@aragon-agent/cli 0.6.47 → 0.6.48

package/dist/team/activity.d.ts

@@ -1,67 +0,0 @@
-/**
- * Activity strings for the live roster (team-live-activity F-1).
- *
- * ASCII ONLY: `src/team/**` is inside the glyph scanner's scope. That rule binds
- * the LITERALS IN THIS FILE, not the runtime data flowing through it - a child
- * may well be reading a CJK path, and `run.description` already carries model
- * text into the same row.
- */
-/**
- * The only part of a tool call this module will ever render (P1-4).
- *
- * A PICKED SUBSET, NOT THE ARGS OBJECT. Storing `event.args` whole would put
- * `write_file`'s `content` - a whole file - on the run, and `TeamRuntime.publish`
- * shallow-copies the run into the live snapshot AND into every `agent_update`
- * event, and `dispatch()` copies it again into `DispatchOutcome.runs`, which
- * `TeamCard` holds in the transcript for the rest of the session. Two facts make
- * that unacceptable rather than merely wasteful: the retention is unbounded in
- * the file's size, and one `JSON.stringify` anywhere downstream would put a file
- * body into a log - the exact liability the logging section spends a paragraph
- * avoiding.
- *
- * This costs nothing in maintenance: the switch below is ALREADY a per-tool
- * table, so a new tool is one entry either way.
- */
-export interface ActivityArgs {
-    path?: string;
-    pattern?: string;
-    command?: string;
-    to?: string;
-    from?: string;
-}
-/** Pick at capture time. Called once per `tool_execution_start`. */
-export declare function pickActivityArgs(args: unknown): ActivityArgs;
-/**
- * Strip anything that could corrupt the frame, collapse whitespace, and clamp.
- *
- * THE CONTROL-CHARACTER STRIP IS A CORRECTNESS BOUNDARY, NOT TIDINESS. This
- * string is model-chosen and reaches an Ink <Text> unescaped. A child running
- * `bash` with an ANSI escape in the command line would otherwise write that
- * escape into the roster: at best a mangled row, at worst a cleared screen or a
- * moved cursor in the middle of someone's transcript. `bash`'s own output never
- * reaches the panel; its ARGUMENTS now do.
- *
- * SURROGATES ARE GUARDED AT BOTH ENDS (P2-3). The trailing guard is the obvious
- * one. The LEADING one is needed because the caller feeds this a rolling tail
- * built with `String.slice(-n)`, which cuts on a UTF-16 code-unit boundary and
- * can therefore hand us a lone LOW surrogate at index 0. Either half of a split
- * pair renders as a replacement character - the same failure `truncateBytes`
- * guards one layer down (report.ts).
- */
-export declare function sanitizeActivity(raw: string, maxChars: number): string;
-/**
- * One short phrase describing a tool call.
- *
- * `wide` false is the narrow-terminal form (D-4). For a path-valued argument
- * that is the basename; for `bash` it is the PROGRAM NAME (P1-6). Running the
- * command line through `basename` is wrong twice over: `basename` splits on `/`,
- * so `npm test -w cli` comes back whole and `./scripts/build.sh --prod` comes
- * back as `build.sh --prod`. Neither is shorter and the second is a lie about
- * what ran.
- *
- * Anything not in the table renders as its own name. That is the correct
- * degradation for `skill_find` and for any tool a later round registers on a
- * child: a name is always true, and inventing an argument mapping for a tool
- * this module has not been taught is how a formatter starts lying.
- */
-export declare function describeToolActivity(toolName: string, args: ActivityArgs | undefined, wide: boolean): string;

package/dist/team/activity.js

@@ -1,130 +0,0 @@
-/**
- * Activity strings for the live roster (team-live-activity F-1).
- *
- * ASCII ONLY: `src/team/**` is inside the glyph scanner's scope. That rule binds
- * the LITERALS IN THIS FILE, not the runtime data flowing through it - a child
- * may well be reading a CJK path, and `run.description` already carries model
- * text into the same row.
- */
-import { TEAM_LIMITS } from './limits.js';
-/** Pick at capture time. Called once per `tool_execution_start`. */
-export function pickActivityArgs(args) {
-    const a = (args ?? {});
-    const str = (v) => typeof v === 'string' && v.length > 0 ? v.slice(0, TEAM_LIMITS.activityArgChars) : undefined;
-    const picked = {};
-    const path = str(a.path);
-    const pattern = str(a.pattern);
-    const command = str(a.command);
-    const to = str(a.to);
-    const from = str(a.from);
-    if (path !== undefined)
-        picked.path = path;
-    if (pattern !== undefined)
-        picked.pattern = pattern;
-    if (command !== undefined)
-        picked.command = command;
-    if (to !== undefined)
-        picked.to = to;
-    if (from !== undefined)
-        picked.from = from;
-    return picked;
-}
-/**
- * Strip anything that could corrupt the frame, collapse whitespace, and clamp.
- *
- * THE CONTROL-CHARACTER STRIP IS A CORRECTNESS BOUNDARY, NOT TIDINESS. This
- * string is model-chosen and reaches an Ink <Text> unescaped. A child running
- * `bash` with an ANSI escape in the command line would otherwise write that
- * escape into the roster: at best a mangled row, at worst a cleared screen or a
- * moved cursor in the middle of someone's transcript. `bash`'s own output never
- * reaches the panel; its ARGUMENTS now do.
- *
- * SURROGATES ARE GUARDED AT BOTH ENDS (P2-3). The trailing guard is the obvious
- * one. The LEADING one is needed because the caller feeds this a rolling tail
- * built with `String.slice(-n)`, which cuts on a UTF-16 code-unit boundary and
- * can therefore hand us a lone LOW surrogate at index 0. Either half of a split
- * pair renders as a replacement character - the same failure `truncateBytes`
- * guards one layer down (report.ts).
- */
-export function sanitizeActivity(raw, maxChars) {
-    let src = raw;
-    const head = src.charCodeAt(0);
-    if (head >= 0xdc00 && head <= 0xdfff)
-        src = src.slice(1);
-    const flat = src
-        // eslint-disable-next-line no-control-regex
-        .replace(/[\x00-\x1f\x7f]/g, ' ')
-        .replace(/\s+/g, ' ')
-        .trim();
-    if (maxChars <= 0)
-        return '';
-    if (flat.length <= maxChars)
-        return flat;
-    if (maxChars <= 3)
-        return flat.slice(0, maxChars);
-    let cut = maxChars - 3;
-    const code = flat.charCodeAt(cut - 1);
-    if (code >= 0xd800 && code <= 0xdbff)
-        cut -= 1;
-    return `${flat.slice(0, cut)}...`;
-}
-/** The basename of a path, for the narrow form. Pure string work: no `path`. */
-function basename(p) {
-    const parts = p.split(/[\\/]/);
-    return parts[parts.length - 1] || p;
-}
-/** The program name of a command line, for the narrow form. */
-function program(commandLine) {
-    const first = commandLine.trim().split(/\s+/)[0] ?? '';
-    return basename(first) || commandLine;
-}
-/**
- * One short phrase describing a tool call.
- *
- * `wide` false is the narrow-terminal form (D-4). For a path-valued argument
- * that is the basename; for `bash` it is the PROGRAM NAME (P1-6). Running the
- * command line through `basename` is wrong twice over: `basename` splits on `/`,
- * so `npm test -w cli` comes back whole and `./scripts/build.sh --prod` comes
- * back as `build.sh --prod`. Neither is shorter and the second is a lie about
- * what ran.
- *
- * Anything not in the table renders as its own name. That is the correct
- * degradation for `skill_find` and for any tool a later round registers on a
- * child: a name is always true, and inventing an argument mapping for a tool
- * this module has not been taught is how a formatter starts lying.
- */
-export function describeToolActivity(toolName, args, wide) {
-    const a = args ?? {};
-    switch (toolName) {
-        case 'read_file':
-            return detail('read', a.path, wide, basename);
-        case 'write_file':
-            return detail('write', a.path, wide, basename);
-        case 'edit_file':
-            return detail('edit', a.path, wide, basename);
-        case 'list_dir':
-            return detail('list', a.path ?? '.', wide, basename);
-        case 'glob':
-            return detail('glob', a.pattern, wide, basename);
-        case 'grep':
-            return detail('grep', a.pattern, wide, basename);
-        case 'bash':
-            return detail('bash', a.command, wide, program);
-        case 'team_send': {
-            const to = (a.to ?? '').trim();
-            return to ? `messaging ${to}` : 'messaging';
-        }
-        case 'team_wait': {
-            const from = (a.from ?? '').trim();
-            return from ? `waiting for ${from}` : 'waiting for mail';
-        }
-        default:
-            return toolName;
-    }
-}
-function detail(verb, value, wide, narrow) {
-    const v = (value ?? '').trim();
-    if (v.length === 0)
-        return verb;
-    return `${verb}: ${wide ? v : narrow(v)}`;
-}

package/dist/team/limits.d.ts

@@ -1,308 +0,0 @@
-/**
- * TEAM_LIMITS — the single authority on every bound this subsystem enforces
- * (team-subagents §3.5 / §3.6).
- *
- * ASCII ONLY: `src/team/**` is inside the glyph scanner's scope
- * (`glyphs.test.ts`), so no literal in this tree may hold a non-ASCII byte.
- *
- * TWO KINDS OF NUMBER LIVE HERE AND THEY ARE NOT INTERCHANGEABLE.
- *
- *  - The `TEAM_LIMITS` entries below are STRUCTURAL: they describe what the
- *    report format, the message channel and the panel can physically carry, and
- *    a user has no business tuning them. They are constants.
- *  - The six `team.*` config keys (`config/schema.ts`) are POLICY: how wide to
- *    fan out, how long to wait. They are clamped, persisted and user-facing.
- *
- * `hardMaxSubagents` sits on the structural side ON PURPOSE even though it
- * bounds a config key: the requirement caps the fan-out at 10 (R-g), so a
- * hand-edited `config.json` must not be able to raise it. It is applied twice —
- * once in `clampTeamConfig` and once in `normalizeSubagentSpecs` — because a
- * clamp that only runs on the read path leaves a bad value on disk.
- *
- * BYTES, NOT CHARACTERS, for every report budget (D-8 / I-9). `ToolExecutor`
- * truncates combined tool text at 100 000 BYTES; a 24 000-character CJK report
- * is 72 000 bytes, so a character-denominated cap would let the executor chop
- * the report mid-section and take the failure count with it.
- */
-export declare const TEAM_LIMITS: {
-    /**
-     * The requirement's ceiling (R-g). Enforced in `clampTeamConfig` AND in
-     * `normalizeSubagentSpecs`, so neither a hand-edited file nor a model asking
-     * for thirty can exceed it.
-     *
-     * 10 IS THE CEILING, 8 IS THE DEFAULT (`DEFAULT_TEAM_CONFIG`): different
-     * numbers on purpose (subagent-fanout D-1 / D-2). The default is a policy
-     * choice the user may move in either direction from the settings screen; this
-     * bound is what no user may move.
-     */
-    readonly hardMaxSubagents: 10;
-    /** `SubagentSpec.description`, clamped (never rejected). */
-    readonly descriptionChars: 60;
-    /** `SubagentSpec.prompt`, clamped. The child cannot see the conversation, so
-     *  this has to be generous enough to carry a whole brief. */
-    readonly promptChars: 8000;
-    /**
-     * The brief `/subagent <brief>` copies into its forced-directive turn
-     * (subagent-fanout 3.5.4 / R-10). CHARACTERS, not bytes, because this bound
-     * feeds a user-visible truncation marker rather than a wire budget: past it
-     * the directive says `[brief truncated]` so a clipped instruction is never
-     * mistaken for a complete one.
-     */
-    readonly forcedBriefChars: 4000;
-    /**
-     * Parts `/subagent --parts a,b,c` may name BEFORE `maxSubagents` narrows them
-     * (subagent-fanout 3.5.2 rule 3). The second, smaller ceiling is the policy
-     * number and belongs to the command body; this one only bounds what one
-     * argument can carry.
-     */
-    readonly forcedPartsMax: 8;
-    /** Slugified `SubagentSpec.label`. Long enough to be meaningful in the panel,
-     *  short enough that five of them fit one row of an 80-column terminal. */
-    readonly labelChars: 12;
-    readonly reportMaxBytes: 48000;
-    readonly summaryMaxBytes: 6000;
-    /**
-     * Floor a summary is trimmed TO, never below. A section trimmed to nothing is
-     * indistinguishable from a child that reported nothing, which is exactly the
-     * distinction the report exists to preserve.
-     */
-    readonly summaryFloorBytes: 400;
-    /** Deduped `write_file` / `edit_file` paths recorded per child (§3.4). */
-    readonly filesTouchedMax: 30;
-    /** `team_send` calls one child may spend on a whole dispatch. */
-    readonly messagesPerAgent: 6;
-    /** `team_send` calls the whole dispatch may spend. */
-    readonly messagesPerDispatch: 24;
-    /** Minimum gap between one child's sends. */
-    readonly minSendIntervalMs: 15000;
-    /** Quota units a `to: "all"` broadcast costs. */
-    readonly broadcastCost: 2;
-    readonly subjectChars: 80;
-    readonly bodyChars: 800;
-    /** `team_wait` timeout clamp, in seconds. */
-    readonly waitMinSeconds: 5;
-    readonly waitMaxSeconds: 120;
-    readonly waitDefaultSeconds: 30;
-    /**
-     * `agent_update` coalescing window (§5.2). Five children streaming tokens must
-     * not push five React renders per token; phase transitions always pass.
-     */
-    readonly agentUpdateThrottleMs: 120;
-    /** Agent rows the live panel renders before collapsing to `+N more` (§6.1). */
-    readonly panelMaxRows: 5;
-    /** Below this many terminal rows the panel collapses to its header line. */
-    readonly panelCollapseRows: 20;
-    /** Below this many columns the status-bar readout degrades to `[3]` (D-20). */
-    readonly statusCompactCols: 100;
-    /**
-     * STORAGE ceiling for a prose activity tail (P0-1). The most a `SubagentRun`
-     * may carry, NOT the width of the column: `TeamPanel.activityBudget(cols)`
-     * derives the display budget from the terminal and is always <= this.
-     */
-    readonly activityChars: 64;
-    /**
-     * Floor under `activityBudget`. Below roughly 46 columns the roster has bigger
-     * problems than an elided activity line, and a column narrower than this shows
-     * nothing but an ellipsis - at which point the phase word is more informative.
-     */
-    readonly activityMinChars: 12;
-    /** Rolling text buffer kept per child. The only per-token allocation. */
-    readonly activityTailChars: 160;
-    /** Per picked tool argument (P1-4). A path or a command longer than this is
-     *  elided on capture, so a run can never carry more than a few hundred bytes
-     *  of argument however large the real call was. */
-    readonly activityArgChars: 120;
-    /** At or above this many columns the activity line carries a full path or
-     *  command; below it, a basename or a program name. Mirrors `statusCompactCols`.
-     *  This chooses WHICH FACT to show; `activityBudget` chooses how much of it. */
-    readonly activityWideCols: 100;
-    /**
-     * Below this many columns the roster's activity cell gives way ENTIRELY
-     * (tui-subagent-panel §4.2.2).
-     *
-     * STRUCTURAL, like everything else in this block: twelve characters of prose in
-     * a 40-column terminal push the description column below its own twelve-column
-     * floor, and "who is doing what" outranks "what they are writing right now" when
-     * there is room for only one. That is why the cell returns 0 rather than
-     * `activityMinChars` below this width.
-     */
-    readonly activityDropCols: 56;
-    /**
-     * Events kept per child in `SubagentRun.timeline` (§4.5.3). Twenty-four covers
-     * a typical dispatch's complete tool sequence; ten children x 24 x 120 chars is
-     * about 29 KB at peak, and the ring is EPHEMERAL - it never reaches a file.
-     */
-    readonly timelineEntries: 24;
-    /** Display clamp on one timeline line; the same order as `activityArgChars`. */
-    readonly timelineTextChars: 120;
-    /**
-     * Row ceiling of the Agents board's projection. The overlay scrolls, so this is
-     * a memory bound rather than a fit: one row per fact and at most ten children
-     * plus one detail section, which no healthy dispatch can reach.
-     */
-    readonly boardMaxRows: 240;
-    /** Retries of a child that produced NOTHING. One, and not configurable: a
-     *  second retry doubles the worst-case latency of a failing dispatch and buys
-     *  nothing a re-dispatch by the lead would not buy better. */
-    readonly maxColdStartRetries: 1;
-    readonly retryBackoffMs: 2000;
-    /**
-     * Wall-clock budget of everything one SINGLE inspection may carry (the
-     * silence-triggered path). The digest holds run facts, the supervisor's own
-     * memory and the tail of the child's live history; unbounded, one look at a
-     * child that read five big files would make the "cheap" supervision call
-     * cost more than the child (the argument FAST_LIMITS.digestMaxChars makes,
-     * one level up).
-     */
-    readonly overseerDigestBytes: 8000;
-    /** Clamp on a `nudge` guidance text. Advice, not a second brief. */
-    readonly overseerGuidanceChars: 1200;
-    /** Clamp on the decision `reason` recorded in the report. */
-    readonly overseerReasonChars: 200;
-    /**
-     * Per-child slice of a BATCH digest (D-3 / R-P2-5). 6.7x tighter than the
-     * single-inspection budget above: a batch carries up to ten children in one
-     * call, and memory plus the history tail are the first to give way (the
-     * facts-first assembly order already puts them last).
-     */
-    readonly overseerPerChildDigestBytes: 1200;
-    /** Whole-batch digest ceiling (D-3): 10 children x 1 200 B fits exactly. */
-    readonly overseerBatchDigestBytes: 12000;
-    /**
-     * Output budget of ONE batch inspection (R-P0-1). The single-inspection path
-     * keeps `FAST_LIMITS.reviewOutputTokens` (512); a batch must return up to
-     * ten decision objects, and 512 tokens cannot physically hold ten x
-     * (reason <= 200 + guidance <= 1200). A truncated batch reply fails
-     * "first `[` to last `]`" parsing and degrades to all-wait - exactly the
-     * failure this separate, larger budget exists to prevent.
-     */
-    readonly overseerBatchOutputTokens: 2048;
-    /** Tighter `reason` clamp inside the batch protocol (R-P0-1). */
-    readonly overseerBatchReasonChars: 80;
-    /** Tighter `guidance` clamp inside the batch protocol (R-P0-1). */
-    readonly overseerBatchGuidanceChars: 400;
-    /**
-     * LOOKS per child per dispatch (D-5). A look is cost, not mutation: every
-     * inspection of a child - silence-triggered, cadence tick or post-mortem -
-     * spends one, including a call that failed softly. Twelve with the cadence
-     * ladder covers hours of healthy slow work; past the budget the supervisor
-     * says so once (`trigger: 'quiet'`) and stops waking that child up.
-     */
-    readonly overseerMaxLooksPerChild: 12;
-    /**
-     * LOOKS per dispatch, the fool-proof total (R-P1-6):
-     * hardMaxSubagents x overseerMaxLooksPerChild = 120. A mathematical upper
-     * bound, not a tuning knob - the per-child budget always binds first; this
-     * one exists so a bug cannot loop calls forever. When it trips, the whole
-     * MonitorLoop stops with one aggregate notice.
-     */
-    readonly overseerMaxLooksPerDispatch: number;
-    /**
-     * NUDGES per child per dispatch (D-5). Advice is cheap but not free: a
-     * supervisor that steers the same child forever is one that never decided.
-     */
-    readonly overseerMaxNudgesPerChild: 3;
-    /**
-     * Rebuilds PER CHILD per dispatch. One: a replacement that also stalls dies
-     * on its own watchdog rather than ping-ponging fresh children at a task the
-     * supervisor already failed to rescue once.
-     */
-    readonly overseerMaxReplacementsPerChild: 1;
-    /**
-     * Abandons PER CHILD per dispatch (D-5). One, so a supervisor that has
-     * already given up on a brief cannot keep re-killing bodies of it.
-     */
-    readonly overseerMaxAbandonsPerChild: 1;
-    /**
-     * Cadence ladder growth factor (D-2). Each no-action inspection schedules
-     * the next `previous * ratio` out, so a healthy slow child is looked at
-     * logarithmically in its run time rather than every five minutes.
-     */
-    readonly overseerCadenceRatio: 1.6;
-    /**
-     * First cadence check when neither `team.overseerIntervalMs` nor
-     * `team.subagentTimeoutMs` names one (the shipped defaults): the supervisor
-     * still looks even when no ceiling was configured.
-     */
-    readonly overseerDefaultCheckMs: 300000;
-    /** Clamp on a ladder step / model-chosen `nextCheckMs`, lower bound. */
-    readonly overseerNextCheckMinMs: 60000;
-    /** Clamp on a ladder step / model-chosen `nextCheckMs`, upper bound. */
-    readonly overseerNextCheckMaxMs: 900000;
-    /**
-     * A child's OWN idle watchdog, multiplied by this while the supervisor is
-     * active, so the soft silence inspection at `idleTimeoutMs` always precedes
-     * the hard abort. The watchdog stays as the backstop for a child that
-     * wedges past every inspection.
-     */
-    readonly overseerWatchdogFactor: 4;
-};
-/**
- * The pseudo-label the runtime's AGGREGATE supervisor notices travel under
- * (subagent-overseer-v2 D-5 / R-P1-6): the dispatch-total look-budget notice
- * is one fact about the whole roster, not a row any child owns. RESERVED at
- * normalization (`normalizeSubagentSpecs` seeds `dedupeLabel` with it), so a
- * spec asking for the bare label `team` is handed `team-2` instead - the
- * report ledger and the exec stream can then never attribute an aggregate
- * notice to a real child, or swallow that child's own quiet note.
- */
-export declare const TEAM_AGGREGATE_LABEL = "team";
-/**
- * The two comm tools, which are POLICY-EXEMPT inside a child (§3.11).
- *
- * They are not in the lead's registered tool list, so evaluating them against
- * the lead's `allowed-tools` ceiling would refuse them and dead-end the channel
- * — the same dead end `SKILL_TOOL_FLOOR` exists to prevent one level up.
- *
- * Deliberately NOT added to `HOST_TOOL_NAMES` (P2-5). That constant answers
- * "can a skill's `allowed-tools` declaration ever take effect?", and these two
- * are never registered on a lead — so a skill naming them genuinely IS
- * unenforceable, and `aragon skills doctor` reporting it as such is correct
- * rather than a gap.
- */
-export declare const TEAM_SUBAGENT_TOOL_NAMES: ReadonlySet<string>;
-/**
- * Bumped whenever the wording of either prompt block changes, so a behaviour
- * report can be tied to a block revision with one grep.
- */
-export declare const TEAM_BLOCK_VERSION = "v3-2026-10";
-/**
- * The version string on `<forced_team_mode>`'s first line.
- *
- * The block's words are a CONTRACT with the model, so a behaviour report ("it
- * still did the work itself") has to be tieable to one revision. Kept beside
- * `TEAM_BLOCK_VERSION` rather than reusing it: the two blocks answer different
- * questions ("can dispatch" versus "must dispatch") and a shared number would
- * name neither revision.
- */
-export declare const FORCED_TEAM_BLOCK_VERSION = "v1-2026-10";
-/**
- * Length ceiling of the `<forced_team_mode>` text.
- *
- * THE FIRST BLOCK LENGTH IN THIS SUBSYSTEM THAT A TEST ASSERTS: the `<team_mode>`
- * block's 1200 is a comment in `team/prompt.ts`, and unifying the two is a
- * separate small refactor (deliberately out of scope here). Looser than 1200
- * because this block is only ever paid for while the session is ARMED, and it
- * carries three layers of instruction rather than one.
- */
-export declare const FORCED_TEAM_BLOCK_MAX_CHARS = 1400;
-/**
- * Fallback nudges per user turn. ONE, and not configurable.
- *
- * The gate exists because the first two layers (the directive and the prompt
- * block) both depend on the model cooperating. A second nudge would double the
- * worst case of a model that is simply not going to dispatch - at N children per
- * attempt - and D-8's whole argument is that the mechanical layer must be
- * bounded by something other than the model's willingness.
- */
-export declare const FORCED_NUDGE_LIMIT = 1;
-/**
- * Grace window before the fallback nudge fires, in milliseconds.
- *
- * SAME ORDER AS THE TODO FOLLOW-THROUGH GRACE, for a structural reason rather
- * than a stylistic one: both continuations share ONE `followTimer` in `App`
- * (I-9), so "at most one armed continuation" is only explicable while the two
- * numbers stay in the same order of magnitude. **Changing todo's grace means
- * re-reading this constant.**
- */
-export declare const FORCED_NUDGE_GRACE_MS = 3000;