and
+ * must not learn one, so `relieveTail` takes this as a callback. ASCII, per this
+ * file's header.
+ */
+export declare function tailClipMarker(removed: number): string;

package/dist/compaction/limits.js

@@ -0,0 +1,172 @@
+/**
+ * COMPACTION_LIMITS — the single authority on every bound context compaction
+ * enforces (context-auto-compaction §5.4).
+ *
+ * ASCII ONLY: `src/compaction/**` is inside the glyph scanner's scope
+ * (`glyphs.test.ts::inScope`), so no literal in this tree may hold a non-ASCII
+ * byte. That regex is a HARDCODED DIRECTORY LIST, which is why adding this tree
+ * and adding `compaction` to that list are the SAME change (C-7) - a scanner
+ * that silently stops scanning is worse than no scanner. `fast/limits.ts`
+ * records the identical trap one feature earlier.
+ *
+ * TWO KINDS OF NUMBER, AND THEY ARE NOT INTERCHANGEABLE - the distinction
+ * `TEAM_LIMITS`, `TODO_LIMITS` and `FAST_LIMITS` all state:
+ *
+ *  - The entries below are STRUCTURAL. They describe what a digest, a
+ *    summarization call and a card can physically carry. A user has no business
+ *    tuning them.
+ *  - The six `compaction.*` keys in `config/schema.ts` are POLICY: when to fire,
+ *    how much to keep, which model, what to do on failure. Those are clamped,
+ *    persisted and user-facing.
+ */
+export const COMPACTION_LIMITS = {
+    /** Ceiling on the rendered digest, before the summarizer-window check. ~30k tokens. */
+    digestMaxChars: 120_000,
+    /** Output + system prompt + framing, kept out of the digest's share of the window. */
+    summarizerReserveTokens: 8_000,
+    // --- Per-message clips in the digest (§3.6.3). ---
+    userChars: 2_000,
+    assistantChars: 2_000,
+    toolArgChars: 200,
+    toolResultChars: 600,
+    /**
+     * Clip bound on the TAGGED text anchor.
+     *
+     * A CLIP, NOT A SHAPE SWITCH (D-21 / P1-9). v1 of the design kept the anchor
+     * verbatim-standalone past this length and merged it into the block below it,
+     * which made the anchor rule LENGTH-DEPENDENT and therefore non-idempotent: on
+     * the second compaction the "first user message" is a summary. A 4 000-character
+     * first message clipped with a visible marker keeps the rule "one leading
+     * anchor, always recognizable"; a 40 000-character first message is not a goal
+     * statement, and truncating it with a marker is the honest outcome.
+     */
+    anchorChars: 4_000,
+    /** What the summarizer prompt asks for, and what the splice clips to. */
+    summaryMaxChars: 6_000,
+    summaryOutputTokens: 2_048,
+    /**
+     * Wall-clock budget for ONE summarization call.
+     *
+     * Longer than the fast reviewer's 20 s because the input is ~30x larger and the
+     * output ~4x larger, and because the cost of a timeout here is a DEGRADED RUN
+     * rather than a missing critique.
+     *
+     * COUPLED TO THE ENGINE'S `COMPACTION_HARD_TIMEOUT_MS` (120 s). The worst legal
+     * host budget is two calls at this value back to back plus the digest render,
+     * so `2 * callTimeoutMs` must stay strictly under the engine's ceiling. Raising
+     * this without raising that makes the §3.7 ladder unreachable and turns every
+     * slow-but-correct summarization into `manager_timeout`.
+     */
+    callTimeoutMs: 45_000,
+    // --- Anti-loop guards (§3.8). ---
+    /** Turns between compactions. Never two in a row: 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`. */
+    minTurnsBetween: 2,
+    /** Compactions STARTED per run, both triggers, hard. */
+    maxPerRun: 5,
+    /** A splice must project occupancy below `threshold - this`, or it bought two turns. */
+    minReclaimRatio: 0.15,
+    /** Consecutive no-progress PRESSURE compactions before the proactive trigger self-disables. */
+    stuckLimit: 2,
+    // --- Tail relief (hardening §3.3 / W2). ---
+    /**
+     * Clip bound on an oversized `tool_result` body inside the RETAINED tail.
+     *
+     * 2 000 rather than `toolResultChars`' 600, and the two are not the same
+     * question. That one bounds a message in the DIGEST, which the summarizer
+     * reads once and which exists to be lossy. This one bounds a message the model
+     * keeps WORKING FROM: it is the last two-to-four turns, and the run is about to
+     * continue on them. Relief only ever runs when the alternative is a request the
+     * provider cannot accept.
+     */
+    tailToolResultChars: 2_000,
+    /**
+     * Room reserved for the `<compacted_context>` block in the tail-budget
+     * projection.
+     *
+     * THE BLOCK IS WRITTEN AFTER THE PROJECTION, which is why this exists at all:
+     * relief measures a spliced history whose summary body is already in it, but
+     * the `plan === null` path measures one where nothing has been written yet, and
+     * a budget with no allowance would let the block push the result back over.
+     */
+    blockAllowanceTokens: 2_000,
+    // --- The archive (hardening §3.5 / W4). ---
+    /**
+     * Files kept for ONE `runId`, newest first.
+     *
+     * PER RUN, NOT PER DIRECTORY (RV-6). `<home>/compaction/` is shared by every
+     * `aragon` on the machine — `file-sink.ts:94-95` already treats concurrent
+     * processes as normal — so a global count-based prune would delete a live
+     * session's archives out from under it.
+     */
+    archiveMaxFiles: 20,
+    /**
+     * The cross-run sweep: 7 days.
+     *
+     * AGE RATHER THAN COUNT, for the reason above. Age can only ever delete files
+     * whose run finished a week ago; a count can delete the newest file another
+     * process wrote a second ago.
+     */
+    archiveMaxAgeMs: 604_800_000,
+    /**
+     * Per-file ceiling before dropped message bodies are clipped from the OLDEST
+     * end. The metadata and the summary are never clipped: a truncated archive
+     * that still says what happened beats none.
+     */
+    archiveMaxBytes: 8_000_000,
+    // --- Sub-agent compaction (hardening §3.4 / W3). ---
+    /** A child's `keepRecentTurns`, delivered through its config view. */
+    childKeepRecentTurns: 2,
+    /**
+     * A child's guard-2 ceiling, delivered through `Compactor`'s instance bound.
+     *
+     * HERE RATHER THAN AS A `compaction.*` KEY, and the placement is the whole
+     * argument: `maxPerRun` is STRUCTURAL (see this file's header), there is no
+     * config path to it by design, and a child needs a tighter one. Both numbers
+     * therefore stay in the one file that owns structural bounds.
+     */
+    childMaxPerRun: 2,
+    /** Rows the collapsed transcript card renders; Ctrl+O expands. */
+    cardTextRows: 6,
+    /**
+     * Below this many columns the status chip drops.
+     *
+     * 100, matching all four existing peers - `retry-view.ts`, `fast/limits.ts`,
+     * `team/limits.ts`, `todo/limits.ts` (P2-2). A status-bar degradation ladder
+     * that is nearly-but-not-quite uniform is a reading hazard bought for nothing.
+     * The guaranteed reporting surface is `/compact status`, which is what makes
+     * dropping the chip acceptable at all.
+     */
+    statusCompactCols: 100,
+};
+/**
+ * Bumped whenever the wording of the `<compacted_context>` block or the
+ * summarizer system prompt changes, so a behaviour report can be tied to a block
+ * revision with one grep - exactly what `FAST_BLOCK_VERSION` and
+ * `TEAM_BLOCK_VERSION` are for.
+ */
+export const COMPACTION_BLOCK_VERSION = 'v1-2026-08';
+/** Opening tag of the verbatim goal anchor (§3.6.5). */
+export const ANCHOR_TAG_OPEN = '<original_task>';
+export const ANCHOR_TAG_CLOSE = '</original_task>';
+/** Opening tag of the spliced summary block. Recognized on the NEXT compaction. */
+export const BLOCK_TAG_OPEN = '<compacted_context';
+export const BLOCK_TAG_CLOSE = '</compacted_context>';
+/** Marker appended to an anchor clipped at `anchorChars`. ASCII (see the header). */
+export const ANCHOR_TRUNCATION_MARK = ' [truncated]';
+/**
+ * What tail relief writes where it removed text (hardening §3.3.2 rule 4).
+ *
+ * IN THE HISTORY THE MODEL READS, at the exact place the data went missing. A
+ * model that sees this knows the output is partial and can re-run the tool; a
+ * model handed a silently truncated file cannot tell the difference between "the
+ * file ends here" and "the file was cut here".
+ *
+ * THE HOST OWNS IT, not core: `engine/compaction.ts` has no glyph vocabulary and
+ * must not learn one, so `relieveTail` takes this as a callback. ASCII, per this
+ * file's header.
+ */
+export function tailClipMarker(removed) {
+    return `\n[... ${removed} characters removed by context compaction ...]`;
+}