App Server

knightcode app-server hosts many conversation threads in one process. The threads share models and credentials but own separate session files, extension runtimes, tools, approvals, and turn queues. The desktop app can spawn it; scripts can use the same protocol.

The transport is newline-delimited JSON over stdin/stdout. Messages use JSON-RPC 2.0 request, response, and notification shapes without a "jsonrpc" field. There is no protocol version number: the client and the server ship together, and a client that needs a newer server checks serverInfo.version. It is separate from the single-session RPC mode. Worktrees, diffs, git operations and extension dialogs are later phases; this server does not implement them yet.

Starting

knightcode app-server [-e <extension>]... [--session-dir <dir>]

-e / --extension is repeatable. Paths resolve against the process working directory and load in each thread as explicitly trusted CLI paths. Project-local resources still follow project trust. Extension dialogs and custom UI are not yet available in desktop mode; extensions see mode: "desktop" and hasUI: false.

Session storage uses --session-dir, then KNIGHTCODE_CODING_AGENT_SESSION_DIR, then the global sessionDir setting. Without an override, sessions use the per-project layout described in Sessions. stdout carries only protocol messages; logs go to stderr. Closing stdin shuts down the server. knightcode app-server --help (or -h) prints usage on stderr and exits without opening a connection.

Handshake

Send a request with a string or numeric id:

{"id":1,"method":"initialize","params":{"clientInfo":{"name":"example","version":"1.0"},"capabilities":{"optOutNotifications":["item/reasoning/delta"]}}}

The result contains serverInfo and platform: {os, arch, homeDir}. os is win32, darwin, or linux. capabilities.optOutNotifications is an optional array of at most 256 method names, scoped to this connection. Unknown names are tolerated. It filters only notifications, never responses or server requests such as approvals.

Wait for its result, then send the notification:

{"method":"initialized","params":{}}

Only then send other requests. Requests before initialization return notInitialized. Unknown methods return methodNotFound; invalid parameters return invalidParams.

Threads and turns

Start a thread with its project directory:

{"id":2,"method":"thread/start","params":{"cwd":"/path/to/project","approvalMode":"ask"}}
{"id":3,"method":"model/list","params":{}}

Wait for each response. thread/start and thread/fork return {thread, warnings, ui} (ui is the thread's extension UI, see Dialogs and extension UI). A fork is named after its source with (fork) appended. thread/resume also returns warnings; render them from the response, because they can occur before subscription. A warning has {message, source, path?, line?, column?}, where source is settings, extension, trust, or server. Malformed settings, refused extension loads, and undecided project trust do not silently disappear. Later extension errors send a subscriber-only warning with threadId.

Threads and list summaries include createdAt (header creation time) and optional forkedFromId (the parent's header ID, not a UUID inferred from its filename). A loaded thread also carries sessionFile, the session file's path, once it has been saved. List summaries use file modification time for updatedAt; loaded threads use the latest entry time, falling back to the header. Their status is one of:

• {type: "notLoaded"}
• {type: "idle"}
• {type: "active", waitingOn: []}; the array can contain approval and dialog.
• {type: "systemError", message}; ownership loss remains visible after unload until a successful reload.

thread/status/changed broadcasts these objects, including the first opening and last closing of an approval or dialog.

Extension providers appear in model/list after a thread has loaded them. Each model has thinkingLevels, the accepted picker values; non-reasoning models report ["off"]. Pick a model using thread/settings/update, with threadId and model: {provider, id}, before starting a turn if no usable model is selected.

{"id":4,"method":"turn/start","params":{"threadId":"THREAD_ID","text":"Explain this repository","requestId":"client-start-1","clientMessageId":"optimistic-1","images":[{"path":"/absolute/path/screenshot.png"}]}}

A start returns {status, threadId, requestId}. status is started, handled (an extension consumed the input), or cancelled (cancelled before acceptance). It does not wait for the turn to finish. The optional client token requestId is echoed in the response and turn/started; it is distinct from the JSON request id. Session-replacing commands preserve it and return the effective target thread ID. The first acceptance wins even if cancellation follows later.

turn/start and turn/queue/add accept optional clientMessageId (1–128 characters). It is echoed on the first live userMessage of the turn the input starts, after input transformations; it is not persisted. A handled or removed input creates no matching item. Their images accept inline {mimeType, data} base64 images or {path} on the server's shared filesystem. Path images must be absolute, regular PNG, JPEG, GIF, or WebP files of at most 20 MiB; missing, relative, non-image, directory, and oversized inputs return invalidParams.

A plain text turn sends these notifications in order:

1. turn/started
2. item/started and item/completed for the user message
3. item/started for the agent message
4. Zero or more item/agentMessage/delta notifications
5. item/completed for the agent message
6. turn/completed

A turn is one agent run between durable start and end entries. Automatic retries, recovery, and steering may span several low-level agent runs inside it. Turn IDs come from persisted marker entries; item IDs use durable ordinals except tool calls, which retain their call IDs. Final messages decide item contents and statuses. Turn requires numeric startedAt and optionally has completedAt; items optionally carry both. Times are Unix milliseconds, unlike thread creation/modification times, which are ISO strings. Durable turn markers and end entries determine final turn times. Assistant text/reasoning uses message timestamp and duration, and tool calls finish at the tool-result timestamp. User, compaction, and extension items may have only completion time. Nested calls are live-only. An interrupted turn without an end entry has no completion time.

Started-item times and streaming contents are provisional. item/completed replaces times, text, fileChanges, and index; do not just append deltas. index is the item's top-level position in the turn: order by its latest value, because a streaming item's index can change at message end. item/removed withdraws an item dropped by a message-end handler. Nested tool calls carry parentItemId and no top-level index. A saved turn with no end entry reads as interrupted. See Session Format.

Completed edit/write calls include fileChanges: [{path, kind, diff?}]; kind is add, modify, or delete. A non-empty tool-result patch becomes diff. Extension tools can report validated file changes through result details; a malformed report is ignored as a whole. Failed calls report no changes.

turn/completed contains a summary without items, avoiding a second large transcript transfer. Failed turns have error: {message, kind}, where kind is auth, rateLimit, contextOverflow, network, provider, or other, classified from the error text. Summaries and references bound error.message to 2,000 characters; full payloads retain it. Interrupted outcomes take precedence over failures.

turn/steer hands input to the running turn and requires expectedTurnId, the currently running turn ID. A stale ID returns turnMismatch, so a steer never lands in a later turn. A steer cannot be taken back once accepted.

turn/queue/add queues input on the server and returns {queueId}. Each queued input starts a turn of its own once the thread is idle, oldest first; on an idle thread it starts at once. Until it starts it is removable with turn/queue/remove. A queued input that fails to start (no model, signed out) is dropped with an error notification carrying its clientMessageId, if it had one. turn/queue/updated sends {threadId, steering: [text], followUp: [{id, text, attachments?}]}: steers the agent has not delivered yet, and the queued inputs. text is the user's own (not a prompt composed with review comments); attachments counts the images and review comments an input carries that its text alone does not.

turn/interrupt accepts optional turnId; a supplied stale ID returns turnMismatch, while omitting it retains interrupt-current behavior. It drops queued inputs, as Escape does in the terminal; the client still has their text from turn/queue/updated to restore into its composer.

An interrupt answers at once; it does not wait for extension hooks. A start interrupted during its input hooks or an extension command returns cancelled and launches nothing, but hooks already running finish. A start interrupted later, during compaction before the run or before_agent_start, returns started, and its turn completes interrupted without a model request.

Settings changes and compaction serialize with starts; a thread already running refuses another start with turnInProgress. Settlement hooks and their deferred dispatch also retain admission until they finish. Steering input still waiting in interception belongs to that running turn: interruption, unloading, or settlement cancels it. It cannot silently enter a later turn; the waiting request returns invalidParams.

Subscriptions and replacement

thread/start, thread/resume, thread/fork, and turn/start subscribe the requesting connection. thread/unsubscribe withdraws it. Every initialized connection gets thread/sidebar and account updates, including thread/turn/completed: {turn} for completed, failed, and interrupted outcomes. Only subscribers get turn/completed, turn/item notifications, deltas, retries, queue updates, and thread/tokenUsage/updated. Approvals target one subscriber at a time.

Initialize visible sidebar rows from newest-page thread/read, without loading or subscribing to all threads or resolving referenced turn payloads. Keep live outcomes arriving during a read and apply them after its result. An absent saved last-seen turn ID establishes a baseline; live completions can still mark unread. Active/waiting status outranks the previous outcome; systemError stays separate. Desktop OS notifications are for live outcomes, not initial history.

thread/resume returns history and subscribes in the same synchronous send step. Its separate runningTurn is a current snapshot, even when an old history cursor is supplied. Its required queue and usage are also live send-time snapshots, not values from handler entry or the history cursor; so is ui. Buffer queue/usage/ui/* events while awaiting the response, install the snapshots (an empty queue clears stale state), then replay the buffered events. Invalid cursors subscribe nobody. Install runningTurn before applying buffered deltas. thread/replaced moves subscriptions to the new thread; captured old extension contexts become stale. thread/unsubscribe returns {status: "unsubscribed" | "notSubscribed"}.

TokenUsage contains cumulative input, output, cacheRead, cacheWrite, and cost, plus optional contextTokens and required contextWindow. thread/tokenUsage/updated: {threadId, turnId, usage} arrives after every finalized assistant reply, including replies before tool results and compacted-history billing. contextTokens can be absent when the current context is unknown.

History, listings, and large values

thread/read reads saved history without loading or locking the thread and never writes the source. thread/resume loads and owns it. History pages contain at most 50 turns, newest page first; turns within a page are chronological. nextCursor anchors later pages to the original branch prefix, so appends cannot shift the page. Diverged, foreign, or malformed cursors return invalidCursor.

Initial thread/list uses a per-directory .thread-index.json metadata cache. It stats transcripts and rescans only new or changed files; the first cold scan still reads their full history. Cached titles are bounded, never full messages, and fork origins are cached after resolving parent headers. Warm listings open no unchanged transcripts or parent headers. Missing, corrupt, invalid, or oversized (over 16 MiB) indexes rebuild safely; cache writes are best effort. Directories are scanned sequentially, with at most 64 transcripts open.

Continuations use an immutable summary snapshot without rescanning. Up to 32 snapshots are retained with a five-minute sliding expiry; expired, evicted, or filter-mismatched cursors return invalidCursor.

Large turns/items and tool input/output use immutable payload references rather than oversized protocol lines. Call payload/read with the reference ID and a byte offset; concatenate base64-decoded chunks until eof. nextOffset, when present, is the next byte offset, not a character count. References expire after 30 minutes without access; reread the thread to obtain new references. IDs grant access only to server-owned spool files, not arbitrary filesystem paths.

Approvals

Desktop threads default to ask; terminal-mode in-process threads default to fullAccess. Effective mode is persisted on the branch; invalid stored state restores as ask.

Built-in plan and classifier restrictions run before this gate. Approvals are not a filesystem or operating-system sandbox.

The server sends item/toolCall/requestApproval with thread/turn/item IDs, the tool name, input or an input reference, and required decisions. Optional reason explains the gate (for example, an outside-workspace write), and optional prefix is the offered shell command prefix. Render the supplied choices, not a hardcoded list. Reply to its id with a result such as {"decision":"accept"} or {"decision":"decline","message":"Use the project-local file instead"}. A decline message is optional and bounded to 4,000 characters; it reaches the agent as the tool's decline reason. cancel always blocks execution. An unoffered decision is treated as decline and never grants permission.

acceptForSession remembers a non-shell tool name for this thread's load, never across unloads. It is never offered for bash or PowerShell. Shells can instead offer acceptPrefixForSession with a positively understood prefix. For example, a git status grant covers git status --short, not git reset. Only simple, recognized command forms get a prefix; chaining, redirection, substitution, quoted/escaped command names, script dispatchers, interpreters, nested wrappers, and unrecognized subcommand forms fail closed. A later command must independently produce exactly the granted prefix, not merely start with its text. Grants are shell-specific and confined to this thread's load.

The most recent initiating subscriber is asked first. Disconnect or unsubscribe cancels that attempt and migrates it to another subscriber; no subscribers means cancel. serverRequest/cancelled withdraws an outstanding request.

Dialogs and extension UI

What KnightCode asks the user mid-turn reaches a desktop connection as a server request, asked the way approvals are; the thread reports waitingOn: ["dialog"] meanwhile.

item/toolCall/requestUserInput carries ask_user's questions, all at once, with the itemId of the call asking, so a client can move back and forth between them before answering. Each question has an id, its text and its options (label, description); a label ending in (Recommended) is the recommended one. Reply {"answers": {"<id>": {"label": "..."}}} with the chosen option's exact label and an optional note, or {"other": "..."} for the user's own answer; leave out a skipped question. {} dismisses the questions and stops the turn.

ui/dialog carries the rest, such as the classifier gate's confirmation of a risky call and plan mode's review. kind says which: a select carries options, a confirm its message, an input its placeholder, an editor its prefill. Reply with {"value": "..."} for a select, input or editor, {"confirmed": true} for a confirm, or {} to cancel.

Notices such as "Downloaded model" arrive as the notification ui/notify with a level of info, warning or error.

What an extension shows beside the transcript is the thread's extension UI:

• The ui of the thread/start, thread/resume and thread/fork results is all of it as it stands when the response is written: {statuses: [{key, text}], title?, workingMessage?, workingVisible, widgets}. Statuses are sorted by key. Each widget is {key, placement, lines} or {key, placement, view}, placement being aboveEditor or belowEditor; those above come first, each group in the order they were last set.
• ui/state carries every field but the widgets, whole, whenever one changes. An absent title or workingMessage is a cleared one. workingVisible: false asks the client to hide its waiting indicator, and workingMessage replaces its text.
• ui/widget carries one widget set or replaced (it moves to the end of its placement), or {threadId, key} alone when it is removed.

Install ui from the result, then apply the notifications, buffered across a resume as queue is. A /reload clears everything the old extensions showed, with one ui/state and a removal per widget, before the new ones start. The server strips ANSI escapes and keeps text on one line, at most 200 code points; widget lines keep their spacing. A view is data, never HTML; the server checked it before sending.

A button or field in a widget's view answers with the request ui/event (below), which reaches every handler the thread's extensions registered with ctx.ui.onViewEvent. Views on toolCall and extensionMessage items (their optional view) are for display only.

Widget factories, custom() components, header and footer components, raw terminal input and the working indicator draw only in the terminal: the server writes one line to its log per method per thread load and draws nothing. Editor access (setEditorText, getEditorText, pasteToEditor) does nothing. Threads a terminal-mode connection loads have no dialogs or extension UI yet.

Workspaces, diffs and git

A thread runs in a workspace: the user's checkout ({mode: "local"}), or a git worktree of its own ({mode: "worktree"}), as in the Codex desktop app.

thread/start {cwd, workspace: {mode: "worktree", baseRef?, carryLocalChanges?, environment?}} creates a worktree on a detached HEAD at baseRef (default: the checkout's HEAD) under <worktrees.root>/<4 hex digits>/<repository name> and runs the thread there, in the same subdirectory the cwd had. No branch is created until git/branch/create or a handoff makes one. carryLocalChanges brings the checkout's uncommitted changes (untracked files included, ignored files not) along as uncommitted edits. Creation runs no hook, filter or user code, so a failed creation leaves nothing behind. A start whose cwd is inside a managed worktree is a new thread in that worktree.

thread/workspace/progress {threadId, phase, output} streams the output of an environment's setup or cleanup script to the connection that asked. Every Thread carries workspace; ThreadSummary.workspaceMode marks worktree threads in listings.

Diffs come in Codex's Changes scopes: lastTurn, uncommitted (the default), unstaged, staged, committed and branch. Each hunk has a content-hash id; an action naming a hunk that changed since it was read answers gitConflict with data.reason: "diffChanged". Files changed while a turn ran are reported by turn/diff/updated {threadId, turnId, files} after turn/completed, each file as {path, additions, deletions}: the net line totals between the working tree before and after the turn. A read-back turn carries the same list as Turn.files. This means "changed while the turn ran", not "changed by the turn": an edit another local thread or an editor made meanwhile is in the list too. A worktree thread is the only writer in its worktree, so its list is its own.

Review comments are drafts the client keeps. They travel with the next turn/start or turn/queue/add as comments; the server quotes the commented lines from the diff the user reviewed, and the user item shows the comments next to the typed text.

Reverting an added file deletes it, an untracked one included. Clients confirm with the user before a revert that removes files.

thread/handoff moves a chat and its uncommitted changes between its worktree and the main checkout. The destination must be clean: "Stash or commit your local changes to hand off." Archiving a thread deletes its worktree after a snapshot, kept as refs/knightcode/snapshots/<bucket>/<time>, and thread/worktree/restore rebuilds it. Past worktrees.keep worktrees, the least recently used are pruned the same way. Outside a git repository, diff and git methods answer notGitRepo.

gitConflict errors carry data.reason: diffChanged, patchFailed, concurrentChange, pushRejected, localChanges, operationInProgress, branchExists, recoveryRequired, worktreeDeleted, worktreeMissing, worktreeInUse or submodules, with paths, output or recoveryInstructions where they apply.

Method reference

All methods require the handshake. Common errors are notInitialized, invalidParams, and shuttingDown; the additional errors below are method-specific. The generated schema remains the exhaustive field reference.

turn/queue/remove

Params: {threadId, queueId}. Result: {removed}. removed: true means the input will not start; false means it was unknown, already started, or already removed. This does not edit saved history. Errors: threadNotFound.

{"id":10,"method":"turn/queue/remove","params":{"threadId":"THREAD_ID","queueId":"QUEUE_ID"}}

config/read

Params: optional {cwd}. Result: {settings, defaults, origins, versions, paths, errors}. Without cwd, only global settings are read. With cwd, trusted project settings merge over global settings; untrusted project files are not read. origins maps dotted leaf keys to global or project. versions.global and optional versions.project are SHA-256 hashes of raw file contents, or "none" for an absent or unread untrusted scope. defaults is what an unset key means, in the shape of settings; paths names the global and, with cwd, the project file. Settings, origins, versions, and parse errors come from the same scope reads, not cached settings paired with fresh versions. Parse errors appear as structured source: "settings" warnings with an optional path and available one-based line/column; unaffected scopes still read. Storage read/lock failures return internal.

{"id":11,"method":"config/read","params":{"cwd":"/path/to/project"}}

config/write

Params: {scope, key, value, cwd?, expectedVersion?}. scope is global or project; key is a dotted key such as retry.provider.maxRetryDelayMs. value: null deletes the leaf. Result: {version}, after the write and reload. Project writes require cwd and an explicitly trusted project. Writes and ordinary settings setters share a per-file queue across manager instances. Under the file lock, expectedVersion, if present, compares against the current raw-content hash; existing migrations run before the edit, so deleting a canonical key does not resurrect its legacy alias. Corrupt JSON is refused without overwriting it.

Errors: configConflict with data: {version} when the file changed; reread and resolve the conflict rather than blindly retrying. invalidParams rejects untrusted project writes (data.reason: "untrusted"), missing project cwd, malformed keys, and prototype-pollution segments (__proto__, constructor, prototype). Read/parse/write failures return internal.

{"id":12,"method":"config/write","params":{"scope":"global","key":"theme","value":"dark","expectedVersion":"none"}}

tool/list

Params: optional {cwd}. Result: {tools: [{name, description, source}], defaultTools}. The tools a new thread in cwd would register (without cwd, a project-less one), from an in-memory session like command/list's; hidden tools are left out, and MCP tools appear only once a thread connects. defaultTools is the global defaultTools setting resolved (+name and -name applied), or the built-in default read, bash, edit, write. Change it with config/write.

{"id":26,"method":"tool/list","params":{}}

tool/settings/read and tool/settings/write

The KnightCode tools /tools manages, stored in tools.json in the agent directory. tool/settings/read {} returns {tools: [{name, enabled, settings}], classifiers: [{provider, id, name}]}: each tool's state (its default when the file does not set it) and its own settings, with values of keys ending in Key masked to their last four characters; classifiers are the classifier models with credentials, for the classifier gate's model. tool/settings/write {name, key, value} sets one setting; enabled takes a boolean, other keys a string, and null deletes the key. An unknown tool or a malformed key is invalidParams.

{"id":27,"method":"tool/settings/write","params":{"name":"websearch","key":"provider","value":"brave"}}

mcp/list, mcp/update, mcp/add and mcp/remove

The servers of the global mcp.json and, for an explicitly trusted project, its .knightcode/mcp.json (MCP). mcp/list {cwd?} returns {servers, errors, paths}: each server's name, scope, defining file (source), the project file that overrides it (override), transport (stdio or http), target (command line or URL), description, enabled and exposure. mcp/update {cwd?, name, enabled?, exposure?} saves the change where /mcp would: in the project override when there is one, else in the defining file. mcp/add {scope, cwd?, name, config} validates config as an mcpServers entry and adds or replaces it (result {replaced}); auth is refused in a project file. mcp/remove {scope, cwd?, name} returns {removed}. Project writes need cwd and a trusted project (invalidParams, data.reason: "untrusted"). Connections, sign-in and status belong to the MCP extension in each thread (/mcp); loaded threads pick up file changes when they next load.

{"id":28,"method":"mcp/add","params":{"scope":"global","name":"docs","config":{"url":"https://example.com/mcp"}}}

resource/list and resource/set

Extensions, skills, prompt templates and themes as knightcode config lists them. resource/list {cwd?} returns {resources: [{type, path, name, enabled, scope, origin, source}]} for what a thread in cwd (or with no project) would find, installing nothing. Built-in extensions are left out: they are always on, so resource/set refuses them like any unknown path. resource/set {cwd?, type, path, enabled} loads or skips one resource with a + or - pattern in the settings of the scope that defines it: a top-level resource in that scope's extensions, skills, prompts or themes, a package's in that package's filter. Project resources need a trusted project; an unknown path is invalidParams.

{"id":29,"method":"resource/set","params":{"type":"skills","path":"/home/u/.knightcode/agent/skills/review/SKILL.md","enabled":false}}

project/trust/read

Params: {cwd}. Result: {trusted, needsTrust}. trusted is true, false, or null (undecided); needsTrust means project-local resources require a decision. This reads the server-owned trust store, not a separate desktop-side preference.

{"id":13,"method":"project/trust/read","params":{"cwd":"/path/to/project"}}

project/trust/set

Params: {cwd, trusted} with a boolean decision. Result: {}. Writes the server-owned trust store; subsequent loads and config reads use it. Already loaded runtimes are not silently rebuilt. A cwd that is not an existing directory returns invalidParams. Filesystem failures return internal.

{"id":14,"method":"project/trust/set","params":{"cwd":"/path/to/project","trusted":true}}

file/search

Params: {cwd, query, limit?, searchId?}. query is 1–256 characters, limit is 1–200 (default 50), and searchId is at most 128 characters. Result: {files: [{path, score, indices}], truncated} with cwd-relative paths, fuzzy scores, and matched character indices. The server lists paths with fd using an argument array, never a shell. A newer request using the same searchId (including from another connection) immediately supersedes the older one, which returns {files: [], truncated: false} unless shutdown takes precedence.

Listings have a ten-second TTL, at most eight cached cwds, and at most two active listings. Output is capped at 200,000 paths or 32 MiB stdout, with truncated: true when the bounded listing is retained; a 15-second timeout rejects with internal. Partial failed output is usable but never cached; empty failed output rejects. No remaining waiters means abort; shutdown always returns shuttingDown, even if child rejection arrives after superseding. toolMissing means fd is unavailable; other empty listing failures return internal.

{"id":15,"method":"file/search","params":{"cwd":"/path/to/project","query":"src/config","limit":20,"searchId":"composer"}}

command/list

Params: {threadId} or {cwd}. Result: {commands: [{name, description}]}: extension commands, prompt templates, and skills as skill:<name>, without the leading /. With cwd, the server lists what a new thread in that project would offer, from a session it builds in memory and never binds: no extension starts, and nothing is written. Built-in commands that only work in the terminal (usage, llama, undo) are left out.

{"id":17,"method":"command/list","params":{"cwd":"/path/to/project"}}

ui/event

Params: {threadId, widgetKey, id, value?}. Result: {}. A press of the button id, or a change of the field id, in the widget widgetKey's view: value is a string for a text or select field, a boolean for a toggle, and absent for a button. Every handler of the thread gets {widgetKey, id, value?}; a handler that fails becomes a warning. The extension answers by setting the widget again: a field shows the view's value, not what was typed. A press of a widget or control that is gone answers {} and reaches nobody.

Errors: threadNotFound when the thread is not loaded, and invalidParams when the value does not fit the control (a toggle's string, a select value outside its options, a button with a value).

{"id":18,"method":"ui/event","params":{"threadId":"THREAD_ID","widgetKey":"checks","id":"verbose","value":true}}

thread/delete

Params: {threadId}. Result: {}. Permanently removes an archived transcript; it does not use the interactive resume selector's trash command. Archive first, then explicitly delete. The thread must have no running/admitted work. Deletion fences new starts and reloads, rechecks archive/busy state inside that fence, and locks/verifies the raw file before unlinking. Refused rechecks preserve the loaded thread and subscriptions. Success broadcasts thread/deleted: {threadId} to every initialized connection.

Errors: threadNotFound, threadNotArchived, turnInProgress, threadLockedElsewhere, and archiveConflict for hard-linked transcripts; filesystem failures return internal.

{"id":16,"method":"thread/delete","params":{"threadId":"ARCHIVED_THREAD_ID"}}

server/diagnostics

Params: {}. Result: rssBytes, heapBytes, loadedThreads, loadingThreads, payloadFiles, payloadBytes, inactiveFactories, uptimeMs, listingFilesOpened, and listingBytesScanned. The listing counters are cumulative: transcript scans account for file sizes and parent-header lookups add opens, but not bytes. They exclude index reads and filesystem stat work. Diagnostics observe the shared server, not only this connection.

{"id":18,"method":"server/diagnostics","params":{}}

thread/diff/read

Params: {threadId, scope?, against?, commit?, context?, ignoreWhitespace?}. Result: {scope, root, files, filesOmitted?, untrackedSkipped?, turnId?, stale?, against?}. files are {path, oldPath?, status, binary, additions, deletions, tooLarge?, hunks}, their paths relative to the repository's root (not the thread's cwd, which can be a subdirectory of it); each hunk keeps its lines with their " ", "+", "-" or "\" prefix. committed needs commit; branch compares with against, else the worktree's base, the upstream or the default branch, and echoes it. More than 500 untracked files are counted in untrackedSkipped and left out. Files past the response budget come without hunks (tooLarge) or are only counted (filesOmitted). A hunk id hashes the file's header (its blob ids) and the hunk at the read's context: an action must pass the same context (a context: 0 id matches no action), and after an action changes a file, read it again, as its other hunks' ids change too. A hunk id from an ignoreWhitespace read matches no action. Reads take no lock and never write the user's index. Errors: threadNotFound, notGitRepo; invalidParams for a missing commit or no branch to compare with.

{"id":20,"method":"thread/diff/read","params":{"threadId":"THREAD_ID","scope":"uncommitted","context":3}}

thread/diff/revert

Params: {threadId, scope?, against?, commit?, path?, hunkId?, context?, reapply?, turnId?}. With turnId (lastTurn only), a last turn other than that one answers gitConflict diffChanged, so an Undo shown on one turn never reverts a newer one. Result: {paths}. Reverse-applies one hunk, one file, or the whole scope to the working tree; lastTurn without a path undoes the turn, and only while every affected file is exactly as the turn left it. With reapply (lastTurn only, no hunk), the turn's changes apply again instead, and only while every affected file is exactly as the turn found it. Hunks of renamed, added, deleted, binary or mode-changed files revert as whole files only. Refused while a turn runs in the checkout; no turn starts and no tool runs there until it ends. Errors: turnInProgress; invalidParams for staged and committed or a path outside the diff; gitConflict with diffChanged or patchFailed.

{"id":21,"method":"thread/diff/revert","params":{"threadId":"THREAD_ID","path":"src/a.ts","hunkId":"HUNK_ID"}}

thread/handoff

Params: {threadId, to: "local" | "worktree", branch?, environment?}. Result: {thread, branch?, warnings}. To the main checkout: the worktree's HEAD becomes a branch checked out there (the worktree's own branch, else branch, else the previous handoff's branch fast-forwarded, else <worktrees.branchPrefix><title>), the uncommitted changes move, and the worktree stays, detached and clean. Back: the worktree is detached at the main checkout's HEAD and the changes move into it; a thread born in the main checkout gets a worktree. The changes are saved under <agentDir>/handoff/ before anything moves and leave the source only after they landed. The thread is reloaded where it now runs: subscribers receive thread/replaced with the same id, and queued inputs follow it. Errors: turnInProgress; invalidParams for a thread with no message yet or already on that side; gitConflict with localChanges, operationInProgress, branchExists, recoveryRequired (with data.patch) or worktreeDeleted.

{"id":22,"method":"thread/handoff","params":{"threadId":"THREAD_ID","to":"local"}}

thread/worktree/restore

Params: {threadId}. Result: {thread}. Rebuilds a deleted worktree at its old path from its snapshot: detached at the old HEAD, with the working files and the staged state as they were (ignored files were not kept). Then the thread loads there. Errors: threadNotFound; invalidParams when the worktree is not deleted; gitConflict with worktreeMissing (deleted without a snapshot: fork the chat into a new worktree instead) or submodules.

{"id":23,"method":"thread/worktree/restore","params":{"threadId":"THREAD_ID"}}

worktree/list and worktree/delete

worktree/list {cwd?} → {worktrees: [{path, projectCwd, head?, branch?, threadIds, lastUsedAt, deleted}]}, the managed worktrees (of cwd's project when given). worktree/delete {path} → {snapshot?} deletes one after its snapshot. A worktree a thread is loaded in, or one holding a nested repository (an initialized submodule or a repository inside it), is never deleted: archive and pruning skip it, and worktree/delete answers gitConflict with worktreeInUse or submodules. Only registrations in the managed layout are removed; the user's own worktrees are never touched.

{"id":24,"method":"worktree/delete","params":{"path":"/home/u/.knightcode/agent/worktrees/ab12/project"}}

environment/list and environment/save

environment/list {cwd} → {environments, trusted} reads the project's .knightcode/environments.json; environment/save {cwd, environments} writes it (names unique, at most 50). An untrusted project's scripts never run. See Settings.

{"id":25,"method":"environment/save","params":{"cwd":"/p","environments":[{"name":"node","setup":{"default":"cd \"$KNIGHTCODE_WORKTREE_PATH\" && npm ci"}}]}}

git/*

Each acts on the thread's own checkout: its worktree, or the user's checkout. git/status → branch, upstream, ahead/behind, staged/unstaged/untracked/ conflicted counts and an operation in progress. git/stage and git/unstage take {threadId, path?, hunkId?, context?}: no path is everything; paths are literal. git/commit {threadId, message?, includeUnstaged?} stages everything first unless includeUnstaged: false; an empty message is written by the thread's model from the staged diff, outside the conversation, and the commit answers concurrentChange when the checkout changed meanwhile. The user's hooks run. git/push pushes the current branch with one explicit, non-forcing refspec and sets the upstream the first time. git/pr/create {threadId, title?, body?, draft?} runs gh pr create (toolMissing without gh). git/branches, git/branch/create {name} ("Create branch here"), git/branch/switch {name} (refused when it would overwrite local changes or an ignored file the branch tracks) and git/log {limit?} (the commits since the base, for the committed scope) complete the set. Errors: notGitRepo, invalidParams, authRequired, toolMissing; gitConflict with diffChanged, patchFailed, concurrentChange, pushRejected, branchExists or localChanges.

{"id":26,"method":"git/commit","params":{"threadId":"THREAD_ID","includeUnstaged":true}}

Ownership, unloading, and shutdown

Only one process may own a persisted session file. Symlink/junction aliases share ownership; hard-linked transcripts are refused. threadLockedElsewhere means another process owns it. Terminal resume offers to fork instead. Read-only discovery and history require no ownership.

If ownership is lost to another process, the thread closes with an error notification rather than continuing writes. Archive/unarchive owns both paths, refuses destination collisions, and never overwrites another transcript. An archiveConflict can include structured recovery data naming surviving files.

A thread unloads after ten minutes without a subscriber and with no running work. Unloading releases its session lock, owned virtual routes, and extension-cache leases. The factory cache retains at most 256 inactive factories plus actively leased entries. Runtime module unloading is not promised; see extension state and cache behavior.

On stdin EOF or SIGINT/SIGTERM, new input stops immediately, pending login flows cancel, and every thread unloads. Accepted requests still receive answers; work that cannot finish returns shuttingDown. Output drains after accepted requests settle. At nine seconds, unfinished accepted dispatches receive shuttingDown; late results cannot publish a second response. The final second is reserved for output drain before the ten-second hard exit, including a client that has stopped reading stdout. Trusted JavaScript already executing remains cooperative; the server cannot roll back its private effects or preempt a synchronous callback.

Limits and client reference

Limits count UTF-8 bytes:

• 64 MiB per stdin line; an oversized line is dropped with an error notification.
• 64 MiB of pending unread output before the connection aborts.
• 1 MiB per history/list response, including envelope reserve.
• 64 KiB per inline turn/item; larger values use references.
• 48 KiB per payload chunk before base64 encoding.
• 64 KiB per view in a tool result or an extension message, as serialized JSON; a larger or invalid view is dropped and the raw output or text shows.
• 16 KiB per widget (its lines or view), 16 widgets and 50 statuses per thread; status and widget keys of 1 to 128 characters; status text, title and working message cut to 200 code points. A refused widget or status gives one warning per key per load (a load warning when it happened while the thread loaded).
• Views nest stacks and lists at most 8 levels deep.

The in-repository AppServerClient supports typed request, onNotification, onRequest, readPayload, resolveTurn, and resolveItem calls. Its AppServerError preserves code, message, and data exactly; inspect data for archive recovery information instead of parsing message strings. createInProcessClient connects the same server/client over an in-memory transport. A host owns the shared context and ThreadManager: stop input, dispose the context to cancel logins, and await threads.closeAll() on shutdown.

Every message is defined in app-server.schema.json. Regenerate it from the workspace:

cd packages/app-server-protocol
bun run generate:schema