CLI

outl ships one binary — outl — that does everything a workspace needs from outside the TUI: scripts, cron jobs, CI, editor integrations, and LLM agents. This document is the surface contract.

Looking for how to plug outl into Claude Desktop / Cursor / other MCP hosts? → docs/mcp.md. The two share the same handlers under the hood, so this page stays the source of truth for what each command does.

The bet — one binary, one surface

Knowledge bases get integrated everywhere: editors, shell scripts, LLM agents, automation. The wrong move is to grow a separate API per host (REST for one, MCP for another, library for a third). Each new client doubles the surface and drifts.

outl’s bet: everything reachable from outside the TUI is reachable through the outl binary, with a stable JSON envelope. Other protocols (MCP today, anything that comes next) are thin shims that shell out to the same commands. There is one place where logic lives, and that place is outl-actions.

The stack

┌────────────────────────────────────────────────────────────┐
│ Hosts                                                       │
│   shell · cron · CI · editors · Claude Code · Claude Desktop│
└────────────────────────────────────────────────────────────┘
                  │                            │
                  │ subprocess                 │ MCP / stdio
                  ▼                            ▼
┌──────────────────────────────┐  ┌──────────────────────────┐
│ outl <subcommand>            │  │ outl mcp serve           │
│  page · block · daily ·      │  │ (thin shim, declares     │
│  search · query · export …   │  │  tools, calls into the   │
│                              │  │  same handlers below)    │
└──────────────────────────────┘  └──────────────────────────┘
                  │                            │
                  └──────────────┬─────────────┘
                                 ▼
┌────────────────────────────────────────────────────────────┐
│ outl-actions                                                          │
│   block · tree · todo · journal · page · backlinks · sync · template   │
└────────────────────────────────────────────────────────────┘

The MCP server is a subcommand of the same binary. There is no outl-mcp crate, no separate distribution, no parallel logic. A new feature lands once: as a function in outl-actions, exposed by one subcommand and one tool, sharing the same handler.

JSON envelope

Every command that produces machine output emits the same shape so downstream consumers (jq, LLMs, scripts) cache one parser.

Success:

{
  "ok": true,
  "data": { "...": "command-specific payload" },
  "error": null
}

Failure:

{
  "ok": false,
  "data": null,
  "error": {
    "code": "PAGE_NOT_FOUND",
    "message": "page 'foo' does not exist"
  }
}

error.data is present only for a handful of codes that carry more than a message — today, PAGE_MARKDOWN_AHEAD_OF_LOG (root CLAUDE.md invariant 8, RFC 0255): a page write refused because its .md holds content the op log never recorded. error.data then carries { path, lines, sample, recovery_command } so a script or an MCP-driven agent can act on the refusal — name the page, know how many lines are at risk, and run the fix — instead of parsing error.message. Every other error code omits data entirely rather than sending null, so the envelope shape existing consumers already parse doesn’t change.

Error codes are stable strings, listed alongside each command below when relevant. Exit codes follow:

Exit 3 exists for outl sync, which can legitimately do nothing: another process on this device holds the sync endpoint (or the endpoint lease could not be claimed at all), P2P is off, or no device is paired. A script can then tell “I flushed” from “someone else will” without reading either as an error.

Add --json to any command to force JSON. Without the flag, output is human-readable (tables, colored). MCP tools return the same data, projected for an LLM consumer rather than a script. A success reply is content-only: content[].text carries the payload as compact JSON, or — for the two markdown-first tools (outl_export_md, outl_page_render) — the raw .md string. Those two flatten because their payload is {slug, md} and you supplied the slug. The journal reads (outl_daily_today, outl_daily_get) stay JSON on purpose: their outline carries the block ids, a rendered .md has none, and outl_block_update / _move / _delete / _toggle_todo all need one. There is no structuredContent on success — the text already carries the full payload as JSON, so an envelope around it is a duplicate. The MCP copy also drops fields only a GUI renderer reads (an outline node’s tokens, and default-valued collapsed / todo / empty properties); the CLI’s own --json output keeps them. That pruning applies only to outline nodes carrying an id, so outl_export_json’s parser AST comes back exactly as the parser produced it. An error reply is the deliberate exception: it sets isError: true and keeps structuredContent: { ok: false, error }. That lets a caller read error.data — e.g. PAGE_MARKDOWN_AHEAD_OF_LOG’s path / lines / sample / recovery_command — instead of parsing prose.

Commands by domain

The CLI column is what you type at the terminal. The MCP tool column is the name Claude Desktop (or any MCP host) sees.

Page

page get returns page meta plus the outline tree. page render returns the projected .md string (clean, no sidecar fields). page rename updates the page-slug property and renames the on-disk .md/.outl — it does not rewrite [[old_slug]] references in other pages. Affected blocks come back in affected_refs so the caller can decide whether to bulk-rewrite.

page create --content accepts a forest of [{text, children?}, ...] (or a single {text, children?} for ergonomics) so a brand-new page lands with its full outline in one op-log session instead of a chain of block append calls. Pass --content - to read the JSON from stdin. The returned content array mirrors the input and carries the freshly minted block ids, so the caller can keep referencing them in follow-ups.

page create --slugify treats the positional argument as a human name and derives the slug from it through the shared outl_md::slugify rule (lowercase, fold Latin accents, non-alphanumeric → -, collapse + trim). It is opt-in and idempotent on an already-clean slug, so the default path — and the outl_page_create MCP tool — stay literal, keeping hierarchical slugs like ai-agent/learning verbatim. The flag exists so external clients (the Raycast extension’s “New Page”) can ask the user for a name only and let the one owner of the rule generate the slug, instead of re-implementing slugify.

Block

block move is the one user-visible name for Op::Move. Cycle detection still applies: a move that would create a cycle returns { "code": "CYCLE_REJECTED" } and the op still goes into the log (see docs/crdt.md). block toggle-todo walks None → TODO → DOING → DONE → None, same as outl_actions::cycle_todo. One call is one step, so reaching DONE from an unmarked block takes three.

block append-tree writes a root block plus its recursive children in one op-log session. --tree accepts the JSON shape {"text": "...", "children": [{"text": "...", "children": [...]}]}, or --tree - to read the JSON from stdin. The response mirrors the input shape with id at every node so the caller can map back to anything they wrote. Prefer this over chained outl block append calls when authoring structured content from a script or agent.

Daily / Journal

<date> accepts ISO (2026-05-31) and natural ("April 22nd, 2026", "yesterday", "tomorrow"). Range is inclusive on both sides and emits one entry per day in the interval — days that have no materialised journal come back as { exists: false } placeholders so the caller can spot gaps.

Asset

Copies <file> into <workspace>/assets/<hash>.<ext> (content-addressed, so re-importing identical bytes is idempotent) and appends its markdown link as a new block. The link is [name](assets/<hash>.<ext>) for a document, ![name](…) for an image. Target defaults to today’s journal (--daily); pass --page=<slug> to append to an existing page instead — --page and --daily are mutually exclusive (INVALID_ARG), and an unknown slug returns PAGE_NOT_FOUND. A file over [assets] max_bytes in config.toml (default 100 MiB, 0 = unbounded) is rejected before the copy with INVALID_ARG. The MCP tool takes the file as a path argument (MCP is stdio — it receives a filesystem path, not bytes) plus optional page / daily. Only the link enters the op log; the asset’s bytes are a plain blob replicated alongside the .md projections, never through the CRDT.

Search / Query

search is full-text and lives today as the TUI’s workspace search. query is the structured filter (tag, property, date range, kind).

search and backlinks answer from the op log, not by reading pages/ and journals/. Practical consequence: a line that exists in a .md but in no op is not found. That happens when a file was edited outside outl and no reconcile has run yet, or when the page is in the state outl doctor reports as ahead of the log — outl reconcile --ahead-of-log is what turns that content into ops. In exchange, a block the log holds stays findable even while its sidecar is stale, which the previous disk-walking implementation could not promise: it dropped every block after the first one whose hash disagreed. The --raw='…' flag is reserved for the not-yet-implemented query DSL and currently rejects with INVALID_ARG — when the DSL lands it folds into the same outl_query tool, not a new one.

Backlinks / Refs

block embed resolves !((blk-XXX)) recursively, returning the source block plus children — the same expansion the TUI does inline.

Tags / Properties

Properties stay in the key:: value lines at the top of the page; the CLI never invents a new place to put metadata (see docs/markdown-format.md).

Export

export hugo is the pipeline that drives avelino.run: frontmatter from page properties, block refs flattened, code blocks preserved. export md is the same string page render returns. export json is the full AST plus sidecar — the format an external tool would ingest.

Template

template list finds every page with a non-empty template:: property. template apply deep-copies the template’s subtree under a target block with variable substitution. template resolve returns a callable template’s code block language, source, and declared params. template run executes a callable template — injects --params, runs its code block through the shared runtime, and writes the > **result:** subtree under --block. The MCP tool takes params as an object ({ "k": "v" }). --block (and apply’s optional --block) must belong to --page; a cross-page block returns INVALID_ARG (reprojection only touches --page, so a foreign anchor would silently drop the new blocks from disk). See Templates for the full guide.

Batch

batch runs a list of write ops sequentially in one workspace session. Input shape:

{
  "ops": [
    { "op": "page_create",       "args": { "slug": "ideas" } },
    { "op": "block_append_tree", "args": { "page": "ideas",
                                           "tree": { "text": "root",
                                                     "children": [{ "text": "child" }] } } },
    { "op": "page_prop_set",     "args": { "page": "ideas", "key": "icon", "value": "💡" } }
  ]
}

Supported op names: page_create, page_update, page_delete, page_rename, block_append, block_append_tree, block_insert, block_update, block_move, block_delete, block_toggle_todo, daily_append, page_prop_set. Each op’s args mirror the matching standalone tool.

Semantics: stop-on-first-error. When an op fails, earlier ops stay in the op log (they’re already CRDT ops; we don’t roll them back) and the response carries failed_at, failed_op, and error so the caller can decide what to do with the suffix that never ran. CLI exit code is 1 in that case; MCP returns the payload via the normal envelope.

Workspace / Admin

init, serve, reconcile, recover, compact, import, mcp serve, peer, plugin, and sync are CLI-only on purpose — they’re either interactive, long-running, or bootstrap commands that don’t fit a tool-call shape.

Importing a graph has its own page: Importing a graph — sources and dialect translations, what --force does and why re-importing is opt-in, the resumable-without---force rule, and the reconciliation table that accounts for every block the source held.

outl plugin manages the workspace’s JS plugins (under <workspace>/.outl/plugins/), wrapping outl-plugins. init <NAME> [--id <ID>] [--dir <PATH>] scaffolds a buildable starter project (manifest + package.json + tsconfig + src/index.ts + README); run bun install && bun run build inside it for an installable bundle. search [QUERY] lists installable example plugins from the repo (filtered by QUERY when given). list loads every installed plugin and prints each one’s version, enabled state, and the slash commands it contributes. install <SOURCE> takes a local directory holding a plugin.json plus its bundle (the installed shape), or a github:owner/repo[/subdir][#tag] source. A github: source is cloned at an immutable semver tag — the newest published tag when none is pinned, never a mutable branch. It prints the permissions the manifest requests and asks for approval before copying the plugin in and freezing those permissions in the lockfile. Pass --yes to approve non-interactively (required when stdin is not a TTY). run <PLUGIN_ID> <COMMAND_ID> runs a contributed command and re-renders every page’s .md afterwards, because the op log is the source of truth and the files are a projection. config show <ID> / config set <ID> <KEY> <VALUE> read and write the plugin’s plaintext config in the lockfile (value coerced to the field’s schema type). secret set <ID> <KEY> [--value <V>] / secret remove <ID> <KEY> (alias rm) manage the plugin’s secrets in the OS keychain — never in the workspace on disk. secret set prompts for the value on a hidden line when --value is omitted (prefer that: a --value on the command line lands in your shell history). enable <ID> / disable <ID> flip the plugin’s enabled flag in the lockfile without uninstalling it. remove <ID> (aliases uninstall, rm) deletes the plugin’s directory and its lockfile entry. remove <ID> (aliases uninstall, rm) deletes the plugin’s directory and its lockfile entry.

outl peer pair takes an optional --name <NAME> — the label this device advertises to the other (shown in the peer’s outl peer list). It defaults to the machine hostname; the GUI clients default it to “desktop” / “mobile” and let the user edit it before pairing.

--ticket - reads the ticket from stdin instead of argv, for the 400-to-750-character string nobody wants to paste into a shell — pbpaste | outl peer pair --ticket -, or the same piped into docker compose run -i.

The QR, and why it sometimes isn’t printed. A ticket QR runs 77 to 101 columns, because a ticket carries one entry per direct address the endpoint found and a laptop with Tailscale and Docker bridges has plenty. Past about two direct addresses it no longer fits an 80-column terminal, and one wider than the terminal wraps. A wrapped QR is not a degraded QR — no camera decodes it, and nothing on screen says why — so outl peer pair measures the terminal first and prints a single line naming the width it needed instead. outl peer qr [<ticket>|-] is how you get it back: it renders any ticket you already have and prints nothing else, so you can copy the ticket out of a narrow SSH window and render it in a wide one. Unlike pair, it always prints (the width warning goes to stderr), because a command asked for a QR and nothing else has no useful way to refuse. Error correction is L rather than the crate’s M default, which is ~8 columns narrower — M’s redundancy is aimed at print, and this code is on a screen being photographed from close range. Flattening the ticket’s double base64 encoding looks like the bigger win and is not: it saves ~12 columns, still leaves the worst case over 80, and needs a ticket-format bump that would stop a phone on one release pairing with a laptop on another. The reasoning is on outl_sync_iroh::ticket_qr. The mobile app pairs by camera, so for a phone the QR is the only practical route in.

outl peer remove <id> unpairs a device on this machine only — your other devices keep their own peer lists. For a lost or stolen device use outl peer revoke-all, which rotates the workspace identity so nothing that is not re-paired can sync again. It prompts for confirmation (--yes skips it). Scope, caveats and why rotation rather than a broadcast: docs/sync.md → Managing peers.

outl sync forces a one-shot P2P sync pass (bring the iroh transport up, exchange ops with every paired device, exit). It’s for scripts that mutate via the CLI and must flush to peers before the process dies — a normal short-lived CLI mutation can’t keep a connection alive long enough. The long-lived surfaces (outl serve, outl mcp serve, the desktop/TUI apps) sync continuously and don’t need it. If one of them is running on this machine, outl sync says so and exits without doing anything: a device binds one sync endpoint at a time, and that process is already pushing your ops out.

outl serve — the background daemon

Two halves, both on by default:

• the file watcher reconciles external .md edits into the op log (--no-watch turns it off);
• the sync supervisor holds this device’s iroh endpoint so paired peers converge continuously (--no-sync turns it off).

--once reconciles every .md and exits; it implies neither half and conflicts with both flags. Turning both halves off is a usage error rather than a process that runs and does nothing.

A third job, and it runs in the other direction. The watcher carries .md → tree; nothing carried tree → .md in bulk, so when ops arrived by sync a page’s .md stayed wrong until somebody opened it — 704 stale pages on a real 2,574-page workspace.

So serve sweeps: after its initial scan (which makes --once the batch projection pass for free) and after each peer-ops reload, with a 30s floor. It writes only pages whose re-projection removes nothing from disk; a page that would remove content lines is withheld for outl doctor --repair, which backs every file up and applies volume ceilings. That gate also makes a torn op log safe without the sweep knowing anything about op-log health — a truncated replay renders less than disk holds, so every page it touches is withheld. Measured: 710 pages re-projected in 30s, outl doctor from 745 warnings to 9, zero content lines lost across 2,848 files. It reports the change, not the state — named on first sight, again when the set differs, once when it empties. Full reasoning: RFC 0260.

--no-watch does not sweep: its contract is holding the endpoint and nothing else, and it exists to sit beside a GUI that already projects the page it opens.

The sync half defers. One endpoint per device identity, elected not assigned — so the supervisor asks for the lease every 30s and treats a refusal as a normal state. A desktop or TUI that already holds the endpoint keeps it, and the daemon takes over when that process exits. It does not hand the endpoint back: once the daemon holds it, a GUI started later loses and stays degraded (sync indicator never green, Refresh cannot force a pass) until the daemon stops. Deferring on the way in is not the same as yielding once held, and only the first is implemented. It also stands down entirely when no devices are paired, since holding the endpoint to sync with nobody only denies it to a GUI that could be using it to pair. It re-reads .outl/peers.json and rebuilds the transport when that file changes, so a device paired after the daemon started is actually synced with.

Which flag to run permanently. The watcher emits ops, so it takes the exclusive per-actor write lock, and any process that loses the race for the device actor gets a fresh ephemeral actor and its own ops-<ulid>.jsonl. That is cheap for an occasional overlap and expensive for a daemon: outl serve running forever holds the device actor, so every later GUI or TUI launch mints one more op-log file.

The sync half has no such cost — the transport writes peer ops to ops-<peer>.jsonl and uses your actor id only for the vector clock it offers. So:

• outl serve --no-watch — the mode to leave running beside a GUI you also use. No write lock, no ephemeral actors.
• outl serve — both halves, for a headless box where nothing else opens the workspace.

That headless box is packaged: the repo ships a Dockerfile and a docker-compose.yml that run exactly this. See Self-hosting an always-on peer — the pairing direction and the two persistent volumes are the parts that are not guessable.

--no-watch exits non-zero when [sync] transport is "file", because holding the P2P endpoint is then the only job it had and there is none. Exiting 0 there would read as success to a process manager, which would restart it into the same config forever. Plain outl serve only warns: the watcher still has work.

What it converges is the op log, not the .md files. The daemon never re-projects: peer ops land in ops/, the materialised tree is reloaded in memory, and the .md on that machine keeps whatever text it had. They catch up the next time a client opens the workspace, or when outl doctor --repair runs. Not on outl reconcile: with no flags that command lists the pages holding an unreconciled external edit and writes nothing, and its two writing modes (--ahead-of-log, --allow-bulk-delete) both go .md → op log, which is the opposite direction. doctor --repair is the one that projects, writing every page whose .md is absent or drifted under the same invariant-8 guard. So a headless box stays a correct replica; it is not a place to read current notes off disk. Re-projecting from a daemon has to clear invariant 8 first — overwriting a .md that holds content the log never saw is the one failure this project treats as unrecoverable — so it is deliberately not done here.

SIGTERM and SIGINT both shut down cleanly, releasing the endpoint lease. That release matters: a lease left held by a killed process locks every outl process on the device out of an endpoint.

outl doctor

The integrity check you run before trusting a migration, and after any sync weirdness. Read-only by default — it reports, it never fixes, unless you pass --repair. Exit code is 1 when the report carries any error, so it drops straight into a script or CI step.

Full check list, what --repair is allowed to touch, and the ceilings that make it stand down: doctor.md.

Snapshots. --repair reclaims the boot cache the snapshot GC judges droppable, not only the undecodable ones — outl_core::snapshot::gc owns that verdict and doctor only phrases it. A superseded snapshot (it decodes, but another outranks it) is reported as info, not a warning: nothing is wrong, there is just disk to reclaim. An unusable one is a warning. One that could not be judged — unreadable, or written by a newer build — is reported and never deleted: a file we could not read is not a file we read and proved bad. Measured: 54MB in 4 files down to 13MB in 1. --repair also collects abandoned snapshot scratch files — snap-*.bin.tmp.<ulid>, and the shared snap-*.bin.tmp older builds left behind (kind prune_snapshot_tmp, not backed up — a half-published write never became a record). Reasoning: RFC 0258.

outl reconcile

The .md → op log direction, in three modes.

outl reconcile [<path>] [--ahead-of-log] [--allow-bulk-delete]

• No flags — a read-only listing of .outl/orphans.log, the blocks a past reconcile could not match to anything in the log. Writes nothing.
• --ahead-of-log — reconcile the pages whose .md holds content that exists in no op, bypassing the sidecar hash gate (it clears last_synced_hash so reconcile_md stops short-circuiting). Such a page is hash-faithful, so the ordinary reconcile reads it as in-sync and never looks at it. Opt-in, because it writes ops for content the log has never seen — a deliberate write, not a repair.
• --allow-bulk-delete — apply a deletion the orphan-volume guard refused (more than 500 blocks of a page, or more than 75% of one). It reaches both modes: the plain pass, which is where a refused page actually lives, and --ahead-of-log. Check what the .md holds before reaching for it — the guard fires on exactly the case where the file is the thing that is wrong.

Pages it could not judge are counted and named. There are three states, not two: a page can be reconciled, read and found clean, or never read at all. An .md that exists and will not open, a sidecar that will not parse, and a sidecar that records no block text are all the third. Each used to be dropped without a line of output, so a workspace where every page was skipped printed no page holds content outside the op log and exited 0 — a refusal that never reached the user, on the command you run precisely because a silent refusal already cost you content (invariant 8). Both writing modes now end with the list:

2 page(s) could NOT be judged and were not reconciled:
  alpha — its `.md` could not be read
  gamma — its sidecar will not parse
None of these is clean — each was skipped before its content was read.

Two things are deliberately not on that list, because they are answerable rather than unreadable. A page with no .md at all: there is no file, so nothing on disk can be outside the log, and that is the normal state of every page on a freshly paired device. And an empty page whose sidecar records no text — outl init leaves two of those behind, and every page created for a [[link]] that was never filled is the same shape. A skip list that fires on the happy path is one users learn to skim.

outl recover

Recovers block text an Op::Edit truncated — the mirror of outl reconcile --ahead-of-log, reading a different source. Both close the same gap (RFC 0210), and neither replaces the other:

• reconcile --ahead-of-log reads the .md. It recovers content still on disk but in no op. It cannot help a page whose .md was already overwritten.
• recover reads the op log. The producer bug did not lose text into thin air: the reconcile that followed it emitted an Op::Edit carrying the truncated text, and the edit before that one — carrying the full text — is still in the append-only log. So a page whose .md was already overwritten before the guard existed is unreachable by the first and recoverable by the second.

Read-only by default. A plain outl recover lists what it found and writes nothing. --apply is the only writing mode, and what it writes is a new Op::Edit per block — the log is never rewritten (root CLAUDE.md invariant 1).

outl recover [<path>] [--apply] [--min-lines=N]

• --min-lines (default 1) — only report a block that lost at least this many non-blank lines. A one-line loss is often ordinary editing, so the default is noisy on purpose: a false positive costs a line of output, a false negative costs the content. Raise it when the listing is too long to read.
• --apply — write the recovered text back. Additive by construction: a block only qualifies when its current text is a prefix of the revision being restored, so nothing the block shows today is dropped. outl_actions::restore_truncated_block re-checks that at write time rather than trusting the scan, and refuses a block that changed in between.

A --apply run re-projects the .md of every page it touched, best-effort. The common, harmless outcome there is the page’s .md already carrying the full text (only the log was truncated). The re-projection guard reports and skips that page rather than overwrite it — nothing is lost, since the .md is the party that is ahead. What it leaves behind is a stale sidecar: it still lists the truncated text, so the guard keeps refusing that page’s future re-projections, and outl serve --once does not clear it (measured: 0 ops applied, the .md unchanged). outl doctor names each such page and points at the fix, which is outl reconcile --ahead-of-log — that clears last_synced_hash and rebuilds the sidecar.

Run reconcile --ahead-of-log first, then recover. Reconcile is the wider net (it recovers whole blocks the log never saw, not just truncated ones) and it refreshes sidecars, so running it first leaves recover with only the cases the .md genuinely cannot answer. Measured on the workspace that surfaced issue #210: reconcile-first recovered 40 pages / 623 ops and left recover reporting 4 blocks holding 4 lines, all of them one-line editing artifacts. Recover-first restored the same content as 4 blocks / 77 lines but parked two pages behind the guard until a reconcile ran anyway.

Cost is one op-log read per node in the workspace, since there is no cheaper prefilter that wouldn’t itself be a second opinion about which blocks deserve a look. Measured at ~4.7s over 67,213 nodes / 214k ops on the workspace that surfaced issue #210.

outl compact

Drops provably-inert ops from the op log (issue #110). The log is append-only, replays on every boot and ships whole to every newly paired device, so its size is the ceiling on both — a real 2,574-page workspace carried 262MB of ops/ for 28MB of markdown.

It removes one shape and only that shape: a Move that restates the placement its own Create made immediately before it. Every other op is copied through byte for byte.

outl compact [<path>] [--apply] [--no-horizon] [--force]

Read-only by default. A plain outl compact reports and writes nothing.

The predicate is six conditions, not one, and RFC 0256 is where the argument lives. The short reason: Op::Create is idempotent, so an adjacent Create + Move pair reads two ways that are indistinguishable in the file. When the Create really created the node the Move is inert; when the node already existed the Create is the inert one and the Move is doing the work — which is what a trashed-then-restored block looks like, since deletion is Move(node, TRASH_ROOT). On that workspace the naive rule matched 99.9% of Move ops and the sound predicate matches 94.7%; the 2,074 it declines are exactly the trashed-and-restored ones.

• --apply — rewrite the log. Refuses while any other outl process holds the workspace (exclusive flock on .outl/.lock plus a write lock on every actor), backs every file up to .outl/compact-backup/<timestamp>-<ulid>/ and fsyncs before touching anything, then replaces via temp + rename. Deletes the index sidecars rather than rebuilding them — compaction invalidates every byte offset, and deleting has no failure mode where a rebuild writes a wrong one. Refuses outright on an unparseable record, on a directory it cannot read, and on the PerPage layout — and refuses before any file is rewritten, because “a damaged log is never rewritten” has to mean no file, not this file. Rewrites only this device’s ops-<actor>.jsonl; every other actor file is reported and left alone. A peer’s file is that device’s own append-only log, and every file transport (iCloud, Syncthing, a shared filesystem) reconciles per path, last-write-wins. A shortened copy therefore wins the merge and takes with it every op that device had not yet shipped. Run the command on each device.
• --no-horizon — also compact history newer than the 30-day settling horizon. The horizon is not about the ops, which are inert the moment they are written; it is about peers that have not delivered theirs yet. It is measured against min(newest op, now): anchoring it on the log alone would let one peer’s wrong clock silently turn every run into --no-horizon. Drop it only on a workspace with no other device.
• --force — also rewrite other devices’ op logs. Safe only when no file transport carries this workspace (a transport = "iroh" workspace inside a Dropbox folder is exactly the trap, which is why the config is not what decides). Also the only route to an ops-<ephemeral>.jsonl this device wrote in a past session: those are ours in fact, but an ephemeral actor leaves no binding behind, so they are not ours provably.

Measured: 62,209 of 217,811 ops dropped (28.6%), ops/ 262MB → 195MB, materialized tree, properties and collapsed flags byte-identical, outl doctor reporting integrity OK.

outl page history / outl block history

What the op log says happened, newest first. Read-only — neither writes an op.

$ outl page history buser-cto --limit 3
history of `buser-cto`
  2026-08-07 11:44  edited
      - **Volta do Slack**
      + **Volta do Slack** :slack:
  2026-07-16 12:24  deleted
      - 32 dias, até o abril/23
  2026-07-16 12:24  created

showing the 3 most recent of 174 events — `--limit` for more

--limit (default 50) caps the listing, never the count: the last line always names the total, so a truncated history can’t read as a complete one.

Two rules worth knowing before you read the output, both owned by outl_actions::timeline:

• A page’s history includes the blocks deleted out of it, with the text each one held when it went. That is usually why someone opens a history at all. A block moved to a different page goes with it and shows up there; outl block history follows one block wherever it has lived.
• Not everything in the log is a change. Folding a block, snoozing a reminder, a page-slug write, an Op::Edit that re-emitted a block’s existing text, and a re-emitted Create / Move that moved nothing are all skipped. A reconcile produces those in volume — on the reference workspace one block’s history was six rows of them around a single real edit.

For scripting, --json gives one flat object per event: change is created / edited / deleted / restored / moved / property, with from / to / text / key beside it, plus physical_ms and logical if you need the real HLC ordering rather than the rendered local time.

The desktop has the same read behind the ⏱ button in the page header.

MCP

Every machine-shaped command above is also exposed as an MCP tool through outl mcp serve — same binary, same handler, same JSON shape. Claude Desktop, Cursor, and any other MCP host plug straight into it.

→ docs/mcp.md covers the wiring, resources, prompts, and troubleshooting. This document stays focused on the surface; how to attach it to a host lives over there.

What does not map 1:1 (and that’s fine)

• Interactive commands (init, reconcile, recover, mcp serve) stay CLI-only. A wizard inside a tool call is the wrong shape.
• Long-running watchers (serve) stay CLI-only. MCP tools are request/response; the file watcher is a process, not a tool.
• Destructive commands (page delete, block delete) accept --confirm on the CLI and require confirm: true in the MCP input. Without it, the tool returns { "code": "CONFIRM_REQUIRED" } and the operation is a no-op.
• Importers (outl import …) stay CLI-only — they’re one-time migrations, not workspace ops.
• page history / block history — read-only, and their handlers print rather than returning a Value, so they need the ordinary fn(ctx, …) -> Result<Value, ApiError> extraction before they can be registered. Tracked with #241.

Layout

The CLI and shim are siblings inside outl-cli. Everything below delegates to outl-actions.

outl-cli/
└── src/
    ├── main.rs              # clap entry, dispatches to commands/
    ├── output.rs            # JSON envelope, --json flag, exit codes
    ├── commands/
    │   ├── page.rs
    │   ├── block.rs
    │   ├── daily.rs
    │   ├── search.rs
    │   ├── query.rs
    │   ├── tag.rs
    │   ├── export.rs
    │   ├── workspace.rs
    │   └── mcp.rs           # `outl mcp serve` shim
    └── mcp/
        ├── server.rs        # stdio transport
        ├── tools.rs         # tool registry → handlers
        ├── resources.rs     # outl:// URIs
        └── prompts.rs       # /outl-* prompts

commands/*.rs and mcp/tools.rs both reach into outl-actions. No business logic lives in either layer — they format input and output, that’s it.

Status

Shipping today:

• outl init, outl serve, outl doctor, outl reconcile, outl recover, outl import logseq|obsidian|roam, outl theme.
• outl (no subcommand) opens the TUI.
• outl page get|create|update|delete|list|rename|render (create accepts --content to seed the outline in one call)
• outl block get|append|append-tree|insert|update|move|delete|toggle-todo|tree
• outl daily today|get|append|range
• outl search "<query>" [--in=blocks|pages|all]
• outl query [--tag] [--priority] [--since=Nd] [--kind] [--prop k=v]
• outl backlinks page|block|embed
• outl tag list|pages
• outl page prop set|get|list
• outl export hugo|md|json
• outl template list|apply|resolve|run
• outl batch — stream {ops: [...]} from stdin (or --ops=…)
• outl workspace info
• outl mcp serve — full MCP protocol surface (tools, resources, prompts) over stdio.

Still ahead:

• Richer outl query --raw='…' DSL (today returns INVALID_ARG).
• Per-page block-level property surface beyond the well-known keys the prop list probe enumerates.

The order of landing matched the order of unlocking real workflows (scripts → LLM agents in Claude Code → Claude Desktop → blog publishing pipeline).