docs/generated/vocabulary.mdgenerated from code
Reference · 56 min read
Constants: limits
Vocabulary: Constants in limits.
Read as
Source: crates/lvcs-vocab/src/limits.rs.
| Name | Value | Meaning |
|---|---|---|
BULLETIN_SECS | 300 | How long notifications stay in a document's recent-notification bulletin (discussion:owner/438). |
ANCHOR_WAIT_MS | 10_000 | How long a received notification waits for its range's anchor to sync. A notification rides gossip and the change it anchors on rides the document's sync stream, so it may arrive first. Until this daemon's copy of the document resolves the range, the notification is held back and resolved again at each render; past this it is delivered with no range. Milliseconds. Not configurable. |
JOURNAL_SET_ASIDES_KEPT | 0 | How many set-aside journal generations a project keeps beside its journal (discussion:owner/731): none before v0.1.0, each deleted once the fresh journal holds a record. |
JOURNAL_TORN_KEPT | 1 | How many torn journal tails (journal.cbor.torn-<n>) a project keeps, the newest (discussion:owner/731). |
JOURNAL_SWEEP_MAX_FILES | 256 | The most files one opening's journal sweep deletes (discussion:owner/731): past it, the next opening takes the rest, so a sweep's record stays small however many deletes failed before. |
TREE_PENDING_RENAMES_MAX | 64 | Most watcher rename pairs a tree file machine keeps while a save or namespace move is awaiting its answer (f-break/6). The oldest pair then falls back to ordinary reads of both paths, preserving disk bytes at the possible cost of document identity. One watcher batch can contain many pairs, but 64 outstanding pairs bound its durable context. |
TREE_REFUSED_BODIES_MAX | 1_024 | Most refused disk bodies the tree remembers for delayed watcher reads. The oldest falls back to ordinary disk judgment on overflow. |
TREE_REFUSED_BODIES_PER_PATH_MAX | 8 | Most refused disk bodies kept for one path before its oldest is evicted. |
ADD_WAIT_MS | 30_000 | How long add waits for main to hold the file it saves before it is answered recover=wait (o-decouple/1, discussion:owner/653). |
QUIET_MS | 10_000 | How long main's visible frontier stays unchanged before it is quiet. A runner serving a change action's channel then prepares a job for it (discussion:owner/157 and discussion:owner/161), and navigator reviews take the same debounce. lvcs.toml [runner] quiet_ms may raise it, up to QUIET_MAX_MS. |
CHANGE_CADENCE_MS | 120_000 | The longest main's visible frontier waits for a change job while main keeps moving. A change runner opens a job at the earlier of its quiet period's end and the first boundary of this period (a multiple of it since the Unix epoch, which every runner shares) after main's first move since its last job opened, so steady saves never starve it and runners attest the same frontier; a [runner] quiet_ms above it raises it to quiet_ms. Not configurable. |
CHANGE_CADENCE_SLACK_MS | 2_000 | How late after a cadence boundary a change runner's Tick may come and still open a job there. A runner busy at the boundary, or woken later than this, waits for the next boundary, so the runners open their jobs together, on the frontier main holds then, and reach a quorum on it. Not configurable. |
STOP_RETRY_MS | 5_000 | The retry_after_ms of a pending stop_gate (discussion:owner/217). Its caller asks again after this long. Not configurable. |
NOTIFICATIONS_WAIT_DEFAULT_MS | 600_000 | How long notifications.wait waits with no timeout_ms ( discussion:owner/424): ten minutes, after which it answers with no notification and its caller runs it again. |
NOTIFICATIONS_WAIT_MAX_MS | 3_600_000 | The longest timeout_ms notifications.wait takes (discussion:owner/424): an hour; a longer one waits this long. |
QUIET_MAX_MS | 600_000 | The most lvcs.toml [runner] quiet_ms raises QUIET_MS to (discussion:owner/161). |
VISIBLE_COALESCE_MS | 1_000 | The shortest interval between two moved frontiers sent to a runner serving change channels. Its daemon coalesces renders of main's file documents into at most one Visible per interval (discussion:owner/161). Not configurable. |
RANGE_CONTEXT_MAX_CHARS | 80 | Code points a range's near line keeps (discussion:owner/160): git's hunk-header line, which git cuts at 80 bytes; a code-point cut never splits a character. Not configurable. |
DISK_BASE_RING | 4 | Recent disk bases the tree keeps per file as DiskEdit candidates (discussion:owner/47). |
SAVE_SETTLE_MS | 250 | How long a changed file that reads as zero bytes stays unchanged before the tree reads it where the file watcher reports no close after a write (FSEvents, kqueue, Windows; discussion:owner/131, discussion:owner/150, and discussion:owner/464), so a truncate followed by a write within it is one save. Every other change settles after [SAVE_SETTLE_QUIET_MS]. inotify reports the close, and there a file is read at it. Milliseconds. Not configurable. |
IMPORT_TURNS | 4 | How many of a project's documents import a sync batch at once (discussion:owner/453, net-join/6): an import allocates heavily, and many at once on every core contend on a process-wide allocator lock until none finishes. The rest wait for a turn, in order. Not configurable. |
IMPORT_HEAVY_BYTES | 64 << 10 | The size of a sync batch that takes an import turn ([IMPORT_TURNS]), in bytes of its payloads: an edit's few bytes never wait behind a join's imports. Not configurable. |
SAVE_SETTLE_QUIET_MS | 10 | How long a changed file's size and modification time hold still before the tree reads it where the file watcher reports no close after a write, each new event for the path starting it again; a file that then reads as zero bytes waits [SAVE_SETTLE_MS] instead (discussion:owner/464 and 467). A file whose size or modification time moved during the read is read again once it has held still this long. Milliseconds. Not configurable. |
UNSEEN_RENDER_MS | 50 | How long after a render the tree wrote landed a file may still have been written without its author seeing that render (discussion:owner/398). A text save with no hook's read whose bytes were written (the file's modification time) before the render's landed write's modification time plus this is never based on that render, nor on any write in flight. It covers an editor's gap between checking the disk and writing its file (tens of milliseconds for a file of a few hundred KB) and the filesystem's timestamp granularity (the coarse clock's tick, 4 ms at most, on Linux; finer on APFS and NTFS), and stays far below the time a human takes to reload, edit, and save. Measured on the files' own modification times, so the watcher's delay and SAVE_SETTLE_MS never widen it. Milliseconds. Not configurable. |
HELD_RETRY_MS | 250 | How long the tree waits before writing a file again after another process held it open against the write (Windows: a sharing violation or access denied, the held-rename row refining discussion:owner/252). Milliseconds. Not configurable. |
HELD_DEADLINE_MS | 10_000 | How long the tree keeps writing a held file again, from the first time a write of that render was held, before it stops and reports the path as write_held (the held-rename row refining discussion:owner/252); a later render starts a fresh wait. Milliseconds. Not configurable. |
INTENT_TTL_MS | 60_000 | How long a peer's declared intent to change a file in the host's tree stands (intent, row 440): the tree's next change of the file within it is that peer's, against the base it named. Milliseconds. Not configurable. |
TRANSPORT_PANICS_MAX | 3 | How many panics of a project's transport within TRANSPORT_PANIC_WINDOW_MS end it for good (discussion:owner/676): each earlier one restarts it with a fresh swarm, and this one fails the project. Not configurable. |
TRANSPORT_PANIC_WINDOW_MS | 600_000 | The window TRANSPORT_PANICS_MAX counts a transport's panics in (discussion:owner/676). Milliseconds. Not configurable. |
TRANSPORT_RESTART_WAIT_MS | 1_000 | How long a panicked transport waits before it restarts, times the panics within TRANSPORT_PANIC_WINDOW_MS so far (discussion:owner/676). Milliseconds. Not configurable. |
MTIME_SLACK_MS | 2_000 | The coarsest modification-time granularity among the filesystems the tree may live on (HFS+ and FAT keep whole seconds): the slack an intent's window allows a file's modification time. Milliseconds. Not configurable. |
PEER_SIGN_PER_SEC | 100 | Changes per second the daemon signs for one peer key, the owner's and runners' included (discussion:owner/146): a request whose op signs is refused rate_limited past it, and a tree save waits for a token. The same for every peer: a tenfold margin over the gating suite's busiest signer, the owner, at 8 a second (discussion:owner/146). Raisable or lowerable in protected lvcs.toml [limits]. |
PEER_SIGN_BURST | 1_000 | Burst of changes the daemon signs for one peer key above PEER_SIGN_PER_SEC (discussion:owner/146): more than ten times the longest back-to-back run of signs the gating suite makes (15). |
PEER_GOSSIP_PER_SEC | 50 | Gossip messages per second forwarded from one signer, the owner included, before the excess is dropped, and the rate a local announce or publish takes its token at (discussion:owner/146). |
GOSSIP_BURST | 100 | Burst of gossip messages above the per-second rate, for any signer (discussion:owner/146). |
GOSSIP_FRESH_MS | 30_000 | How far an envelope's signed ts may be from the shell's now_ms, in either direction, before the gossip actor ignores it as stale (discussion:owner/112); an accepted envelope is remembered until ts + GOSSIP_FRESH_MS, so a replay within that span is a duplicate and a later one is stale. 30 s: the clock skew between NTP-synchronized nodes plus gossipsub propagation, which holds nothing longer than its few-second message cache. Milliseconds, the unit of the envelope's ts. Not configurable. |
GRANT_RENEW_RETRY_MS | 10_000 | How long the grants actor waits before asking again for the renewal of a windowed grant it already asked to renew (discussion:owner/148): the issuer's daemon dropped that renewal, or its caps render hasn't yet listed the renewal in the grant's place. Milliseconds, the unit of the renewal tick's now_ms. Not configurable. |
GRANT_TTL_MIN_SECS | 10 | The shortest window grant issues (discussion:owner/148): a ttl_secs below it is refused INVALID_PARAMS, so an issuer's daemon renews a lineage at most once every GRANT_TTL_MIN_SECS / 2 seconds, each renewal one more grant in the caps document. Seconds. Not configurable. |
QUORUM_DEFAULT | 1 | Attestations from distinct runner devices required to move a change action's channel (discussion:owner/297). Raisable, never lowerable. |
QUORUM_PROTECTED | 2 | Attestations from distinct runner devices a frontier needs in each change channel a gated action's after reaches before it takes it, and in every change channel before a protected-path approval counts (discussion:owner/297). Raisable, never lowerable. |
ACTOR_IDLE_SECS | 300 | A document actor with no open handle, subscriber, or sync node for this long is stopped (discussion:owner/584). |
LOADED_BYTES_BUDGET | LOADED_ROOM_BYTES - LOADED_ROOM_BYTES / 4 | The stored bytes (their ledger files' sizes) of the documents a project's daemon keeps loaded before it asks settled ones to rest early (session-29a3231b/2, under discussion:owner/632's memory cap): a load that takes the total past it asks up to 128 oldest loaded documents at once, each resting only when asked twice with no step between. 32 MiB (session-9468d544/36): with a 64 MiB budget a 601-document, 256 KiB local import exceeded discussion:owner/632's 1 GiB daemon cap even with 128 concurrent rest asks. The host retained up to 75 MiB of stored documents and peaked at 1,064,584 KiB RSS. Starting the rest sweep earlier gives import and sync room for in-flight allocations. The e2e cap assertion checks the resident result. In memory's units since owner/776 as amended: each loaded document counts for its stored bytes at [RESIDENT_FACTOR], at least [LOADED_DOC_MIN_BYTES] (a measured resident figure), against three quarters of the room ([LOADED_ROOM_BYTES], the resident mark), 672 MiB at the 2 GiB cap; a stored-scale budget against a resident-scale floor held about 230 small documents where the mark holds some 3,000. Not configurable. |
RESIDENT_FACTOR | 10 | How much memory a loaded document holds per stored byte of its ledger, until the canary measures it (caps-transport on owner/762: a loaded Loro document is about ten times its encoded size). A small ledger's document holds at least [LOADED_DOC_MIN_BYTES] whatever its size. Not configurable. |
LOADED_ROOM_BYTES | crate::budget::DAEMON_IDLE_RSS_BYTES / 2 | The memory of loaded documents the rester holds before the bars give way (owner/776, option C): the resident mark, half of DAEMON_IDLE_RSS_BYTES, 896 MiB at the 2 GiB cap, each document counted as [LOADED_BYTES_BUDGET] counts it. Past it a barred document rests, at most one per loaded document a sample; a load waits only under the byte budget's pressure. Not configurable. |
LOADS_IN_FLIGHT_MIN | 4 | How many document loads the daemon's rester admits past LOADED_BYTES_BUDGET at once, each counted until it rests (daemon-memory/7): a load that would take the loaded bytes past the budget waits its turn, in order, for a rest to make room, unless fewer than this many loads admitted past the budget are still loaded. So a project can always open, and one document larger than the budget loads alone, while the daemon's loaded documents stay within the budget and this many documents more, however many a join, an open, or an init loads. The caps, the namespace, and the approvals, which every other load waits on, never wait. Not configurable. |
LOADED_DOC_MIN_BYTES | 128 << 10 | The least a loaded document counts for against LOADED_BYTES_BUDGET, whatever its ledger's stored bytes (daemon-memory/7): a loaded file document holds about 118 KB of heap however small its ledger (heaptrack of a 3,500-file agentflare init, about 10 KB stored each), so counting stored bytes alone let some 2,500 small documents load at once. 128 KiB, that measurement rounded up. Not configurable. |
COMPACT_SECS | 600 | Interval of the compaction tick that makes a document actor emit a compacted save (discussion:owner/582). |
PRESENCE_TTL_SECS | 30 | A peer counts as present on a document until this long after its last presence or edit there (discussion:owner/438). |
ACTIVITY_RECENT_SECS | 300 | How far back project.summary counts a project's activity ( discussion:owner/303). |
HOST_GATHER_MS | 2_000 | How long a host op (project.summary, inbox) waits for each open project's answer before it lists that project with no data ( discussion:owner/303). |
HOST_SESSIONS_PER_KEY | 8 | Most host-level host node streams one device key holds at once (discussion:owner/300). |
LOCAL_CONNECTIONS_MAX | 1_024 | Most connections the per-user socket holds at once (discussion:owner/564): one past them is closed unread. Only the daemon's own user connects there, so this bounds a runaway client, not a stranger. Not configurable. |
REQUEST_VALUES_MAX | 65_536 | Most JSON values one request line may hold (discussion:owner/564), counted before it is parsed: each is tens of bytes once parsed, so this bounds a line's parsed size to a few times its own. An edit's text is one value however long. Not configurable. |
SUBSCRIPTIONS_PER_CONN_MAX | 256 | Most document and event subscriptions one connection holds at once (discussion:owner/561): one more is refused rate_limited, and the client unsubscribes one first. Each is held by the project's router and pushed to once per change it matches. Not configurable. |
ROUTER_FILE_WAITERS_MAX | 1024 | Most questions the router holds waiting for a path the namespace hasn't named yet (discussion:owner/653): one more is answered at once as a path it doesn't hold is. Each asks about a protected path, a few per project. Not configurable. |
MAIN_VERSIONS_KEPT | 200 | Most of main's newest versions the router keeps (discussion:owner/687): older ones stay reachable through blame and the actions ledgers. A few hundred bytes each. Not configurable. |
MAIN_VERSIONS_MAX | 50 | Most versions one main.versions answers (discussion:owner/687). |
MAIN_VERSION_PATHS_MAX | 20 | Most paths one version names (discussion:owner/687); a save of more counts the rest. |
HOST_CONNECTIONS_MAX | 256 | Most connections the host node holds at once, from any node (discussion:owner/557): one more is refused at the handshake. Not configurable. |
HOST_CONNECTIONS_PER_NODE | 4 | Most connections the host node holds from one node id (discussion:owner/557). Not configurable. |
HOST_PENDING_CONNECTIONS_MAX | 32 | Most host node connections still in their handshake at once (discussion:owner/557). Not configurable. |
HOST_NEGOTIATING_STREAMS_MAX | 16 | Most streams of one host node connection negotiating their protocol at once (discussion:owner/557). Not configurable. |
HOST_HELLO_DEADLINE_MS | 10_000 | Milliseconds a host node stream has, from its challenge, to send its first request (discussion:owner/557): one that sends none in time is closed, so a stream that never says hello holds nothing for long. Not configurable. |
HOST_REACH_WAIT_MS | 30_000 | Milliseconds host.enable waits for a non-loopback address before answering with loopback alone. |
HOST_HELLO_LINE_MAX_BYTES | 64 << 10 | The longest line a host node stream may send before its first request is answered (discussion:owner/557): a hello is a key, a signature, and a root, so a longer one closes the stream. Its later lines may be as long as any connection's, but none is read until the first is answered. Not configurable. |
HOST_HELLOS_HELD_MAX | 1_024 | Most hellos the host actor holds while the daemon loads and starts (discussion:owner/653; secfix/25); one more is refused rate_limited. Not configurable. |
HOST_ASKS_HELD_MAX | 64 | Most init, join, and project requests the host actor holds on one root until the close or join in flight there ends (discussion:owner/653; secfix/25); one more is refused to try again. Not configurable. |
HOST_SWITCHES_HELD_MAX | 64 | Most host.enable or host.disable requests the host node switch holds until the switch in flight settles (discussion:owner/653; secfix/17); one more is refused to try again. Not configurable. |
NODE_LOAD_ASKS_HELD_MAX | 1_024 | Most node.load askers one node's keeper holds until its fetch ends (discussion:owner/653; secfix/18); one more is answered that the node's load is unavailable. Not configurable. |
GRANTER_ASKS_HELD_MAX | 1_024 | Most questions a project's Granter holds waiting for a caps render (discussion:owner/653; secfix/6): one more is answered that it can't wait. Not configurable. |
GRANTER_RENDER_WAIT_MS | 60_000 | How long a project's Granter holds a question waiting for the caps document's first render (discussion:owner/787): past it, the question is answered that it can't wait, so a caps document that never renders keeps no document from opening for ever. Not configurable. |
MAILBOX_HELD_MAX | INBOX_MAX_PER_SENDER | Most notifications a mailbox holds for a peer no caps render has named yet (discussion:owner/653; secfix/8): as many as one sender may keep unread (INBOX_MAX_PER_SENDER), which covers what a new key is told between its file landing and the render that names it, and every render ends the wait. One more is dropped with a warning. Not configurable. |
GOSSIP_HELD_MAX | 256 | Most received gossip messages a project holds for the caps render that may admit them (discussion:owner/653; secfix/7), counted together: those whose signer no render admits yet, and those out of the scope rendered so far. A grantee's first notices race its grant by the time one caps document takes to sync, so a few per new grantee are enough; at GOSSIP_MAX_BYTES each this is at most 16 MiB, and signed notices are far smaller. One more is dropped as it was before holding. Not configurable. |
RUN_HISTORY_MAX | 256 | Most runs the runs actor keeps per channel, newest kept ( discussion:owner/303). |
CHANNEL_MOVES_MAX | 1_024 | Most moves of one channel's value the runs actor keeps for channel.history, newest kept (discussion:owner/303). |
RUN_LOG_TAIL_LINES | 50 | Most lines of a run's log run.view answers (discussion:owner/312). |
STEP_LOG_TAIL_LINES | 20 | Most lines of one step's masked output a run's runner keeps for run.view and the run push (discussion:owner/312). |
STEP_ERROR_LINES_MAX | 40 | Most lines that look like errors a run's runner keeps of one step's masked output, wherever in it they came, the first and the last half of them kept (discussion:owner/422 and discussion:owner/481). |
STEP_ERROR_LINES_SHOWN | 10 | Most of a step's kept error lines its tail shows ahead of its last STEP_LOG_TAIL_LINES lines, in run.view and the run push (discussion:owner/422). |
SEARCH_RESULTS_MAX | 200 | Most lines one search answers (discussion:owner/303). |
CHANGE_SIGNERS_MAX | 64 | Most change ids one change.signers asks about (discussion:owner/685). |
SEARCH_FILE_MAX_BYTES | 4 * 1024 * 1024 | Largest file search reads, in bytes; a larger one is skipped (discussion:owner/303). |
GIT_IMPORT_POLL_MS | 3_000 | How often the owner's daemon reads each mapped git branch's head, to import commits made outside LVCS (discussion:owner/309). |
GIT_IMPORT_FETCH_MS | 15_000 | The least time between two fetches of one [git] import ref by the owner's daemon (discussion:owner/318); a poll between them reads the ref as last fetched. |
RESOURCE_STEP_SECS | 60 | The step between two of the daemon's resource samples (discussion:owner/303). |
RESOURCE_SAMPLES_MAX | 60 | Resource samples the daemon keeps in memory: one hour of steps (discussion:owner/303). |
NODE_LOAD_SAMPLES_MAX | 60 | Most resource samples one node.load answer carries: the latest of the answering daemon's steps of RESOURCE_STEP_SECS (discussion:owner/400). |
JOIN_SETTLE_SECS | 120 | How long agentflare join waits with no document settling before it ends incomplete (discussion:owner/215). It waits that long for a connected node to settle the documents still unsettled, or for the caps document to have a connected node again, then exits 1. Exchanges that settle nothing (a refused node's redials, live edits on settled documents) do not extend the wait. A node that leaves mid-sync settles none of its documents, and a redial of a bootstrap node comes at most 9 transport ticks of 5 s (45 s) apart, so this is more than twice that wait plus margin: a cut that heals within about a minute still re-syncs. Not configurable. |
JOIN_UNANSWERED_SECS | 60 | How long a join whose every bootstrap address failed its dials, with no node of the project ever identified, keeps redialing before it fails (launch-qa/8, r3-e2e/17). The transport redials the bootstrap nodes on its doubling backoff meanwhile, so a node that answered slowly (a slow link, its own dials closed before identify) is reached; an address nothing answers at still fails well inside JOIN_BOUND_SECS. Not configurable. |
JOIN_BOUND_SECS | 300 | How long agentflare join runs at most before it ends (discussion:owner/336). A join whose caps document named the project's root by then ends incomplete, its project record written, so the daemon's opening that follows finishes the sync (discussion:owner/215): a busy project's live edits may keep its documents from ever going quiet, and its settles keep JOIN_SETTLE_SECS from ever passing. A join that heard no root by then fails. Not configurable. |
RUN_READ_RETRY_MS | 2_000 | How long a gated run waits before it reads a document again whose read was refused because this node doesn't hold the version the run's frontier names yet (discussion:owner/357): it reads again until RUNNER_READ_WAIT_MS after its reads went out, then fails naming the document. Not configurable. |
STEPS_IN_FLIGHT | 256 | Most steps of one machine whose journal entries aren't acknowledged yet, or whose outputs haven't run (discussion:owner/439): the machine steps on while earlier steps close in order, so the journal's writer batches their entries; past this, it waits. Not configurable. |
DOC_STEP_BATCH_MAX | 32 | Most pieces of one document's waiting work one run takes ( discussion:owner/354): its events right behind each other, stepped in order on one blocking thread, so a backlog costs one shell turn a batch. Not configurable. |
WIRE_MAX_BYTES | 16_777_216 | Largest encoded envelope a node accepts, on the sync stream or over gossipsub; a larger one is rejected malformed before it is decoded. Chosen at the M3 freeze to bound decoding of untrusted input. |
JOURNAL_RECORD_MAX_BYTES | 268_435_456 | Largest journal record body, 256 MiB: a record is one actor event with its snapshots, and its largest payload is one sync message or change (each at most WIRE_MAX_BYTES) or one file's text a tree saves, so the bound leaves 16 times the wire's room. A write of a larger record fails (and stops the daemon); at start, a length prefix past the end of the journal is a torn tail only when it is at most this, and any larger prefix refuses the start. |
JOURNAL_ROTATE_BYTES | 67_108_864 | The bytes of records past its head (format header and segment record) a journal file holds before the daemon closes it as segment journal.cbor.1 and starts a fresh one (discussion:owner/319), so a daemon's journal stays bounded whatever its uptime, and its rotations stay linear in the bytes it writes however large a head is (g-rename/1). |
JOURNAL_KEEP_SEGMENTS | 32 | Closed journal segments kept beside the current one (discussion:owner/319); older ones are deleted, so a journal holds at most JOURNAL_KEEP_SEGMENTS + 1 files of JOURNAL_ROTATE_BYTES, each with its head (the live incarnations it carries) and one batch more. |
JOURNAL_SEGMENT_FACTS | 256 | Fact records (discussion:owner/233) the journal writer keeps from the run so far, newest last: an entry naming one its segment doesn't hold has it written ahead of it, so a document entry naming a recent fact still finds it. A segment's head copies none (g-rename/1). |
JOURNAL_SEGMENT_FACT_BYTES | 8_388_608 | The most bytes of fact records the journal writer keeps (JOURNAL_SEGMENT_FACTS), 8 MiB; the newest is always kept. An entry naming an older fact in a segment that begins mid-run carries its incarnation, which that segment doesn't replay (discussion:owner/319). |
JOURNAL_SHARED_MIN_BYTES | 256 | Shortest byte or text string an actor's journal entry shares within a segment (discussion:owner/342): a later entry of the same segment holding the same string names it by hash instead of copying it, so a file body each tree reports, saves, and renders is journaled once per segment. |
JOURNAL_BODY_MIN_BYTES | 4096 | Fewest bytes of a string an actor's journal entry holds that is written once to the journal's content-addressed body store, .lvcs/bodies/<hash> (discussion:owner/452), and named by its SHA-256 in every entry, across segments and runs alike; a shorter string of at least JOURNAL_SHARED_MIN_BYTES is shared within its segment (discussion:owner/342). |
JOURNAL_SYNC_MS | 1000 | Most milliseconds the journal's written records go unsynced ( discussion:owner/455): a batch is acknowledged once written, and the writer syncs the file with a plain fsync at most this long after its first unsynced write, and fully at shutdown. A power cut loses at most this window, recovered as a torn tail. |
POLICY_PENDING_MAX | 256 | Most requests of one project that wait for main's approved lvcs.toml (discussion:owner/654), counted apart: saves, one per path, and the scans, messages, and approvals the allowlist holds. The oldest past it is refused policy_unknown; a refused save is held and goes through once the policy is known. |
POLICY_WAIT_MAX_MS | 30_000 | Longest a request waits for main's approved lvcs.toml (discussion:owner/812, amending owner/653): past it the allowlist refuses it policy_unknown with recover=wait, retry_after_ms, and status's reason, never leaving its caller to time out. 30 s: well under REQUEST_TIMEOUT_MS, so the caller hears why. Saves keep holding. |
PROJECT_MAX_DOCUMENTS | 262_144 | Most documents a project holds, 2^18 (discussion:owner/111): the largest project a node syncs, and the most documents the sync actor counts as joined with one node; one more disconnects the node over_limit. Not configurable. |
SYNC_HELD_MAX_MESSAGES | 2 * PROJECT_MAX_DOCUMENTS | Most sync frames the sync actor holds for one node while the node's .lvcs/caps round is unsettled (discussion:owner/111): this daemon's other documents' messages for the node, sent when the round settles. One more disconnects the node over_limit. 2^19: in one caps round an honest exchange holds at most one answer from each document joined with the node (at most PROJECT_MAX_DOCUMENTS), since a document sends a node nothing more for a document until the node answers unless the document's heads change, plus one message for each local change made during the round, which lasts only while the caps document handles the node's last caps message; twice the project bound covers both. Not configurable. |
INVENTORY_HOLDERS_MAX | 64 | Most peer keys one sync inventory names as held by its sender's daemon (discussion:owner/314); one naming more disconnects the node malformed. Not configurable. |
HOST_NAME_MAX_BYTES | 64 | Most bytes of the machine name one sync inventory carries ( discussion:owner/314); one carrying more disconnects the node malformed. Not configurable. |
SYNC_HELD_MAX_BYTES | 64 * WIRE_MAX_BYTES | Most bytes of sync frames (each frame's document address and payload) the sync actor holds for one node while the node's .lvcs/caps round is unsettled (discussion:owner/111); a frame that would pass it disconnects the node over_limit. 2^30 (1 GiB): 64 of the largest envelopes the wire admits, and 4 KiB on average for an answer from every document of the largest project. An honest round's answers carry heads, a Bloom filter of the changes since the last sync, and only the changes the node lacks, each change sent to a node once. Not configurable. |
SYNC_IN_FLIGHT_MAX_BYTES | 8 << 20 | Most payload bytes the sync actor hands one node's documents before they take them (session-29a3231b/2): once the node's frames delivered and not yet taken reach it, the transport stops reading the node's sync stream, so its flow control pushes back to the sender, until the documents take them down to [SYNC_IN_FLIGHT_LOW_BYTES]. 8 MiB: what one node's share of the daemon's in-flight byte budget admits, in raw payload bytes (twice the share over the budget's charge factor, about 6.2 MiB at the 2 GiB machine cap, rounded up; 4 MiB at the 1 GiB cap; discussion:owner/762, l4nma4df). A delivery past its node's share only waits for a permit holding its payload, so more in flight buys no throughput (daemon-memory/1: the dev-2 canary's eight stuck nodes held 263 MB at 32 MiB each); the daemon asserts the two agree. A frame is never refused for it: the frame read is delivered, however large, and only the next waits, so a node always gets one frame through. The worst case a daemon holds in delivered frames is its synced nodes times this plus one frame each. Not configurable. |
SYNC_STREAM_QUEUE_BYTES | 2 * WIRE_MAX_BYTES as u64 | Bytes of outgoing frames the transport queues for one sync stream, each with its 4-byte length, before the sync actor waits for room (session-29a3231b/2): two of the largest envelopes the wire admits. The sync actor writes a node at most this much ahead of what the stream's writer has written, and waits, holding what comes next, until the queue is down to half; the writer's measured room corrects the actor's count (which leaves out each envelope's seal and length) every round. A full queue never closes the stream: a dead link is closed by ping and the connection's own failure. Not configurable. |
SYNC_BLOCKED_MAX_BYTES | 4 * SYNC_STREAM_QUEUE_BYTES | Most bytes (each frame's address and payload) the sync actor holds for one node while it waits for room in the node's stream (session-29a3231b/2): four of the stream's queues. Past it, or past SYNC_HELD_MAX_MESSAGES frames, the node is disconnected over_limit: it reads too slowly for what this side has to send it. A node's documents answer one input at a time, so an honest node's backlog is a window of answers. Not configurable. |
SYNC_ANSWERS_DAEMON_MAX_BYTES | 192 << 20 | Most bytes of answers the daemon holds for its nodes at once (r3-memory/12): what each link's writer has written that its stream hasn't drained, and what it holds waiting for room. Each link's share is this divided by the nodes the daemon admits (NodesAdmitted), at least SYNC_ANSWERS_MIN_BYTES and at most SYNC_STREAM_QUEUE_BYTES and SYNC_BLOCKED_MAX_BYTES together, which a link never told keeps; past its share a node is disconnected over_limit, unpenalized, and its documents rejoin when its stream reopens. Beside the rester's resident mark (896 MiB), twice the in-flight budget (about 299 MiB), and the daemon's measured idle state (245 MB on the canary), 192 MiB keeps the daemon under its 1,792 MiB ceiling at the 2 GiB machine cap: about 1,620 MiB. It bounds the answers the daemon's links hold, which grow with the nodes it serves, not with the cap, so it stays a number of its own. Not configurable. |
SYNC_ANSWERS_MIN_BYTES | WIRE_MAX_BYTES as u64 | The least share of SYNC_ANSWERS_DAEMON_MAX_BYTES a link's answers get however many nodes the daemon admits: one of the largest envelopes the wire admits, so any single frame can always go to a node that holds nothing else. Not configurable. |
SYNC_IN_FLIGHT_LOW_BYTES | SYNC_IN_FLIGHT_MAX_BYTES / 2 | The payload bytes in flight to a paused node's documents at or below which the transport reads its sync stream again (session-29a3231b/2): half of [SYNC_IN_FLIGHT_MAX_BYTES] (discussion:owner/762, l4nma4df), so a paused node resumes once its documents took half of it, not after every frame. Not configurable. |
SENDER_FAULTS_MAX | 8 | Sender faults a node's one sync connection may carry before it is disconnected (discussion:owner/459): each signed change it sends that is unsigned, bad_signature, or wrong_peer is dropped and counts one. Past it, the node is refused as a protocol break is, and waits out the transport's doubling backoff before it is heard again (discussion:owner/320); a new connection starts at zero. A change whose signer the project doesn't name (no_capability: the owner's grant cited by another key, or a grant whose grantee is another key) counts one too, since no honest node ever holds one (discussion:owner/554). Refusals of a change itself (unauthorized, orphaned, equivocation) never count: an honest relay forwards them. Not configurable. |
LEDGER_REFUSED_MAX | 4_096 | Most refused peers one document remembers (discussion:owner/554): its durable peer → first refused counter map, which every sync payload carries. Only a change whose signer the project names (the owner, or the grantee of the grant its record cites) is refused here; a stranger's change is the sender's fault and nothing of it is kept. A refusal past this bound is not kept either: the change stays out, its sender is told over_limit, and its peer's later changes wait on it within the node's deferral bounds. A node's payload naming more refusals breaks the protocol. Not configurable. |
SIGN_RETRY_MAX | 5 | How many times a document asks again for the signature of an own change after this daemon's own keyring failed to sign it (discussion:owner/706): a locked keyring, an I/O error, a key not loaded yet. After the last one fails the change is dropped as after a write device's refusal. Raised only by a decision. |
SIGN_RETRY_BASE_MS | 500 | The wait before the first such retry; each later one waits twice as long, up to [SIGN_RETRY_CAP_MS]: 500 ms, 1 s, 2 s, 4 s, and 8 s, 15.5 s in all. |
SIGN_RETRY_CAP_MS | 8_000 | The longest wait before one such retry. |
UNREADABLE_PEERS_MAX | 64 | Most peers one document remembers holding changes it can't read (discussion:owner/671 rule 2): each peer's first unreadable counter and how many of its changes were set aside, in memory only. An unreadable change from one more peer is not remembered: its sender is told over_limit, no sender fault, and nothing is stored. A node's payload naming more is cut to this many. Not configurable. |
LEDGER_COMPACT_AFTER | 1_024 | How many of a document's signed changes may be past its compacted snapshot before a CompactTick writes a new one (discussion:owner/615): a load imports the snapshot, then at most about this many changes one by one or in one batch. Not configurable. |
LEDGER_IMPORT_OPS | 65_536 | Most Loro ops one batch of a document's load imports, a change holding more alone (g-rename/1): Loro checks a batch out through its richtext tracker at once, which for one 1.4 MB insert and the edits after it peaked at 1,088 MB, while batches of at most this many ops load every ledger of a 4,448-document project at no more than its largest document's size, and four times faster. Not configurable. |
LEDGER_IMPORT_BYTES | 4 * 1024 * 1024 | Most stored change bytes one batch of a document's load or rebuild imports, a larger change alone (daemon-memory/1): an op of a channel's ledger can carry a whole log or run record, so [LEDGER_IMPORT_OPS] alone let one batch hold a whole history's bytes at once, beside the decoded records and the Loro document, about three times a large ledger's stored bytes at its load. Not configurable. |
LEDGER_SEGMENT_MAX_BYTES | 4 * 1024 * 1024 | Size of a segmented entry ledger's newest segment, by its stored ledger's bytes on disk, at which the next append goes to a new segment (discussion:owner/638): a reader loads only the segments it reads, so memory follows what is read, not history. Raised only by a decision. |
APPROVALS_CARRY_MAX | 256 | Most approvals an approvals segment's carry holds for one protected path (discussion:owner/805): the newest, as many as a protected path remembers former approved heads, so a rollback to any of the last this many approved versions needs no new approval. Raised only by a decision. |
LEDGER_SEGMENT_RETAIN | 1 | How many segments of a segmented ledger with a carry kind a node keeps before the newest segment holding a carry, unless its plugin's manifest says another retain (discussion:owner/695): the ledger holds at most this many plus 2 segments on disk. |
LEDGER_EVIDENCE_MAX | 64 | Most equivocation evidence one document keeps (discussion:owner/554), each piece one refused peer's two signed changes; evidence past it, or past [LEDGER_EVIDENCE_MAX_BYTES], is not kept (its change is still refused while [LEDGER_REFUSED_MAX] has room). A node's payload carrying more pieces breaks the protocol. Not configurable. |
LEDGER_EVIDENCE_PER_SIGNER | 2 | Most equivocation evidence one document keeps against one signer (discussion:owner/554): one piece proves it equivocated, so no single key, however many peers it mints, crowds out the evidence against another. Not configurable. |
LEDGER_EVIDENCE_MAX_BYTES | WIRE_MAX_BYTES / 4 | Most change bytes the evidence of one document holds, both changes of each piece counted (discussion:owner/554): a quarter of the largest envelope, so the payload a joining node is sent, which carries all of it, always fits the wire. Not configurable. |
SYNC_JOINS_IN_FLIGHT | 32 | Most known documents a node's sync joins with one node at once that haven't settled with it yet (discussion:owner/569): serving a joiner, the rest wait their turn, each joining as one in flight settles, so the answers a join makes are a window, not every document at once. The count inventory sends is still every document. Not configurable. |
SYNC_UNSETTLED_MAX | 2 * SYNC_JOINS_IN_FLIGHT | Most documents a node's own frames bring in while documents are in flight with it (daemon-memory/19): joined and not settled since, in either direction, counted together; the window's own joins (SYNC_JOINS_IN_FLIGHT) count toward it but are bounded by the window, and a place is given up after SYNC_JOIN_STALE_TAGS as the window's are. A node's frame past it for a document not joined is not delivered; the document is kept, its address only (not at all when the node's heads match this side's), and joined by this side as places free, so the node sends what this side lacks again. Each such document is loaded here, about 0.8 MB resident on the canary: 64 hold about 50 MB a node, where a 0.2.14 seed's frames held 290 to 530 at once. Not configurable. |
SYNC_UNSETTLED_DAEMON_MAX | 192 | Most documents node-started joins hold in flight across the daemon (daemon-memory/19): each link's cap is this shared by the nodes the daemon admits (NodesAdmitted), at least SYNC_UNSETTLED_MIN and at most SYNC_UNSETTLED_MAX. At about 0.8 MB resident a loaded document, about 150 MB. Not configurable. |
SYNC_UNSETTLED_MIN | 16 | The least a link's cap on node-started joins falls to, however many nodes the daemon admits, so each still syncs. Not configurable. |
SYNC_JOIN_STALE_TAGS | 16 * SYNC_JOINS_IN_FLIGHT as u64 | How many tags a node's sync stream mints, its joins and deliveries, before a document in its join window that never settled gives up its slot (discussion:owner/569): its join stands, but the window moves on, so a frame its node dropped never freezes the window. Not configurable. |
SYNC_JOIN_STALE_MS | 30_000 | How long a node's join window may stand unchanged, its stream's writer free, before every slot in it is given up (caps-transport/3): the joins stand, but the window moves on. [SYNC_JOIN_STALE_TAGS] counts tags, and a window whose every answer was lost mints none, so without a time bound it would freeze the stream for good. 30 s: a third of the e2e gate's no-progress bound. Not configurable. |
STATUS_NET_WAIT_MS | 2_000 | The most status waits for its project's network facts before it answers without them, saying the network is busy (request-flood/6). The executor answers the question ahead of its sync backlog, so this binds only behind one long input step. 2 s: within status's 3 s median, so a status never waits out its 30 s bound on the network. Not configurable. |
SYNC_ASK_QUIET_MS | 30_000 | How long a document waits, after it asked a node to resend (resend: true) and its own have made no progress toward what the node's showed, before it may ask once more (discussion:owner/810 rule 3), so an ask lost with no stream reopen never stalls the pair. 30 s: above the catch-up gap of any flood (request_flood's backlog moves every second), under the e2e upgrade suite's 120 s round gap (idle_mixed_version's ROUND_GAP). Not configurable. |
SYNC_REFUSED_MAX_DOCUMENTS | 4 * SYNC_JOINS_IN_FLIGHT | Most documents a node's sync remembers a node sent frames for before it admitted the node (authz/2): each frame is dropped, never delivered, but once the node is admitted each remembered document joins it, so the node's offer of a document this daemon doesn't hold, made before the admission landed, is answered and not lost. An honest node offers at most its namespace and a join window before it hears back; past the bound a frame is only dropped. Addresses only, never payloads, so a node never admitted costs at most this many addresses. Not configurable. |
SYNC_WAITING_MAX_FRAMES | 8 * SYNC_REFUSED_MAX_DOCUMENTS | Most frames from a node whose admission is undecided that wait for it once [SYNC_REFUSED_MAX_DOCUMENTS] documents are remembered (secfix/11): past it, or past [SYNC_WAITING_MAX_BYTES], the node is disconnected over_limit. It bounds only what an unadmitted node makes a node hold. Not configurable. |
SYNC_WAITING_MAX_BYTES | WIRE_MAX_BYTES | Most bytes (addresses and payloads) of the frames waiting for a node's admission (secfix/11): one frame of the largest envelope fits. Not configurable. |
SYNC_RESUME_GRACE_MS | 90_000 | How long a node's sync stream stays suspended after it drops, in milliseconds, before every document leaves the node (discussion:owner/576): a stream that reopens within it resumes where it was, rejoining only the documents that hadn't settled. Above the reconnects seen (a relay circuit's in seconds, a restart's in tens of seconds). Not configurable. |
SYNC_SUSPENDED_MAX | 32 | Most nodes a daemon keeps suspended streams for at once (discussion:owner/576): past it the oldest suspended node's documents leave it. Each keeps only its link (documents joined, those unsettled, its incarnation). Not configurable. |
SYNC_STREAM_SILENCE_MS | 30_000 | Milliseconds a sync stream may go with no frame read from it before its holder closes it (discussion:owner/364). Each transport tick (5 s) writes a keepalive frame on every stream it holds, so only a stream whose node stopped reading, or whose link died unseen, goes quiet this long; the opener then opens a new one, which replaces a dead one the other side still holds. Not configurable. |
REFUSED_SAVE_VERSIONS | 128 | Most versions of a file, newest first, that a document weighs a tree save the tree can't base against (discussion:owner/365): a save whose body the file rendered as at one of them is a disk the tree's renders have not reached, and is answered with the file's render. Each version weighed renders the file once, so this bounds what a fresh tree's start costs. Not configurable. |
LANDED_VERSIONS | 16 | How many versions a tree remembers holding at one path (discussion:owner/668): a disk save's base is only ever one of them, and another author's line in all of them is one the writer saw. A buffer older than the oldest is taken as made from it. Not configurable. |
MERGED_SAVES_KEPT | 64 | How many of its merged saves a document remembers (discussion:owner/728): tree saves it committed with other text than the disk held, whose insertions name no author for the deletion guard. The newest are kept, in memory and at load; four times [LANDED_VERSIONS], so a guard window, which never starts before a tree's oldest landed version, sees every merged save in it unless more than this many were made in it. Not configurable. |
DEFER_MAX_MESSAGES | 64 | Most changes (and pieces of evidence) from one node one document holds deferred until it can judge them, their deps or grants known (discussion:owner/108 and discussion:owner/111). One more is dropped, never the sender's fault, and the node sends it again once the document answers with its have unmoved (discussion:owner/571): an honest first sync of a long document holds more than this before a joiner's grants render. Not configurable. |
CHANGE_MAX_BYTES | WIRE_MAX_BYTES / 2 | Most bytes one change may hold (discussion:owner/611): half an envelope, leaving room for its record, the payload, and evidence. A save or an edit that would make a larger one is refused change_too_large (a tree save of a file whose content is larger, since any change of it may carry it whole); a received one is malformed; and a document holding one from before skips it in sync, logged once per node. Not configurable. |
SYNC_ANSWER_BYTES | WIRE_MAX_BYTES / 4 | Most change bytes one sync answer carries to a node: a document whose changes a node lacks past this sends them a causal prefix at a time, the next once the node's answer reports its have moved, so every answer's envelope stays well under WIRE_MAX_BYTES however long the document's history. Each change counts its bytes plus SYNC_CHANGE_OVERHEAD_BYTES; an answer always carries at least one change. Not configurable. |
SYNC_CHANGE_OVERHEAD_BYTES | 1024 | What one change counts toward SYNC_ANSWER_BYTES beyond its own bytes: a generous bound on its record, deps, digest, signature, and encoding. Not configurable. |
DEFER_MAX_BYTES | CHANGE_MAX_BYTES | Most payload bytes from one node one document holds deferred ( discussion:owner/108 and 111): one change of the largest size a change may be, so any change can wait for its deps, and no node holds more in one document (discussion:owner/762). Not configurable. |
DIFF_HUNK_MAX | 4_096 | The most code points, both sides together, a code point diff compares inside one changed run of lines: a larger run is replaced whole, so no diff costs more than this squared per run, however much of a file a save rewrote. Not configurable. |
RUNNER_READ_WAIT_MS | 120_000 | How long a change runner waits for a document to answer one of its reads, a job's read of a file its actions read or a failure window's read, before it goes on without that document: a job then ends with nothing attested and its runs cancelled, releasing the executor, and a window names no lines of it. A document that never answers never holds the runner's executor for good. Not configurable. |
CHANNEL_QUERY_WAIT_MS | 30_000 | How often the shell starts a new period of the channels actor's open questions: a descent or overlap question still unanswered after a whole period is sent again, so an answer the shell never delivers can't hold a channel's value or a path's approved heads for good. Not configurable. |
CHANNEL_QUERY_WAIT_PERIODS_MAX | 16 | The most CHANNEL_QUERY_WAIT_MS periods the channels actor waits before it sends an unanswered question again: the wait doubles from one period at each sending, up to this. Not configurable. |
FOLD_READ_DEADLINE_MS | 60_000 | How long the actions plugin's folders' executor gives one ledger read of a folder (perf-mem, on the folders' one executor): a read core hasn't answered by then ends, its folder is told it failed and reads again FOLD_READ_RETRY_MS later, and the executor goes on with the other folders' reads, so one hung read never holds every channel's fold. Not configurable. |
FOLD_READ_RETRY_MS | 5_000 | How long a folder of the actions plugin waits after one of its ledger reads failed or passed FOLD_READ_DEADLINE_MS before it reads again, whether or not another event reaches its ledger. Not configurable. |
RUNNER_RETRY_MAX | 5 | How many times in a row a change runner tries a job on one frontier again when a document didn't answer its reads within RUNNER_READ_WAIT_MS: each try waits that long first, so a document that never answers costs at most this many of them; past it the job waits for a change, a new frontier, or a newly served channel, as any job that couldn't be prepared. Not configurable. |
CHECK_TIMEOUT_SECS | 600 | Wall clock lvcs sandbox gives one check command before it ends it (discussion:owner/121), and the CPU time it may use (RLIMIT_CPU, discussion:owner/134). |
CHECK_MEMORY_BYTES | 4 << 30 | Address space one check command may map (RLIMIT_AS, discussion:owner/134), and the memory.max of its run's cgroup leaf when the daemon has a delegated cgroup (the sandbox-hardening row): 4 GiB. Runners must agree on it for a quorum, since a check that fits under one limit and not another would pass on one runner and fail on the next. Not configurable. |
CHECK_PROCESSES | 512 | Processes one check command may run at once (RLIMIT_NPROC, discussion:owner/134), and the pids.max of its run's cgroup leaf when the daemon has a delegated cgroup (the sandbox-hardening row). Runners must agree on it for a quorum. Not configurable. |
CHECK_FILE_BYTES | 1 << 30 | The largest file one check command may write (RLIMIT_FSIZE, discussion:owner/134): 1 GiB. Runners must agree on it for a quorum. Not configurable. |
CHECK_OPEN_FILES | 16_384 | File descriptors one check command's processes may each hold open (RLIMIT_NOFILE, the sandbox-hardening row). Runners must agree on it for a quorum. Not configurable. |
CHECK_CORE_BYTES | 0 | The largest core file one check command may dump (RLIMIT_CORE, the sandbox-hardening row): none, so a crash never writes its memory, which may hold a gated run's secrets, to disk. Not configurable. |
CHECK_CPUS | 4 | CPUs' worth of time one check command may use at once, as the cgroup v2 cpu.max of its run's leaf when the daemon has a delegated cgroup (the sandbox-hardening row; with none, only RLIMIT_CPU bounds it). Not configurable. |
CHECK_LOG_TAIL_BYTES | 16_384 | Bytes at the end of a check's output kept in its job directory's log for the operator (discussion:owner/121); a failed action's action_failed notifications carry a readable text made from it, within CHECK_OUTPUT_TAIL_BYTES (discussion:owner/193 and discussion:owner/242). |
LOG_CHANGE_MAX_BYTES | 2 * CHECK_LOG_TAIL_BYTES + 8_192 | The most bytes of one signed change to an action's log document (discussion:owner/500): its one put of a failed run's log, at most CHECK_LOG_TAIL_BYTES of error lines and as much of tail, with room for the entry's ids and Loro's encoding; every node refuses a larger one. Not configurable. |
ENTRY_MAX_BYTES | 65_536 | The most bytes of one plugin entry's canonical CBOR (discussion:owner/514): an entry ledger's change appends one entry of at most this, and every node refuses a larger one. Not configurable. |
ENTRY_CHANGE_MAX_BYTES | ENTRY_MAX_BYTES + 4_096 | The most bytes of one change to an entry ledger: its one entry, its author, plugin, and time, and Loro's encoding. |
PLUGIN_ASSET_MAX_BYTES | 4 * 1024 * 1024 | The largest file plugin.asset serves (discussion:owner/520): a surface's module or one it loads, 4 MiB. |
LVCS_TOML_MAX_BYTES | 1 << 20 | The largest lvcs.toml the daemon parses (discussion:owner/563): a larger one is refused as one that doesn't parse, before TOML builds its tables. Not configurable. |
PLUGIN_MANIFEST_MAX_BYTES | 256 << 10 | The largest plugin manifest (plugin.toml) the daemon parses (discussion:owner/563). Not configurable. |
EVENT_READ_PAGE_MAX | 256 | Most events one event.read page holds (discussion:owner/562): a read asking more, or naming no page, gets this many and the position to read on from (next). Each event is at most ENTRY_MAX_BYTES, so a page is at most 16 MiB however long its ledger. Not configurable. |
PLUGIN_DORMANT_MIN_MS | 5_000 | The least time a dormant plugin serve stays down after it exited (discussion:owner/790), in milliseconds: events its wake filters match sooner start it once, when this has passed, so a burst of synced events wakes it once rather than once each. A wake time sooner than this is moved to it. Not configurable. |
PLUGIN_WAKE_FILTERS_MAX | 16 | The most wake filters one plugin.dormant names (discussion:owner/790): each is one event subscription of the plugin's key while its serve sleeps, so it stays well under SUBSCRIPTIONS_PER_CONN_MAX. Not configurable. |
PLUGIN_SAMPLE_MS | 10_000 | The step between two samples of a running plugin process's resident memory and CPU time (discussion:owner/790 item 3), in milliseconds: each keeper's executor reads its serve's (/proc/<pid>/status and stat on Linux, one ps on macOS), and the admitter's its admitted commands', and only a sample that moved (a MiB of memory, or PLUGIN_SAMPLE_CPU_STEP_MS of CPU) reaches the admitter. Bounded and cheap: one read per process per step. Not configurable. |
PLUGIN_SAMPLE_CPU_STEP_MS | 100 | The CPU time a plugin process spends between two of its samples that counts as activity (discussion:owner/790 item 3), in milliseconds: less is quiet, and isn't sent. Not configurable. |
PLUGIN_FIRST_PEAK_BYTES | 16 * 1024 * 1024 | What the plugin admitter charges a plugin process it has no peak for yet, a serve's first run or a command's first (discussion:owner/790 item 3), in bytes: 16 MiB, about a fresh serve's peak; its samples correct it. Not configurable. |
PLUGIN_REST_GRACE_MS | 15_000 | How long a serve the admitter asked to rest (plugin.rest) has to answer plugin.dormant and exit, or plugin.busy, before its keeper ends it (discussion:owner/790 item 3), in milliseconds. Only a serve whose manifest opts in (rest = true) is asked. Not configurable. |
PLUGIN_BUSY_HOLD_MS | 30_000 | How long a serve that answered plugin.busy is left out of the admitter's rest asks (discussion:owner/790 item 3), in milliseconds. Not configurable. |
PLUGIN_ADMIT_WAIT_MS | 30_000 | How long a plugin command waits for room in the plugins' share before plugin.admit answers that it still waits, keeping its place for its process to ask again; and how long that place is kept without the ask (discussion:owner/790 item 3, ruling lnhoqmrf (a)), in milliseconds. Not configurable. |
PLUGIN_COMMANDS_MAX | 64 | The most plugin command processes the admitter holds room for at once (discussion:owner/790 item 3); past it a plugin.admit waits as for bytes. Not configurable. |
PLUGIN_STDERR_LINE_MAX_BYTES | 64 << 10 | The longest line of a plugin's stderr the daemon holds (discussion:owner/559): each is copied to the daemon's own stderr cut to this, the rest dropped up to its newline, so a plugin that never writes one grows nothing. Not configurable. |
CHECK_OUTPUT_TAIL_BYTES | 2_048 | The most bytes of the readable text an action_failed notification carries as its content (the command, its exit status, its failing tests, the lines that look like errors, and the end of the output, with stack frames left out), read only through notifications.read (discussion:owner/193 and discussion:owner/242); within INBOX_ITEM_MAX_BYTES. |
CHANNELS_MAX | 256 | Most combinations lvcs.toml's [channels] may declare ( discussion:owner/297 and 508); one more refuses the file at parse. An abuse bound against a hostile synced lvcs.toml, sized so no real project's pipeline reaches it. Not configurable. |
GATED_ACTIONS_MAX | 256 | Most gated actions lvcs.toml's [[actions]] may declare (discussion:owner/141, discussion:owner/245, and discussion:owner/508); one more refuses the file at parse. An abuse bound against a hostile synced lvcs.toml, sized so no real project's pipeline reaches it. Not configurable. |
ACTION_SECRETS_MAX | 32 | Most secret names one gated action's secrets may list ( discussion:owner/141 and 245); one more refuses the file at parse. Not configurable. |
RUNNER_LABELS_MAX | 16 | Most labels a runner may declare (discussion:owner/648); one more refuses its serve's start. Not configurable. |
ACTION_REQUIRES_MAX | 8 | Most labels one action's requires may name (discussion:owner/648); one more refuses the file at parse. Not configurable. |
ACTION_ENV_BYTES | 256 << 10 | Most bytes of environment one step of an action may run with (discussion:owner/245): the sum, over every variable its run sets after the sandbox's fixed ones, of its name's and value's lengths plus 2, and over its GITHUB_PATH entries of each one's length plus 1. An action whose env alone passes it is refused at parse; a step that would pass it fails. Not configurable. |
STEP_FILE_BYTES | 64 << 10 | Most bytes one step file (GITHUB_ENV, GITHUB_PATH, GITHUB_OUTPUT, or GITHUB_STATE) may hold when its step ends (discussion:owner/245); a longer one fails the step. Not configurable. |
ACTION_STEPS_MAX | 64 | Most steps one action may declare (discussion:owner/248). Not configurable. |
IMAGE_MANIFEST_BYTES | 4 << 20 | Most bytes of an image's manifest, index, or config a runner fetches (discussion:owner/346); a longer one fails the pull. Not configurable. |
IMAGE_LAYERS_MAX | 128 | Most layers one image may have (discussion:owner/346); an image with more fails the pull. Not configurable. |
IMAGE_BYTES_MAX | 16 << 30 | Most bytes of an image's layers, compressed, a runner fetches, and most bytes its unpacked root may hold (discussion:owner/346); an image past either fails the pull. Not configurable. |
IMAGE_FETCH_SECS | 1800 | Longest wall clock one fetch of an image's manifest or blob may take, in seconds (discussion:owner/346). |
IMAGE_META_FETCH_SECS | 60 | Longest wall clock one fetch of an image's manifest, index, or config, or of a registry's pull token, may take, in seconds (discussion:owner/346), so a slow registry holds a pull only this long before a layer is fetched. |
IMAGE_MEMBERS_MAX | 1 << 20 | Most members one image layer's tar stream may hold (discussion:owner/346); a layer with more fails the pull. Not configurable. |
IMAGE_ENTRY_BYTES | 4 << 10 | What each member of an image's layers, and each directory its unpack creates on the way to one, costs against IMAGE_BYTES_MAX besides its data (discussion:owner/346), so entries that hold no bytes are bounded too. |
IMAGE_PATH_DEPTH_MAX | 128 | Most components one path in an image's layers may have (discussion:owner/346). |
ACTION_MATRIX_MAX | 32 | Most cells one workflow job's strategy.matrix may expand to (discussion:owner/245); one more refuses the job too_many. Not configurable. |
WORKFLOW_BYTES | 1 << 20 | Most bytes one .github/workflows/ file may hold for LVCS to map it (discussion:owner/245, stage C); a larger one is refused too_many. Not configurable. |
ACTION_PATTERNS_MAX | 256 | Most patterns one action's paths, uses, or writes may hold; one more refuses the file at parse. Not configurable. |
GENERATION_OUTPUTS_MAX | 1_024 | Most output addresses one generation record may hold, sorted and unique; a record with more is refused by the channel verifier. |
DERIVED_STATUS_MAX | 64 | Most derived paths status lists for a caller: those its trees hold that are stale or failed, or were superseded for it in the last RETRY_WINDOW_SECS, stale and failed first. |
STATUS_STUCK_MAX | 32 | Most documents status lists as stuck (discussion:owner/671 rule 6), in document address order. |
STATUS_REMOVALS_ASKED_MAX | 32 | Most protected paths absent from main's frontier that one status asks the namespace whether main removed (discussion:owner/806), in path order: a fresh joiner lacks many that main holds, and only a removal in history is listed as one, so a status stays cheap however many haven't arrived. agentflare approve --waiting asks of every one. |
ROUTER_SETTLE_MS | 100 | How long after main first moves the router computes main's id (o-decouple/8): every move within it is in that one computation, so a flood of saves hashes main at most about ten times a second however many files it holds. Main's frontier and held answers never wait for it; its id, its frontier_moved push, and its versions do. |
ROUTER_SETTLE_US_PER_PATH | 50 | How much longer, per path main holds, the router waits before computing main's id, in microseconds (o-decouple/8): hashing main costs about 5 us a path, so a wait of 50 us a path keeps it near a tenth of the router's time however large main grows, where a fixed wait made each event cost more as main grew. Never less than ROUTER_SETTLE_MS: 400 ms at 8000 paths, 5 s at 100,000. That is how long main's id, its frontier_moved push, and its versions list may lag a save; main's frontier, held, and path answers never wait for it. An incrementally kept hash of main, a cost per change rather than per path, removes the lag after 0.3.0. |
RUNNER_PRESENCE_SECS | 10 | How often a runner publishes its presence (discussion:owner/142 and discussion:owner/245). A runner publishes it on each gated action's channel it serves, and the runner actor gets a PresenceTick, this often. Seconds. Below PRESENCE_TTL_SECS (30), so a live agent's presence never lapses between two of its publishes. Not configurable. |
STEP_STOP_GRACE_SECS | 60 | How long a sandboxed step's processes have after SIGTERM, at its wall clock or when its run is cancelled, to clean up what they started outside the sandbox (an e2e script's trap sweeping its containers) before its namespace is killed (orchestrator/59: a run cut off by SIGKILL left its containers running for 50 minutes). Seconds. A step so ends by its wall clock plus this at the latest. Not configurable. |
RUNNER_ONLINE_MS | 2 * 1_000 * RUNNER_PRESENCE_SECS as i64 | How recently a runner's presence on a channel must have arrived for status and channel.list to name it online (discussion:owner/321): two of its publishes. Not configurable. |
RUNNER_SEEN_KEEP_MS | 10 * 1_000 * RUNNER_PRESENCE_SECS as i64 | How long the runs actor keeps a runner's last presence on a channel (discussion:owner/321): ten of its publishes; older entries are dropped. Not configurable. |
RUNNER_SEEN_PER_CHANNEL_MAX | 64 | Most runners the runs actor keeps a last presence of per channel (discussion:owner/321); past it, the oldest is dropped. Not configurable. |
REACH_NODES_MAX | 64 | Most known nodes the reach actor keeps per project (discussion:owner/326); past it, the one seen longest ago is dropped, an offline one first. Not configurable. |
STATUS_REVOCATIONS_MAX | 16 | Most revocations status lists as not held by every known node (discussion:owner/326), newest first. Not configurable. |
GATED_TIMEOUT_SECS | 3600 | Wall clock the run executor gives one gated action's run before it ends it, recorded as a failed outcome (discussion:owner/142 and discussion:owner/664). Seconds. Not configurable. |
REQUEST_SKEW_MS | 30_000 | How far past the clock a new run request may be dated (discussion:owner/141). run dates a run request after the greatest time its channel's value ranks by (its run requests and its passed trigger runs, discussion:owner/269), and refuses, to retry once the clock nears it, when that is later than this past now, so an honest daemon never signs a far-future change. 30 s: the clock skew between NTP-synchronized nodes. Milliseconds, the unit of time_ms. Not configurable. |
INBOX_MAX_ITEMS | 1_024 | Most pending inbox items per peer, and most delivered ids it remembers per peer, oldest forgotten first (discussion:owner/118). Past it, the newest item of the sender with the most pending items is dropped. |
INBOX_MAX_BYTES | 1_048_576 | Most pending inbox text bytes per peer (discussion:owner/118). Past it, the newest item of the sender with the most pending items is dropped. |
INBOX_MAX_PER_SENDER | 64 | Most pending inbox items per peer from one sender; past it, that sender's newest item is dropped (discussion:owner/118). |
INBOX_ITEM_MAX_BYTES | 4_096 | Bytes of an inbox item's text, cut at a character boundary with truncated set (discussion:owner/118). |
INBOX_SKEW_MS | 30_000 | How far past the shell's now_ms an inbox item's signed ts may be; an item whose ts is later than that, or more than BULLETIN_SECS before now_ms, is dropped at Item (discussion:owner/118). Milliseconds, the unit of ts. |
RETRY_WINDOW_SECS | 600 | How long the daemon keeps an answer under its request id ( discussion:owner/163). The same id sent again within it gets the first answer, and the request is not applied twice. Seconds. Not configurable. |
RETRY_ENTRIES_MAX | 1_024 | Most answers the retry window keeps per peer (discussion:owner/163); past it, the oldest is forgotten first. Not configurable. |
REQUEST_ID_MAX_BYTES | 128 | Most bytes of a request's request id (discussion:owner/163): a longer one is refused invalid_params. Not configurable. |
INTERNAL_RETRY_MS | 1_000 | The retry_after_ms of an internal_error from a transient fault (discussion:owner/163). Such a fault is a part of the daemon not yet running or a document that failed to load; the caller sends the same request again after this long. Milliseconds. Not configurable. |
PURGE_FILE_MIN_VERSION | 0.2.9 | The first lvcs release whose caps judge takes a deleted file's purge, keyed by its bare file id (discussion:owner/614). file.purge refuses to purge a file while a connected node runs an older or unknown release: an older judge refuses the change and stops the owner's caps writes for good. Not configurable. |
PURGE_LEDGER_MIN_VERSION | 0.2.10 | The first lvcs release whose caps judge takes a plugin ledger's purge, keyed by its .lvcs/x/... address (discussion:owner/614), as [PURGE_FILE_MIN_VERSION] gates a file's. Not configurable. |
ADDRESS_MAX_BYTES | 4_096 | Most bytes of a document address (discussion:owner/183), Linux's PATH_MAX. A longer one is refused invalid_params. Not configurable. |
KAD_RECORD_TTL_SECS | 3_600 | How long a Kademlia record lives after it is put on the project's Kademlia protocol (discussion:owner/175 and discussion:owner/187). Seconds. Not configurable. |
KAD_RECORD_MAX_BYTES | 64 << 10 | Most bytes of one Kademlia record's value on the project's Kademlia protocol (discussion:owner/175 and discussion:owner/187). Not configurable. |
KAD_RECORDS_MAX | 64 | Most Kademlia records, provider keys, and providers per key a node's store holds (discussion:owner/565). Kademlia only routes for LVCS, so no record another node puts is stored at all; this bounds the store whatever happens. Not configurable. |
GOSSIP_MAX_BYTES | 64 << 10 | The largest gossipsub message a node sends or takes (discussion:owner/565), stated rather than left to libp2p's default: every gossip topic carries small signed notices. Not configurable. |
GOSSIP_QUEUE_MS | GOSSIP_FRESH_MS.unsigned_abs() | Milliseconds gossipsub keeps a published or forwarded message queued for a node (discussion:owner/653): the queue drains on admission, and this is only the failure bound, GOSSIP_FRESH_MS, past which the node ignores it as stale. Not configurable. |
GOSSIP_QUEUE_MAX | 5_000 | Most messages gossipsub queues for one connection (discussion:owner/565), libp2p's default stated: past it a message is dropped and the node counted slow. Not configurable. |
AUTO_UPDATE_DEFAULT | true | Whether auto-update is on when agentflare update --auto was never run (the updater row). Off is notify-only: the daemon still checks, and records a release as available without applying it. |
UPDATE_CHECK_INTERVAL_MS | 86_400_000 | How often the daemon's updater checks for a release, after the check it makes once the daemon is ready: daily. Milliseconds. |
UPDATE_JITTER_MS | 3_600_000 | The most jitter added to each daily check, so daemons don't all ask at once: one hour. Milliseconds; the shell draws it. |
PLUGIN_NOTES_MAX | 16 | Most lines of one plugin's notes status lists (discussion:owner/672): the rest of its notes file ([PLUGIN_NOTES_FILE]) is not read. |
PLUGIN_NOTES_FILE | notes | The file, under a project's .lvcs/plugins/<plugin>/, where a plugin keeps the lines status names it with (discussion:owner/672), one per line; it removes the file when it has none. |
PLUGIN_NOTE_MAX_BYTES | 240 | Most bytes of one plugin note status lists; a longer line is cut at a character boundary. |
PLUGIN_COUNTS_MAX | 2 | Most counts one plugin declares for a project card ([[counts]], discussion:owner/684); a manifest declaring more is refused. |
PLUGIN_COUNT_LABEL_MAX_BYTES | 24 | Most bytes of one card count's label ([[counts]] label, discussion:owner/684). |
PLUGIN_COUNT_KEY_MAX_BYTES | 32 | Most bytes of one card count's key ([[counts]] key, discussion:owner/684): lowercase ASCII letters, digits, -, and _. |
PLUGIN_COUNTS_FILE | counts | The file, under a project's .lvcs/plugins/<plugin>/, where a plugin keeps its card counts (discussion:owner/684): one <key> <n> line per declared key, rewritten whole. |
CARD_COUNTS_MAX | 3 | Most counts a project card draws after Online (discussion:owner/684). |
PLAN_STEPS_MAX | 32 | Most steps in one issue's plan (discussion:owner/708): the issues fold ignores a planned event past it, and agentflare issue plan refuses it. |
PLAN_STEP_TEXT_MAX_BYTES | 200 | Most bytes of one plan step's text (discussion:owner/708): the issues fold ignores a planned event holding a longer one, and agentflare issue plan refuses it. |
UPDATE_SEEN_CHECK_MIN_MS | 3_600_000 | The least time between two checks the updater makes because a node identified with a newer release than this one (short-ids-674/13), each newer version seen checked once: one hour. Milliseconds. |
UPDATE_QUIESCE_RETRY_MS | 60_000 | How long the updater waits to try again for every sandbox slot while a sandbox runs, before it replaces the binary. Milliseconds. |
UPDATE_QUIESCE_TRIES | 60 | How many times the updater tries for every sandbox slot before it gives up until the next check. |
UPDATE_MANIFEST_MAX_BYTES | 65_536 | The largest manifest or signature the updater reads, in bytes. |
UPDATE_ARCHIVE_MAX_BYTES | 268_435_456 | The largest release archive the updater downloads, in bytes. |
UPDATE_FETCH_TIMEOUT_SECS | 600 | The longest one updater fetch may take, in seconds; connecting gets UPDATE_CONNECT_TIMEOUT_SECS of it. |
UPDATE_CONNECT_TIMEOUT_SECS | 30 | The longest the updater waits to connect, in seconds. |
UPDATE_EDGE_CHECK_INTERVAL_MS | 300_000 | How often a node on the edge channel checks for a build: five minutes. Milliseconds. |
UPDATE_EDGE_JITTER_MS | 60_000 | The most jitter added to each edge check: one minute. Milliseconds; the shell draws it. |
UPDATE_EDGE_EXPIRY_SECS | 1_209_600 | How long an edge manifest counts after its build: lvcs release-sign --build sets expires to the build's seconds plus this, and a node refuses a manifest past it, so a host can't hold the fleet on an old build. Seconds: 14 days. |
UPDATE_EDGE_STAGGER_MS | 10_000 | How far apart edge slots apply one build: a node of slot s applies build b once the clock reaches b + s × this, and checks s × this after the edge channel moves, so the fleet never restarts at once. Milliseconds. |
UPDATE_TRIAL_MS | 600_000 | How long a staged binary has, from its first trial start, to become healthy and promote itself before its supervisor rolls it back. Milliseconds. |
UPDATE_TRIAL_STARTS | 3 | How many starts a staged binary gets before its supervisor rolls it back. |
UPDATE_TRIAL_POLL_MS | 1_000 | How often a trial's supervisor reads the update status for its child's promotion. Milliseconds. |
UPDATE_TRIAL_KILL_MS | 15_000 | How long a trial's supervisor waits after SIGTERM before it kills its child. Milliseconds. |
UPDATE_UNDO_MAX | 4_096 | The most entries an update trial's undo record holds (discussion:owner/768): each recorded journal and rotated segment, and each file the trial created. Past it, recording a created file fails, and so does the trial. |
PAIRING_TTL_SECS | 600 | How long an agentflare auth login waits for a device to claim and be confirmed before it expires, in seconds (discussion:owner/300). |
REMOTE_SIGN_TIMEOUT_MS | 60_000 | How long a write device has to answer a push sign with auth.sign before the record it was asked to sign is dropped unsigned, in milliseconds (discussion:owner/300): long enough for a browser tab in the background to wake, short of a request's own timeout. |
APPROVE_HELD_WAIT_MS | 60_000 | How long approve waits for this node to hold the heads its frontier names for a path before it refuses that path as not held (session-8faca4ed/32, discussion:owner/653 rule 2): each path's document answers once the heads sync, so an approval sent before its change arrived waits for it rather than asking to be sent again. Milliseconds, added to the request's own deadline. Not configurable. |
REQUEST_TIMEOUT_MS | 120_000 | How long the daemon waits to answer a request before it answers it with a fault the client may retry (owner/15), in milliseconds; ops that wait by design add their own wait to it. |
QUERY_MAX_BYTES | 1_048_576 | Most bytes of one frame on a QUERY_PROTOCOL stream, a query or its answer; a longer one ends the stream (discussion:owner/422). |
QUERY_INBOUND_PER_NODE | 4 | Most queries from one node a daemon holds unanswered at once; a stream past it is closed unread (discussion:owner/422). |
QUERY_TIMEOUT_MS | 10_000 | How long a query may take, from opening its stream to reading its answer, and how long a daemon holds one unanswered, in milliseconds (discussion:owner/422). |
DEVICE_NAME_MAX_BYTES | 24 | Most bytes of the device name auth.pair gives, which becomes its alias's stem (discussion:owner/300). |
PAIRINGS_MAX | 16 | Pairings open at once in one project (discussion:owner/300). |
OTEL_FILE_MAX_BYTES | 64 * 1024 * 1024 | Bytes one OTLP JSON lines file under OTEL_DIR_ENV grows to before it is rotated (discussion:owner/435). |
OTEL_FILE_KEEP | 3 | Rotated OTLP JSON lines files kept per signal beside the current one (discussion:owner/435). |
WATCHDOG_STALL_SECS | 120 | Seconds the daemon's watchdog (lvcs_runtime::watchdog, e2e-suites/3) lets pass with work pending and no machine making progress, or with one machine in a single step, before it writes a stall sample to the state directory and exits with [WATCHDOG_EXIT_CODE] so the service manager starts it again. Equal to [REQUEST_TIMEOUT_MS]: the longest anything waits for a machine's answer, so a daemon stalled longer has already failed every caller. The longest legitimate single step today is a document importing one batch of sync messages ([DOC_STEP_BATCH_MAX] of at most [WIRE_MAX_BYTES]), seconds on the maintainers' Mac; project opening, Loro loads, compaction, and the journal's fsync run off the machines' steps (the blocking pool, the writer's and syncer's own threads), so they never count. [WATCHDOG_STALL_ENV] overrides it. |
WATCHDOG_STALL_ENV | LVCS_WATCHDOG_STALL_SECS | The environment variable that overrides [WATCHDOG_STALL_SECS] for one daemon, read once at its start: whole seconds, 0 turns the watchdog off. Any other value is ignored with a warning. |
WATCHDOG_TICK_MS | 1_000 | How often the watchdog's thread samples the runtime, in milliseconds. |
WATCHDOG_EXIT_CODE | 70 | The exit code of a daemon its watchdog stopped (e2e-suites/3): 70, sysexits' EX_SOFTWARE, apart from every code lvcs exits with otherwise (1 for a failure, 2 for usage, 127 for a missing binary). LVCS has no exit-code vocabulary yet, so it sits beside the watchdog's other limits. |
WATCHDOG_SAMPLE_PREFIX | watchdog- | The stall sample's file name in the daemon's state directory, before its Unix time in seconds and .txt: watchdog-<secs>.txt. |
WATCHDOG_E2E_STALL_FILE | e2e-watchdog-stall | e2e builds only (the e2e-clock feature): a file of this name in the daemon's state directory arms the watchdog's stall hook. The watchdog reads it within a tick, removes it, takes the whole seconds it holds as its stall threshold, and from then on every machine step blocks forever, so the daemon stalls as the 2026-11-07 Mac daemon did. A release build has no hook and never reads the file. |
WATCHDOG_SAMPLE_SCHEDULERS_MAX | 256 | Most in-flight schedulers one stall sample lists, longest first. |