crewly 1.20.358 → 1.20.360

package/dist/backend/backend/src/constants.js

@@ -1780,6 +1780,119 @@ export const AGENT_MEMORY_REFRESH_CONSTANTS = {
     /** Kill switch env var; the value `false` disables the refresh. */
     ENV_KILL_SWITCH: 'CREWLY_AGENT_MEMORY_REFRESH',
 };
+/**
+ * Automatic per-task work records (#1349): one `(record)` entry per finished
+ * WorkItem in the project vault log, compiled into the weekly digest.
+ */
+export const WORK_RECORD_CONSTANTS = {
+    /** Hourly tick: catch-up for finished items without a record, blocked items, deferred items. */
+    TICK_MS: 60 * 60 * 1000,
+    /** Delay before the boot-time pass, after the other wiki jobs settled. */
+    BOOT_DELAY_MS: 3 * 60 * 1000,
+    /** A terminal event waits this long before the record is assembled (ticket sync, verify item, evidence land first). */
+    EVENT_DELAY_MS: 20 * 1000,
+    /** Catch-up looks at items that ended within this many days. */
+    CATCHUP_WINDOW_DAYS: 7,
+    /** Cap on records written per pass (the rest wait for the next tick). */
+    CATCHUP_MAX_PER_PASS: 200,
+    /** A record waiting for the owner's OK on a ticket is written anyway after this long. */
+    DEFER_MAX_MS: 72 * 60 * 60 * 1000,
+    /** An item parked in `blocked` for longer than this gets a blocked record. */
+    BLOCKED_AFTER_MS: 24 * 60 * 60 * 1000,
+    /** Hard cap on one record's body (the `meta:` line always fits). */
+    MAX_CHARS: 2_500,
+    /** Characters of the ask quoted. */
+    QUOTE_MAX_CHARS: 160,
+    /** Characters of the result summary kept. */
+    DID_MAX_CHARS: 400,
+    /** Cap on output lines and characters per line. */
+    MAX_OUTPUTS: 8,
+    OUTPUT_MAX_CHARS: 160,
+    /** Cap on error lines and characters per line. */
+    MAX_ERRORS: 5,
+    ERROR_MAX_CHARS: 160,
+    /** Cap on helper lines and characters per line. */
+    MAX_HELPERS: 6,
+    HELPER_MAX_CHARS: 120,
+    /** Bytes read from the end of the agent's transcript for the task window. */
+    TRANSCRIPT_TAIL_BYTES: 2 * 1024 * 1024,
+    /** Transcripts are consulted only for items that ended within this long (older windows are not in the tail). */
+    TRANSCRIPT_MAX_AGE_MS: 48 * 60 * 60 * 1000,
+    /** Slack around the task window when slicing the transcript. */
+    TRANSCRIPT_WINDOW_SLACK_MS: 60 * 1000,
+    /** Cap on trace events read per record. */
+    TRACE_MAX_EVENTS: 2_000,
+    /** State file (recorded keys), under CREWLY_HOME. */
+    STATE_FILE: 'work-records-state.json',
+    /** Cap on remembered record keys. */
+    STATE_MAX_KEYS: 5_000,
+    /** Kill switch env var; the value `false` disables work records. */
+    ENV_KILL_SWITCH: 'CREWLY_WORK_RECORDS',
+    /** Source type of a record entry in the vault log. */
+    SOURCE_TYPE: 'work_record',
+    /** Prefix of a record's `ref:` (`work-record:<workItemId>#<phase>`). */
+    REF_PREFIX: 'work-record:',
+    /** Records the startup briefing lists. */
+    BRIEFING_MAX: 5,
+    /** Characters of one record line in the briefing. */
+    BRIEFING_LINE_MAX_CHARS: 240,
+    /** WorkItem types that are bookkeeping, never recorded. */
+    SKIPPED_TYPES: ['check', 'reconcile', 'confirm', 'ticket_triage', 'autopilot_retro', 'goal_replan', 'ticket_hygiene_review', 'autopilot_self_review'],
+};
+/**
+ * Memory-aware boot (crewly#1353). On 2026-10-17 a 16 GB Mac kernel-panicked
+ * a minute after a Crewly restart: 10 agents restored back-to-back, then the
+ * wiki, memory and work-record boot passes on fixed delays, while the machine
+ * was already swapping. Pressure mode only engages after two 30 s samples,
+ * so the boot restore reads the raw snapshot itself and the boot maintenance
+ * passes wait for a quiet machine. Env overrides exist for the two memory
+ * thresholds and the launch spacing.
+ */
+export const BOOT_PRESSURE_CONSTANTS = {
+    /** Swap in use at or above this share counts as tight for the boot (`CREWLY_BOOT_PRESSURE_SWAP_PERCENT`). */
+    SWAP_USED_PERCENT: 80,
+    /** Free memory below this counts as tight for the boot (`CREWLY_BOOT_PRESSURE_MIN_FREE_MB`). */
+    MIN_FREE_MB: 3 * 1024,
+    /** Minimum gap between two boot launches, pressure or not, so RSS growth is visible before the next (`CREWLY_RESTORE_LAUNCH_SPACING_MS`). */
+    MIN_LAUNCH_SPACING_MS: 15 * 1000,
+    /** Under pressure: how long after an agent registered its process counts as settled before the next launch. */
+    SETTLE_AFTER_READY_MS: 20 * 1000,
+    /** Under pressure: how often the snapshot is re-read while a launch waits for the machine to calm down. */
+    PRESSURE_POLL_MS: 10 * 1000,
+    /** A held launch is logged on the first poll and then every this many polls (once a minute at the default poll). */
+    PRESSURE_HOLD_LOG_EVERY_POLLS: 6,
+    /** Under pressure: an owner-waiting agent waits at most this long for the machine to calm, then launches anyway. */
+    PRESSURE_HOLD_MAX_MS: 10 * 60 * 1000,
+    /** Under pressure: at most this many boot launches in flight at once. */
+    PRESSURE_MAX_CONCURRENCY: 2,
+    /** Boot maintenance passes re-check the restore queue and the snapshot this often. */
+    MAINTENANCE_POLL_MS: 30 * 1000,
+    /** Boot maintenance passes give up waiting after this long and run reduced. */
+    MAINTENANCE_WAIT_MAX_MS: 30 * 60 * 1000,
+    /** Work records written by the boot pass when it runs reduced (instead of CATCHUP_MAX_PER_PASS). */
+    WORK_RECORD_REDUCED_BATCH: 50,
+    /** A vault log larger than this is rotated by streaming (tail read, header count) instead of being read whole. */
+    ROTATION_STREAM_BYTES: 10 * 1024 * 1024,
+    /** Bytes read from the end of a streamed log to pick the entries carried forward. */
+    ROTATION_TAIL_BYTES: 2 * 1024 * 1024,
+    /** Chunk size of the streaming entry count. */
+    ROTATION_CHUNK_BYTES: 1024 * 1024,
+    /** The owner gets one line when boot restores have been held back for longer than this. */
+    OWNER_NOTE_AFTER_MS: 10 * 60 * 1000,
+    /**
+     * Boot launches outside the restore queue (the deferred-start retry) are
+     * held for memory at most this long after boot; afterwards the pressure
+     * mode cap alone decides, because macOS swap stays full for hours after
+     * the pressure is gone and a held agent is still owed something.
+     */
+    RETRY_HOLD_MAX_MS: 60 * 60 * 1000,
+    /** Env override of SWAP_USED_PERCENT. */
+    ENV_SWAP_PERCENT: 'CREWLY_BOOT_PRESSURE_SWAP_PERCENT',
+    /** Env override of MIN_FREE_MB. */
+    ENV_MIN_FREE_MB: 'CREWLY_BOOT_PRESSURE_MIN_FREE_MB',
+    /** Env override of MIN_LAUNCH_SPACING_MS. */
+    ENV_LAUNCH_SPACING_MS: 'CREWLY_RESTORE_LAUNCH_SPACING_MS',
+};
 /**
  * Startup briefing content (#1345): wiki pages, recent lessons and open
  * owner threads an agent gets at registration, on top of its session memory.
@@ -1883,6 +1996,16 @@ export const WORKTREE_CONSTANTS = {
     INCLUDE_MAX_ENTRIES: 50,
     /** Heavy directories symlinked (not copied) from the repo root when present and untracked. */
     DEFAULT_SHARED_DIRS: ['node_modules'],
+    /**
+     * Workspace packages' install dirs (`backend/node_modules`,
+     * `packages/ui/node_modules`) are symlinked too when the lock file is
+     * identical in repo and worktree (crewly#1354): searched this deep.
+     */
+    NESTED_INSTALL_DEPTH: 2,
+    /** Lock files compared (first one present in the repo decides). */
+    LOCK_FILES: ['package-lock.json', 'pnpm-lock.yaml', 'yarn.lock'],
+    /** Directories never descended into while looking for nested install dirs. */
+    NESTED_INSTALL_SKIP_DIRS: ['node_modules', '.git', '.crewly', '.claude', 'dist', '.next'],
     /** Marker line for the entries this feature adds to `.git/info/exclude`. */
     EXCLUDE_MARKER: '# crewly worktree shared paths (#814) — never commit these',
     /** Timeout for local git commands (checkout of a large repo can take minutes under load). */
@@ -2753,7 +2876,7 @@ export const WORKTREE_JANITOR_CONSTANTS = {
     /** Timeout of the `lsof` that finds a file's writers (ms). */
     RUNAWAY_LSOF_TIMEOUT_MS: 10_000,
     /** One line added to every agent's start prompt. */
-    PROMPT_HINT: 'Disk hygiene: never let a command or script write an unbounded file. When you redirect or concatenate output (`> file`, `>> file`, `cat a b > all.md`), cap it (`| head -c 50M`, `| head -n 100000`) and never build a file from its own output. A file in your scratch dir or worktree that grows past a few GB is stopped by Crewly and the owner is alerted. Worktree hygiene: deploy only from a commit that is on main (merge first, never from a side branch); remove your worktree once its ticket is done.',
+    PROMPT_HINT: 'Disk hygiene: never let a command or script write an unbounded file. When you redirect or concatenate output (`> file`, `>> file`, `cat a b > all.md`), cap it (`| head -c 50M`, `| head -n 100000`) and never build a file from its own output. A file in your scratch dir or worktree that grows past a few GB is stopped by Crewly and the owner is alerted. Worktree hygiene: deploy only from a commit that is on main (merge first, never from a side branch); remove your worktree once its ticket is done. In a new worktree, do not run `npm install` when package-lock.json is unchanged: symlink `node_modules` (root and per-package) from the main checkout instead (`ln -s <main>/node_modules node_modules`). Crewly strips node_modules from worktrees untouched for 3 days and archives untouched ones as a patch after 14 days.',
     /** State file under CREWLY_HOME remembering when the owner was last told about low disk. */
     STATE_FILENAME: 'disk-janitor-state.json',
     /** File (in CREWLY_HOME) remembering which unmerged worktrees their lead was already told about. */
@@ -2792,8 +2915,12 @@ export const WORKTREE_JANITOR_CONSTANTS = {
     REGENERABLE_DIRS: ['node_modules', '.next', 'dist', 'build', '.turbo', 'coverage', '.cache', 'logs'],
     /** File suffixes treated as regenerable output (logs). */
     REGENERABLE_FILE_SUFFIXES: ['.log'],
-    /** Directories deleted from kept (dirty / not-merged) worktrees under low disk — reinstalled by `npm install` / rebuilt by `next build`. */
-    RECLAIM_DIRS: ['node_modules', '.next'],
+    /**
+     * Directories deleted from kept (dirty / not-merged) worktrees by the strip
+     * tier — reinstalled by `npm install` / rebuilt by the build. A directory
+     * git tracks anything under is never deleted (a repo that commits `dist`).
+     */
+    RECLAIM_DIRS: ['node_modules', '.next', 'dist', '.turbo', '.cache'],
     /** A kept worktree must be idle this long before its regenerable dirs are reclaimed under low disk (24 hours). */
     RECLAIM_MIN_IDLE_MS: 24 * 60 * 60 * 1000,
     /** How deep under a worktree reclaimable dirs are searched for (root = 0; `web/node_modules` = 1). */
@@ -2818,6 +2945,76 @@ export const WORKTREE_JANITOR_CONSTANTS = {
     GH_PR_LIMIT: 20,
     /** Max bytes of command output kept (stdout/stderr each). */
     MAX_OUTPUT_BYTES: 16 * 1024 * 1024,
+    // --- Age tiers for kept (unmerged / dirty) worktrees (crewly#1354) ---
+    /** A kept worktree untouched (no file, index or commit newer) for this long loses its regenerable dirs (3 days). */
+    STRIP_MIN_IDLE_MS: 3 * 24 * 60 * 60 * 1000,
+    /** Under low disk, or once its WorkItem finished, the strip tier needs only this much idle (1 day). */
+    STRIP_MIN_IDLE_LOW_DISK_MS: 24 * 60 * 60 * 1000,
+    /** A kept worktree untouched for this long is archived as a patch and removed (14 days). */
+    ARCHIVE_MIN_IDLE_MS: 14 * 24 * 60 * 60 * 1000,
+    /** Directory under CREWLY_HOME holding archived worktree patches (`<repo>/<name>-<date>.patch`). */
+    ARCHIVE_DIR: 'worktree-archive',
+    /** Cap on one archive patch; untracked files that would push it past the cap are listed in the header instead (64 MB). */
+    ARCHIVE_PATCH_MAX_BYTES: 64 * 1024 ** 2,
+    /** Cap on untracked files written into one archive patch. */
+    ARCHIVE_MAX_UNTRACKED_FILES: 2000,
+    /** Ref namespace that keeps the commits of an archived detached worktree alive (`refs/crewly-archive/<name>-<date>`). */
+    ARCHIVE_REF_PREFIX: 'refs/crewly-archive/',
+    /** How many levels below a worktree root count for its "last touched" time in the tier decision. */
+    TIER_MTIME_DEPTH: 2,
+    /** State file under CREWLY_HOME: per-worktree strip / archive / finished-work timestamps. */
+    TIER_STATE_FILENAME: 'worktree-tiers.json',
+    /** Finished-work records (task-end hook) older than this are dropped from the state file (30 days). */
+    TIER_STATE_MAX_AGE_MS: 30 * 24 * 60 * 60 * 1000,
+    /**
+     * Every delete added in crewly#1354 (strip, archive, transcripts, tmp, logs)
+     * is dry-run only until this long after the first real pass that saw it
+     * (one day): the first pass logs what it would do, the next day it does it.
+     */
+    RETENTION_ARM_DELAY_MS: 24 * 60 * 60 * 1000,
+    // --- Transcript retention (Claude Code conversations Crewly launched) ---
+    /** An agent transcript (and its sibling tool-results dir) untouched for this long is pruned (14 days). */
+    TRANSCRIPT_MAX_AGE_MS: 14 * 24 * 60 * 60 * 1000,
+    /** The transcript sweep runs at most this often (24 hours). */
+    TRANSCRIPT_SWEEP_INTERVAL_MS: 24 * 60 * 60 * 1000,
+    /** Bytes read from the head of a transcript to find Claude Code's `agentSetting` entry. */
+    TRANSCRIPT_HEAD_BYTES: 256 * 1024,
+    /** Handover files newer than this are scanned for conversation ids (30 days). */
+    TRANSCRIPT_HANDOVER_SCAN_MAX_AGE_MS: 30 * 24 * 60 * 60 * 1000,
+    /** Kinds that handover file names carry between the session name and the timestamp (`<session>-<kind>-<ISO>.md`). */
+    HANDOVER_KINDS: ['restart', 'runtime-switch', 'runtime-revert', 'runtime-upgrade'],
+    /** Sub-directory of a Claude home that holds per-cwd transcript folders. */
+    TRANSCRIPT_PROJECTS_DIR: 'projects',
+    /** Transcript file extension. */
+    TRANSCRIPT_EXT: '.jsonl',
+    // --- CREWLY_HOME/tmp and logs retention ---
+    /** Directory under CREWLY_HOME that agents and skills use as scratch space. */
+    TMP_DIR: 'tmp',
+    /** A top-level entry of CREWLY_HOME/tmp untouched for this long is removed (7 days). */
+    TMP_MAX_AGE_MS: 7 * 24 * 60 * 60 * 1000,
+    /** Sub-directories of CREWLY_HOME/tmp swept per child with their own age (ms); the directory itself stays. */
+    TMP_CHILD_POLICIES: {
+        'slack-files': 30 * 24 * 60 * 60 * 1000,
+        'slack-images': 30 * 24 * 60 * 60 * 1000,
+        'slack-pdfs': 30 * 24 * 60 * 60 * 1000,
+        'crewly-web-release': 2 * 24 * 60 * 60 * 1000,
+        'skill-output': 7 * 24 * 60 * 60 * 1000,
+    },
+    /** How many levels below a tmp entry count for its "last touched" time. */
+    TMP_MTIME_DEPTH: 2,
+    /** The tmp / logs sweep runs at most this often (6 hours). */
+    TMP_SWEEP_INTERVAL_MS: 6 * 60 * 60 * 1000,
+    /** Cap on CREWLY_HOME/logs/archive; the oldest archives go first (300 MB). */
+    LOG_ARCHIVE_CAP_BYTES: 300 * 1024 ** 2,
+    /** Cap on CREWLY_HOME/logs/sessions; the oldest logs of sessions that are not live go first (100 MB). */
+    LOG_SESSIONS_CAP_BYTES: 100 * 1024 ** 2,
+    /** A log written within this long is never deleted by the cap (1 hour). */
+    LOG_CAP_MIN_AGE_MS: 60 * 60 * 1000,
+    // --- Low-disk thresholds relative to the disk size (crewly#1354) ---
+    /** Low disk is free < max(LOW_DISK_BYTES, this fraction of the disk) (8 %). */
+    LOW_DISK_FRACTION: 0.08,
+    /** Critical is free < max(CRITICAL_DISK_BYTES, this fraction of the disk) (3 %). */
+    CRITICAL_DISK_FRACTION: 0.03,
 };
 /**
  * Constants for cron next-run computation (`getNextRunTime`).