zidane/chat/pure

Resolve a profile's accent token to a concrete hex color via the caller's color palette. Renderer-agnostic — accepts the four-role subset of ThemeColors that's relevant to accent surfaces, so this helper composes against either useColors() (TUI) or a CSS-variable lookup (GUI). Omitted / unknown tokens fall back to accent.

The problems of an action.

The plan worth a glance: the session scope's applying plan, else its most recent stopped one. null when neither exists.

The snapshot the footer indicator shows for a provider: the one probed with the credential requests use — the session's pinned account when it has one, else the provider's active credential — else the first reported. null when none.

The space to show: the remembered one when it still exists, else the first space in the list (Default unless the user moved it).

The relative-age ladder every "how long ago" label climbs: under a minute, then whole minutes, hours and days, then (from dateAfterDays) a date. Only a duration reads with "ago" in a sentence: "Updated 5m ago", but "Updated now" and "Updated Mar 4".

Compact relative-time formatter — "just now / 5m ago / 3h ago / 2d ago".

The visible text of a line.

Coalesce a markdown delta into owner's open streaming block, or open one.

Coalesce a thinking delta into owner's open lane, one row per line.

The scope's applying plan, if any.

Replace [span.start, span.end) in text with insertText. Returns the mutated text and the new cursor position (end of insertion).

Applies one event to a bag. Pure: the input bag is never mutated. An event whose id was already applied changes nothing; a refused event returns error and the input bag.

Settings with a change applied. A setting that turns on also turns on the one that it needs, so Session titles turns on Decisions, and a feature turns on what it includes, so Browser tab turns on Agent can use the browser. A setting that turns off also turns off the ones that need it.

The card's one-line headline, in plain English (the TUI's words; the GUI words it through labels with the same shape).

Whether the card also offers signing in again next to its retry: a retry that keeps failing on a stuck refresh is fixed by a fresh sign-in.

The card's main action. A refresh stuck in another process or one that could not reach the provider is retried first; anything else (expired, revoked, missing, rejected) needs a new sign-in.

The provider key a failure is about, for matching it against the host's provider list (a descriptor's key, or its credentials key).

The automatic text model among one provider's models; undefined when nothing qualifies (the caller then uses the main model).

The chords one binding lists. A binding may name several, any of which fires the action: "ctrl+w | cmd+w" (or, in keybindings.json, an array). A | right after a + / - is the key itself (ctrl+|).

Blend two hex colors in HSL space with shortest-path hue interpolation. t ∈ [0, 1]; t=0 returns from, t=1 returns to.

The command that runs a block in the session's folder, with no input, for Run here.

The key of a block's run: its reply and its code. An edited or regenerated reply is a new turn, so its blocks start without a run. Two identical blocks in one reply share their run.

What is typed in a terminal to run a block there, ending with its Enter. The command keeps the terminal's input, so it can ask for some.

Static gradient ramp of length n going from from (index 0) to to (index n-1) in HSL space. For the cycling A→B→A→B ramp the throbber uses, see buildCycleRamp in src/tui/crush-throbber.tsx.

What a finished run writes: the settings change, the provider pick, the look and the records.

Built-ins first, then extension panels in registry (load) order.

The message the Resume button sends: the agent continues the plan by writing its step ids, which reopens it (the gate's resume rule).

Builds a SessionSearchIndex, skipping turns with no searchable text.

Yielding index builder for TUI callers — never monopolizes a frame.

The message a step's Retry sends (a steer mid-run, a prompt when idle). Quotes the step so the agent matches it against its own list, and names the status: retrying a completed step differs from reviving a cancelled one.

Builds a unified diff for an edit payload. Line numbers in each @@ header are synthetic, since hunks carry only snippets. A write_file that creates a file uses --- /dev/null. The result ends with a newline.

The hint for an option: re-reads ~452k tokens when picking it would re-read a long prompt. Nothing for the current option, a switch that keeps the cache, a cache already expired (it is gone anyway), a short prompt, or an unknown estimate.

The factor that corrects this project's estimates: the median of actual / estimated over its recent finished runs, kept within CALIBRATION_MIN..CALIBRATION_MAX. null with too few runs.

A key name as the parser canonicalizes it ('Enter' → 'return', 'ArrowLeft' → 'left').

Whether the user can skip the current step. Welcome and Ready have nothing to skip, and the TUI cannot run without a provider.

The short tags a host's row shows: the tools found and the cloud kinds available, at most max.

output cut to its last BLOCK_RUN_OUTPUT_CAP characters.

Bound a transcript event array for display. Identity below the cap (no re-render churn on the hot append path); above it, drops the oldest events down to TRANSCRIPT_TRIM_TO_EVENTS — extended to the next turn boundary — and prepends a cumulative hidden-count marker. Callers: the stream buffer's drain commit and every eventsFromTurns replay site.

Whether entry may join a chain led by chain[0] (one account per provider; see filterChainFallbacks).

Identity of a chain entry — provider, account and model, not object identity.

1-based chain position of a provider+model(+account) pick, or null when not in the chain.

Compares two chains element by element using chainEntryKey.

Same entries in the same order.

Extracts the #token chat references from text. A token starts after whitespace or at the start of the text. Trailing punctuation is stripped, tokens are lowercased and duplicates are dropped.

Classifies a connection attempt. HTTP servers that required auth or answered 401/403 become token-rejected when the config sends its own headers, else needs-auth. Any other error becomes error. A clean connection is ok.

A line cut to max visible characters, the last one an ellipsis when it is cut.

"14:05" when at falls on now's day, "Oct 12" in the same year, else "Oct 12, 2025".

"14:05", local time, 24-hour.

Closes the panel and keeps the active tab.

Collect every provider's references in one pass. Convenience wrapper — the TUI textarea component calls this on every keystroke to highlight in-prompt mentions.

Ran <gist> for a shell command: heredoc bodies dropped, setup steps (cd, export, set -e) skipped, env and wrappers peeled off, program plus subcommand kept. Further meaningful steps read as +N.

compactingParts as one plain string, for a slot that cannot hold badges: Compacting... ~12k, 8s.

Numeric, segment by segment; a longer version wins a tie on the shared prefix (5.5 > 5).

Orders two present values: numbers, then strings (natural order, case ignored, then exact), then booleans (false first), then anything else by its JSON text.

Compiles a query into a predicate. Relative bounds ('-7d') resolve against options.now each time the predicate runs, so a long-lived predicate stays current. Throws SessionQueryError for an invalid query.

Computes word-level segments for a paired removed and added line. Both segment lists keep the same order. Tokens that differ have changed: true. Very long lines fall back to one fully changed segment.

Computes a line diff between two strings with an LCS table. Common leading and trailing lines are trimmed first. When the changed core is too large for the table budget, the result is all removals followed by all additions.

How a connector signs in, when it is not the usual browser sign-in.

A connector by id, or undefined.

Why zidane cannot add a connector with this input, or undefined when it can. The one reason is a missing token that the connector needs.

Whether a connector is already installed: a server with its name, its endpoint, or its command and arguments.

The MCP server a connector writes to mcps.json.

The connectors that an import brings over: the MCP servers that it copies and that have the id of a connector. The Connectors step marks them, and the run does not add them twice.

The names a copy of name tries, in order: itself, a copy.ts, a copy 2.ts, … (a folder or a dotfile keeps no extension apart).

Count turns before / after the one identified by turnId in the given list. Returns null when the id is missing. Used to label the turn-details modal with N before, M after.

How many calls a script makes, when every call can run once at most: each tools.<name>(…) or sh(…) sits outside loops and functions (callbacks, helpers, methods). undefined when one may run more than once, when the script reaches tools or sh another way (a reference, a computed name), or when it makes no call. A call in a branch that does not run makes the count high, never low, so a percent from it never runs ahead.

Creates an incremental AnsiLines reader. The style carries across chunks and lines, as in a terminal.

Creates a completion provider for # chat references. Each chat gets the shortest id prefix, at least 6 characters, that identifies it in the catalog. Suggestions are fuzzy-matched on title, token and project folder.

Build a commands completion provider against a live command list. The getter is re-evaluated per suggest so registry toggles / cwd rebuilds take effect without re-instantiating the provider.

The entry actions over port, whose paths are relative to the folder (the caller's port maps them where they live).

Creates an empty index of which host holds each id.

A new run at its first step.

Starts a run of the setup wizard.

Creates an empty OutputTail.

rankPaletteItems for a palette being typed into: it remembers the last query, and when the next one only adds to it (one more character, one more word) it searches just what matched before. A longer query never matches an item the shorter one missed, so the result is the same.

Creates an empty presence merge that keeps the last snapshot of each host.

Creates an empty ScriptProgressTracker.

Schedules: the wire shapes, the checks and the time math that the host, the TUI and the UI share. The cron math (./cron, on croner) and the scheduler (./scheduler) are imported by path, so this entry stays free of them. See docs/SCHEDULES.md.

Keys (provider\0model) of the current models: per provider and family, the newest version, one entry per size. keep entries (the active model, fallback chain, off-catalog ids) are always included.

entries in their original order, current models only (see currentModelKeys).

The current step.

Every route for a typed model id, one per provider: the provider the query names first (xai grok-9), then fallbackProviderKey (the current pick), then the rest in the given order. Lets a picker offer "use this id via…" for each signed-in provider instead of guessing one. Empty when the query names no model id.

Cycle to the next / previous tab (wraps). Counts as a user pick.

The day of the year (1-366) of d: a seed stable within a day.

The count days ending at end (inclusive), oldest first, as local YYYY-MM-DD.

Whether the agent gets decisions: the user turned them on, or Model decision approvals (approvalMode: 'decide') need their answers.

The default catchUp of a trigger: a missed reminder still arrives, a missed recurring fire does not.

The default source pinned to a release, Tahul/zidane#v<version>: the native extensions that match the running app's API.

The selection that Bring your setup starts on. It has every piece with something new to copy. It has the sessions of this project, or of every project when this project has none.

The agent a new session in spaceId starts on: the space's (spaces on), else the global one (Settings.defaultAgent), else undefined (Build). An id the registry doesn't hold is the caller's to resolve (resolveAgentId).

The effort and options a new session in spaceId starts with, whatever its model: each the space's own (spaces on), else the global one (Settings.defaultEffort / defaultModelOptions), else absent (the last used with the model). A pick in the composer wins over these.

The model a new session in spaceId starts on: the space's default (spaces on), else the global one (Settings.defaultModel), else undefined (the caller's last used), with the effort and options new sessions start with (defaultSessionKnobs). A pick in the composer wins over this (newSessionModel).

Creates a FeatureGate from its setting key, kill-switch env var, and CLI flag.

A duration in words: 90 min → 1 h 30 min.

A worker's end from its spawn-end marker. Reads spawn.end, or parses the text of a marker recorded before it existed (done · 42s · $0.01 · in 3100 / 400 out), whose facts become badges verbatim.

A time for lists: today 09:00, tomorrow 09:00, else Mon 3 Nov 09:00.

A trigger in words, for lists: once at 09:00 tomorrow, every 30 min, cron 0 9 * * 1-5 (Europe/Paris).

Drop one owner's streamed output for a turn: the reasoning, and the text, of a provider attempt the loop abandoned (a stalled stream retried, a connection that dropped mid-answer, or a move to the next model), which the next attempt replaces and the saved turn never holds. The loop only abandons an attempt whose turn has no other output, so nothing kept is lost. Returns the same array when there is nothing to drop.

Scheme, host, port and path of a URL (token-looking path segments redacted): what is safe to show or log.

The text an edit of turn starts from: its text blocks one paragraph apart, after rejoining a cited Claude reply's fragments (mergeReplyText), so a cited answer edits as the one text it is and, saved unchanged through replaceTurnTextBlocks, stays exactly as written.

The mode a run ends in: Quick start turns a full run into a quick one.

The list with name turned on, or undefined when it already is.

The list without name, or undefined when it isn't listed.

The kill-switch env vars engaged in env, sorted. The GUI main process hands this to its renderer (which has no process.env) so settings rows can say why they are inert.

The last segment of path.

The folder holding path ('': the root).

path as the rules take it (no trailing /), or null when it is not a valid entry path.

A session's environment as its header shows it: local@laptop, worktree@laptop ⎇ feat, ⎇ feat, or the kind.

text as a template that renders to itself: each brace doubled. Use it for user text inside a template (`${escapeTemplate(bot.name)} · {modelId}`).

One level's label, as the decision model reads it.

What view says at now.

When view was made: its estimate's time, or its failure's.

The mark of an extension: icon.gui (unless it is the native catalog's own icon and the catalog has a logo), then brand, then the native catalog's brand, then its icon.gui, then the first letter or digit of the display name.

An extension's space as a Space: read-only, its newSession defaults as the space's own (agent, model, effort).

Extracts an EditPayload from a tool call's raw input. Supports the edit, multi_edit and write_file tools. Returns undefined for other tools or when the input is malformed.

The category a field belongs to.

Maps a file path to a tree-sitter filetype name. Query strings and fragments are ignored, and the extension match is case-insensitive. Returns undefined for unknown or missing extensions.

Retain only records at or above minLevel. Preserves order.

The rows that match where, in input order. now is read once for the whole list.

contentMatches are keyed by keyOf(meta): the bare id by default, a (host, id) session key for a list across hosts (twins are two rows).

Flip any trailing streaming markdown blocks (any owner) to finalized.

Flip the trailing streaming markdown block for one specific owner.

One space looked up by id or name (both case-insensitive), the id match first. undefined for a blank query.

Fit cells to width columns for a renderer that draws them two blank columns apart (the TUI's badge rows): returns the cells that fit whole. A cut cell is truncated with …; anything after it is dropped. Never wraps. prefix is the columns the renderer already spent (glyph + space).

query with nested all in all and any in any merged into their parent, and a one-element all / any replaced by its element. Matches the same rows.

USD cost, dollar-prefixed — 2 decimals, 4 below a cent so tiny spends stay visible (a haiku worker's $0.0007 must not render as $0.00).

Compact token formatter — 12_415 → "12.4k", 1_234_567 → "1.23M".

Compact density: does this item fold into its exchange's one row? Thinking, tool traffic, worker lifecycle, a / command's details and everything a worker produced fold; the prompt, the assistant's prose, errors, notices, the user's own ! runs and session events stay.

A settled fold's header: its own calls in the shared words (Read a file, ran a command +2), with the workers after them when no spawn call already says so (, 1 worker: a script started it). Exchanges without calls say what they did: Delegated to a worker, Thought, Worked.

20.00 / 25.00, or just the remainder when no total is reported.

Byte counts for people: 512 B, 1.5 KB, 2.0 MB.

USD: formatCostLabel below $10 ($0.004, never $0.00), whole dollars below $10k, then $12.3k.

Live cost, readable: sub-dollar runs keep a tenth of a cent (so the figure visibly moves while a cheap run streams), then cents with grouping: $0.123, $4.56, $1,646.17. null for nothing spent, never $0.00.

$0.04, or four decimals under a cent so a cheap run is not shown as $0.00. Same as fmtCost.

A count at a glance: 950, 1.2k, 12.4k, 120k, 4.6M, 1.2B. One decimal (a trailing .0 dropped) below wholeFrom of a unit, whole numbers from there. A value that rounds up to 1000 of a unit reads in the next one (999_999 → 1M, never 1000k).

Balances are dollars for OpenRouter and opaque units elsewhere.

Run time as a person reads it: 42s, 1m 05s, 2h 03m.

A remaining time as a chip says it: ~45s, ~11m, ~1h 20m, ~12h, and 24h+ past a day (the open-ended last step: no number pretends to know).

0-23 → "9 am" / "2 pm" / "12 pm".

The settings string for a pick (inverse of parseModelSetting).

2m18s / 48s / <1s / 1h05m. Compact — this is a receipt, not a profiler.

The receipt's cells, in display order. Renderers show each as a badge, tint it by kind, and truncate to their own width; the CONTENT is fixed here. modelName names the model as the surface's picker does.

The plain-text receipt for slots that cannot hold badges (the event's text, a copy): cells comma-joined, ellipsized to width columns. Never wraps.

1.2k / 48k / 1.3M — token counts are read at a glance, not counted.

42 tk/s; one decimal under 10 so a slow model is not shown as 0 tk/s.

Run a tool's curated formatter and return the result, or null when no formatter is registered / the input shape doesn't match. Renderer decides what to do with null — typically: show ↳ <displayName> with no target / meta tail.

Percent for display: integers print bare, fractions keep one decimal (so a status bar never reads 100% for 99.9).

The first free name for name in folder dir, compared without case (a case-insensitive disk sees A.ts and a.ts as one).

Subsequence match of query in text with the usual bonuses: word starts and the very beginning score higher, gaps cost, runs of adjacent hits are rewarded, and the query found whole (as one run) beats the same letters scattered. Returns null when not every query character appears in order. Case-insensitive; query must already be lower-cased.

One receipt line per exchange: every command-note in an exchange moves to its end and joins into a single event (committed 83dfea7, pushed to origin/main). Returns events itself when there is nothing to merge; the merged event keeps its identity while its notes do not change, so memoised rows stay put.

The group keys of a value: one for a scalar, one per distinct element for a list, none for a missing value or an empty list.

Sections for display. With a query the ranking already interleaves providers, so groups follow the order of their best row. Without a query the recents (ids in recents, present in ranked) form a leading Recent section and the rest keeps provider order.

Groups rows by groupBy.field. Rows keep their input order inside each group, so sort first. Groups are ordered by groupBy.order; the group of rows without a value (key null) comes last. Throws SessionQueryError for an invalid grouping.

Whether text holds anything parseAnsi changes: an escape, a carriage return, a backspace.

Anything to show? A run with no record data renders no line at all.

Whether name is on the host's PATH; undefined while unknown.

A value's level against heatmapThresholds.

The quartile bounds of the values above 0.

Splits text into highlighted and plain runs along ranges.

formatBytes with whole kilobytes (2 KB), for short notices and attachment labels.

Brings items into folder dir through create (one call per entry, so no call carries more than one file): each top-level item takes a free name, what is inside a dropped folder keeps its own. The limits are checked before anything is written; a failure on the way stops the import and says what was made.

Whether a selection brings anything over.

Find a catalog entry's index by its {providerKey, modelId} tuple. Returns -1 when not present. Useful when re-rendering the picker (a query just narrowed the list, where did the selection land?).

Short badge for a segment's origin — the "where does this come from" column. 'file' is intentionally the only one that reads like a location, since it's the only origin the user can act on.

Only http(s) pages (and the blank page) ever load in the dev browser.

Whether a notice is one showCacheNotices hides: a prompt-cache break or expiry, or dropped reasoning (the same prefix churn, seen from the thinking side).

Whether the install existed before the wizard: it has stored settings, credentials, sessions or a last-used provider.

A provider earns a chevron only when there is a real choice to reveal.

Whether spaceId names an extension space (<extension>/<id>).

line closes the fence opened with fenceChar × fenceLen: the same character, at least as many, and nothing but spaces after (an info string makes it a line of code, as in CommonMark).

name is set to '1' in env: the kill-switch test, for a row that only knows the variable.

A loopback address: the dev servers the browser exists for.

Type guard over untrusted metadata.

The host predates the method (models.list, catalog.*, …): callers degrade, never fail.

Whether the program asked for mouse reports.

The isolation a kind gives, named without its machine (the picker names the place apart): In the folder for local kinds, New worktree for worktree kinds, else the kind's own label.

Type guard hosts use before forwarding a payload from an untrusted provider.

Whether value is a relative time: 'now', or a sign, a count and a unit (s, m, h, d, w), such as '-7d'.

Whether view has a timing to count down (not a problem).

Whether value is a SecretAnswer: a value or a decline sent by a client.

Whether name is a valid secret name (SECRET_NAME_PATTERN).

Whether value is a SecretReference.

Whether value is a SecretTtl.

Whether a fence of language runs.

Is this event a user-initiated ! bash exchange (vs. a model-driven shell call)?

A machine name: one path-safe segment, so local@<name> is unambiguous.

name in folder dir.

Up to two names, then a count: "A and B", "A, B and 3 more".

A list of names as a sentence reads it: "A", "A and B", "A, B and C".

"seen 5 min ago" for a machine's lastSeen.

The machine names of the saved servers a local host uses as machines (machines.links.status uses): pickers list those servers once, as servers, not again as local@<server>.

Session cost while a turn streams: settled figure plus the running estimate for the in-flight turn. estimated tells the surface to render it dimmed (the number will move, then snap to the settled value at turn end). null when there is nothing to show — never $0.00.

The IANA time zone of this process (UTC when it has none).

"Monday, Oct 26" for a YYYY-MM-DD day.

Why a machine kind can't be picked now (its machine is offline), else undefined.

machines.listDirs as a folder listing: the parent stops at the root the folder is in.

A machine-qualified environment kind: <base>@<machine name>.

The command a machine runs to join a main server with a join string (formatJoinString).

Whether a machine runs another zidane version than its main server. false when either is unknown.

The machine a session's run is paused on, from its notices (waiting for laptop… until laptop is back. or laptop stayed offline … run stopped), null when none. online: machines known to be back since (a machine:status the window saw), which clear the wait at once.

code with the text of strings, comments and regular expressions replaced by spaces (newlines kept), so a scan sees only code. Template literals keep their ${…} expressions.

Cheap pre-check before splitting: the text may end in a stop marker.

The media a path names, by its extension; null for everything else.

Stored records with new ones merged in. A finished run (any record) clears the in-progress marker.

base with every other host's rows laid over it, by session id.

Merge reference lists from multiple providers into one ordered list with earlier-start-wins disambiguation when spans overlap. Ties broken by insertion order. Spans are sorted ascending so renderers can walk them sequentially with a cursor through the source string.

A message as a person skims it: its words, markdown syntax and code blocks dropped, whitespace folded. For the compact stacks of earlier messages.

Ownership for "one selection stop per message": result-only turns fold into their assistant turn (turnSelectionOwnership), and turns that only did work (thinking, tool calls, worker markers) fold into the answer that follows them in the same exchange. "Read three files, then answered" is one stop, so ↑ moves one visible message at a time. An exchange that ends in work alone stops on its first work turn. Same rule as the GUI's computeSelectStops.

The records an install that predates the wizard gets, so it never opens on its own.

Why the decision model handed a call to the user, for the approval card.

Builds the picker name for a model: the catalog name when given, else the id made readable, with a vendor prefix split off and the provider's own brand word removed.

Identity of a model pick on any shape: the same model on two accounts of a provider is two picks. The one key every picker and chain edit compares by.

Sequences that put a fresh terminal in state's modes.

Length of the mode sequences text starts with: at least the modePrefix put in front of a snapshot. Output that itself starts with mode sequences counts too, so this errs towards less output, never more.

The window closest to its ceiling, which is the one that will bite first and the only one a single-line indicator has room for. null when the snapshot carries no windows.

Move one visible tab delta slots within the visible strip (keyboard reorder), clamped at the ends. Hidden pinned ids keep their slots.

A native extension by name, or undefined.

The listing row of a native extension, from the catalog shipped in the app.

The native extensions as a listing of the default source pinned to version, built from the catalog shipped in the app (NATIVE_EXTENSIONS): what an app shows when the repository cannot be read. No network.

What a native extension needs, in words: "gh (optional)", "the Linear connector".

The install source of a native extension, pinned to a release tag: Tahul/zidane#v<version>:extensions/<name>, the grammar installExtensionFromSpec takes.

What an app does at launch about the wizard (docs/ONBOARDING.md, When it opens).

Whether a script's nested call shows in the transcript: a call that runs a subagent (its worker attaches to it), a plan or a question, or any call with codemodeShowCalls on. Every surface (the host's live rows, isVisible, the kit's rows) applies this one rule.

A persisted record as the events the live run emitted for it.

The tool-result row for a nested call as it ends; the output is cut to a preview.

The tool row for a nested call as it starts.

The model a new session starts on: an explicit pick wins, else the space's default. fallbacks is set only when the result IS the space default (a different pick must not inherit the space's chain); undefined fallbacks leave the caller's usual seeding (lastModelChain) in place.

A fresh, unused space id (space-<n>).

Schedules: the wire shapes, the checks and the time math that the host, the TUI and the UI share. The cron math (./cron, on croner) and the scheduler (./scheduler) are imported by path, so this entry stays free of them. See docs/SCHEDULES.md.

Next min-level in LOG_LEVEL_CYCLE, wrapping around.

Normalize address-bar input to a loadable URL, or null when it is not one.

Coerce whatever a provider returned into well-formed items: entries without an id or title are dropped, strings are capped, unknown kinds fall back to action, and source is stamped with the provider id. Never throws.

Same guard for a run result: anything malformed becomes undefined (nothing to do).

The stored list made safe to use: malformed entries and duplicate ids dropped, and the Default space always present (first when it was missing). Order is kept: it is the switcher order.

A / command's details line: its label, capitalised, or "Details".

Trimmed user-given account name, or DEFAULT_OAUTH_ACCOUNT_LABEL.

"offline since 14:05" (a date when not today) for an offline machine's lastSeen; "offline" when never seen.

The label and description of a feature, from the settings catalog.

The features that a run offers: the ones of its app. An extras run offers only the ones that no other app offers, because the other app set the shared ones.

The parts of an import that the user picked, in words, in the order of the step.

How far a run is.

A record for a run that just ended.

The records that a finished run writes. A full, quick or repeated run writes the shared record and the record of this app. An extras run writes only the record of this app. Connect alone writes none.

The next state of a run.

The settings that a run writes. Only the keys that the run changed count. A value that equals the default is not stored, so it keeps following the default. If the key is stored, the run resets it instead of pinning it. A setting that is off only because the one it needs is off is not written: it comes back as it was when that one turns on again.

The display data of a step.

The features that a run switched on and off. Only the ones it changed: the defaults are not news.

text on one line: line breaks, tabs and runs of spaces become one space.

The agent is blocked on the user (a plan to approve, a question to answer): show the tab regardless of earlier picks. Costs at most a tab switch, and nothing in that tab decides implicitly. Not a user pick, so it does not lock out later proactive opens.

Agent / extension asked to show a tab. Honoured only until the user has picked a tab themselves (a badge that flipped must never steal the tab); otherwise the request is dropped and the state returned unchanged.

The main session id a worker reports to, or null for an ordinary session.

Environment kinds in picker order: this host's own first, then each machine's together (local@ before worktree@), machines by name.

The delivery that is not delivery: what Alt+Enter does while a run works.

Returns the owner of an event: its childId, or parent.

One page of rows. Pass the page's cursor back for the next one. A cursor resumes at the first row that sorts after the row it ended on, comparing sort keys then id, so rows that moved, appeared or went away in between never shift the page. Throws SessionQueryError for a bad limit, a bad cursor, or a cursor made for another sort, filter or scope.

Provider key a row "belongs to" — its own for a provider header, the parent's for a source row.

text as lines of styled spans. A text without a closing newline still yields its last line; an empty text yields one empty line.

The spec's first valid chord (callers that pre-parse one binding). Use parseBindingSpecs to honour every alternative.

Every chord of spec that parses (see bindingAlternatives).

The fence line opens, or null: three or more backticks or tildes, and a backtick fence's info string holds no backtick. Every markdown scanner that must agree on where code blocks are (here, and the desktop app's chunker) reads fences through this and isFenceClose.

Reads a field path apart, or undefined when it names no field and no well-formed facet.

Parse one key/value pair PER LINE. Line-based (not whitespace-split) because values legitimately contain spaces — Authorization: Bearer abc123 is the canonical headers case, and env values like GREETING=hello world are equally valid. The first separator found (the separator that appears first in the line) splits. The rest of the line is the value, trimmed. Blank lines skipped.

{ base, name } of a machine-qualified kind, null for any other.

Splits a model id into vendor, family, version, size, snapshot and routing variant. Parsing is lowercase and never throws.

Parse a model setting; undefined for unset or malformed.

Validate and normalize a host URL (no trailing slash); throws RemoteHostError.

Reads a git or forge URL (or owner/repo) as what git clone would clone; null when it is not one.

The pair key names (a bare id: This computer's).

Reads the time phrase at the start of text. null when the text does not start with one of the forms above.

The duration below which the answer puts probability p (0..1), from the per-level probabilities. null when they don't add up to anything.

Moves entry to position 0, adding it when absent.

The user clicked a tab (or a keybinding cycled to it). Always honoured, always opens.

Add a tab to the host's pinned strip (idempotent, keeps order).

Decide how a file is attached. Images go by type (or extension); PDFs by type, extension or their %PDF- signature; anything else is text when its first bytes read as text, sent as text/plain unless its type already names text (a .ts file the browser types video/mp2t is text).

The steps of one exchange in order, and its counts. Consecutive thoughts are one step; each worker is one step holding its lane's items, merged with the spawn call that started it (the call's line opens onto the worker's progress). A worker's state is its own structured end: failed only when it errored (core describeSpawnEnd), stopped reads done.

The steps that a run shows, in order. It starts from the steps of the mode. It leaves out Bring your setup when zidane found no other agent, and Connectors, Skills and Extensions when installs do not run on this computer. It also leaves out the steps that the host excludes. A connect run always has Connect.

Counts a plan's steps.

Position label for compact surfaces: the active step's position, else the completed count ("3/7"). null for a plan with no steps.

1-based position of key in the chain, or null when absent.

Previews an edit payload the way the edit tool would apply it. Hunks are resolved with the tool's lenient matcher, so recoverable old_string mismatches still render. The result reports a resolution for each hunk.

The folder name of a row's project: the last segment of projectRoot, or of cwd without one (/ and \\ both separate). undefined when neither is set.

The segments cut into the lines the user typed. A flex-wrap row cannot break on a newline inside one of its items, so each line renders as its own row; an empty array is a blank line.

The setup wizard's model: React-free and Node-free, re-exported by zidane/chat/pure and zidane/chat. docs/ONBOARDING.md describes it.

Most-recent-first list of ids, deduplicated and capped.

Filters rows by where, sorts them by sort, then cuts one page. Paged results (a limit or a cursor) break ties by id, so the order is total and a cursor resumes at one place. Throws SessionQueryError for an invalid input.

Rank every item against query. Terms are AND-ed: tog sid needs both tog and sid somewhere in the text, each with its letters in order. A title that holds every term as one run gets a large bonus; description, keywords and group are fallbacks. On an empty query recents come first (most recent on top), then the catalog in provider order. Returns at most 200 rows.

The saved runs of a session, by key. What does not read as a run is left out.

The value of a field path in row (see readParsedField), or undefined for a path that names nothing.

The value of a parsed field path in row. Missing and null both read undefined; a missing boolean field reads false; environment.kind reads 'local' without an environment.

The account a session's primary model runs on, when one is pinned.

A session's stored facet values: {} when it has none or the stored value is not a record.

The raw space id stamped on a session's metadata, if any.

The recent sessions block's content.

The recent sessions block.

next, with every event structurally equal to one of prev replaced by that object (matched one to one, so an event repeated twice keeps two objects). Hands back prev itself when the result is positionally identical, so a store keyed on the array sees no change at all.

What of a model pick a window remembers as its default (lastModelChain: the next new session's model, the next launch's). One rule for the TUI and the GUI. A session on a server runs with the server's credentials, so its pickers list the server's providers: a pick there is remembered only when this computer can run its primary too (a provider only the server has would leave the next local session without a usable model), and the fallbacks this computer can't run are dropped. A pick on This computer is remembered as made. null: nothing to remember.

The template with each placeholder replaced by the row's value: text as is, numbers and booleans as written, lists joined with ', ', anything else (and missing values) empty. Runs of spaces collapse; the result is trimmed.

Move the entry with key by one priority slot; boundary or missing = identity.

Apply a user reorder of the VISIBLE strip to the pinned list. pinned can hold ids the catalog no longer shows (an extension disabled for this cwd); they keep their slots, and the visible ids fill the remaining slots in the new order, so re-enabling the extension brings its tab back where it was. Ids in visible that aren't pinned are ignored. Returns pinned itself when nothing moved.

Whether a toggle's requires setting is on. decisions is also on while approvals use Model decision (decisionsEnabled).

Time to reset, worded like a provider error's (resets in 1 h 45 min, see formatResetDelay), so the panel and a rate-limit line agree. null when the provider reported no reset, resetting once it is due.

Resolve the tab to render: the active one when it still exists, else the first tab, else null (empty catalog). Keeps state stable when an extension is disabled and its tab vanishes.

The space a session is shown in: its stamped id when that space exists, else Default (unstamped, Default itself, or a deleted space).

A bound as epoch ms: a number as is, a relative time from now, an ISO date parsed (UTC without a zone). undefined for anything else.

A run record's estimate as a RunEstimateView: its timing when it has one, else its lasting failure; undefined for neither (never asked, or a failure the next attempt may not hit).

The live work row's status: a retry wins, then a wait that has lasted RUN_WAITING_DELAY_MS, else working.

The estimates of a session's runs in progress, the session's own first: what a session row carries.

Build the receipt for a run record. now stands in for endedAt when the record is not finalized yet (run:end is broadcast a beat before it is).

The header text while a provider call is retried: Retrying 2/3, naming the subagent when it is not the session's own stream. Both apps show these words.

The status a surface times, as a key: one per retry attempt and one per wait, null while nothing is timed. A surface restarts its clock when the key changes.

The same sign-in: the failure is fixed once this one signs in again, so every earlier card for it can say so.

The mode state after text, which starts at stream offset base. Returns state itself when nothing changed.

A schedule's name for lists and the transcript: its own, else the start of its prompt, else its handler.

The session that a schedule's prompt goes to, or undefined (a newSession or an extension action).

A score (a level index, possibly between levels) as a duration: the marks are on a log scale, so a score between two levels is interpolated on it.

The error a failed script threw and the line it names, out of its result (Script error: … then the 12 | source excerpt the sandbox writes).

The loops of a script, by the lines they span, outer ones first (by start, then the longest): bodies of for, while and do, and callbacks of array methods (.map, .forEach, .filter, …). A helper function is not a loop.

The latest tool-progress state of each script, by call id (childId + NUL + callId).

The final progress of a script from its saved call records, or undefined when none carries a line (the script ran without live progress).

Key of a script in scriptProgressByCall.

The view of a script: its lines with their states, the counts, and the percent when it is exact (see countOnceCallSites).

Search-specific serializer. Unlike turnAsText, it never serializes media, document attachments, encrypted reasoning, or base64-looking payloads.

Finds the ranges of text that match query: the whole phrase, else each term, else fuzzy term matches.

Searches turns and aggregates the matches into one SessionHistoryMatch.

Searches a prebuilt index and aggregates the matches into one SessionHistoryMatch.

Like searchSessionHistoryIndex, but yields to the event loop while it searches.

Searches an index and returns the best matching turns.

Yielding matcher for very large indexes; progress is checked per chunk.

Searches turns and returns every match as a SearchSessionResult.

Builds a throwaway index over turns and searches it.

The directory name a session lives in — the same basename label the sidebar groups by. undefined for a directory-less session (the main) or no cwd.

The dimmed second line: the last thing the session said, collapsed to one paragraph and capped. Renderers may clamp further (line-clamp in the GUI, width in the TUI); this is the upper bound and the shared ellipsis rule.

The short label a transcript row shows: api-refactor finished. Status words are chosen to read as a verb phrase after the title.

The key of sessionId on hostKey (null / absent / LOCAL_HOST_KEY: a local session).

The predicate a sessions list filters by, or null (no filtering) while the feature is off. Rows with isMain (the orchestration main) belong to every space and are always kept.

Builds the sync field name for a setting.

A row's search keywords (SETTINGS_KEYWORDS), empty when it has none.

The setting a field names, if it is one.

What runs for a block, or null when nothing does (another language, no command, too long). A block with $ lines runs those lines only, without the prompt, and the lines a trailing `` continues them onto: the other lines are output.

"now" / "5m" / "3h" / "2d" / "6w": the sidebar's age words.

"Jun 12" for a YYYY-MM-DD day.

Schedules: the wire shapes, the checks and the time math that the host, the TUI and the UI share. The cron math (./cron, on croner) and the scheduler (./scheduler) are imported by path, so this entry stays free of them. See docs/SCHEDULES.md.

Six-char short form of a session id for headers and lists.

Short, stable id fragment for the header — enough to disambiguate, not to read.

Whether a snapshot should show its account name: only when another snapshot in the same list belongs to the same provider. One account is the provider.

The parent a side session answers about, or null for any other session.

End of a slice of at most size chars from from, never between the halves of a surrogate pair.

Target chars-per-second for the next smooth-mode tick given the current bucket backlog. Two linear segments (base→burst over the first 400 chars, burst→max over the rest of the ramp) so the function is continuous everywhere and plateaus (rather than cliffs) once backlog passes SMOOTH_MAX_BACKLOG_CHARS.

A sorted copy of rows. Stable: rows equal on every key keep their input order. Rows without a value for a key sort after those with one, whatever the direction. Throws SessionQueryError for an invalid sort.

The space id to stamp on a session created while activeSpaceId is shown, or undefined (feature off, or Default) — pass it to spaceMetadata.

The metadata entries stamping a session into spaceId: none for Default (an absent id IS Default), so creating in Default writes nothing new.

The default model a new session in spaceId starts on, or undefined (feature off, unknown space, or no default set) — the caller then falls back to its usual last-used model.

The space a spaceSwitchItems entry switches to; undefined for any other id.

Feature on? Off by default.

The command palette's "Switch to space" entries: one per space but the shown one. The GUI and the TUI list these same entries, and each runs them through its own switch (spaceOfSwitchItem).

The worker childId's rows belong under: its outermost ancestor that is itself one of the present workers (the ones owning rows in this view). A view that already re-based its owner's rows (a worker opened in a fold) leaves that owner out of present, so the owner's direct children become the lanes there. childId itself when no ancestor is present.

Lane resolver over one view's rows: precomputes the parent map and the present-worker set, then maps a row's childId to its lane.

Worker id → the worker that spawned it, from nested spawn markers.

A worker's start from its spawn-start marker: the task preview, and the child-N it resumes, if any (a badge). Reads spawn.resumes, or parses the resumes child-2 · <task> text of a marker recorded before it existed.

Split a command line into tokens, honoring single/double quotes so npx -y my-server --root "/path with spaces" round-trips. No escape processing beyond quote grouping — this is a form field, not a shell.

Split a string into lines preserving empty lines but dropping the implicit trailing "" produced by a final \n. Exported only for its tests — callers should use computeLineDiff.

Split a reply into what the model wrote and the stop marker at its end, if any.

Folds one output token reading into the rate state. Returns a new state, or the same one for an out-of-order or duplicate sample.

Whole tokens per second for display, or null when there is no honest reading.

text without its escapes, as a terminal would show it.

Aggregate counts for the transcript's summary badges (3 applied, 1 denied, 1 skipped). Exported so renderers don't reimplement the tally. Pure / O(n).

"Ran a command, read a file +2" for a run of calls. Adjacent identical phrases merge, so three reads in a row say "read a file" once. overflow: false leaves the +N off, for a header that counts its calls another way.

When a server last received a push, as the sync status line says it: 12s ago, 5 min ago, then a date.

The marker a server's settings UI shows next to a setting, if any.

Slice n chars off the front of buf without splitting a UTF-16 surrogate pair. Most LLM text is ASCII and never trips this; the guard matters for emoji / extended planes where halving the pair would emit an isolated half-character.

The field paths a template reads, in order, once each.

A thought's line: "Thinking" while it streams, "Thought" once settled.

Collapses runs of three or more newlines to one blank line, trims trailing whitespace and remaps the chip spans.

A title from typed text (a first message): one line, clipped with …. Null when it holds nothing.

Whether a toggle does anything under the other settings: showToolAttachments only in compact density, the only one with a "Working" row. Renderers leave the row out while it does not.

Adds entry to the end of the chain, or removes it when already present.

Why a toggle is on whatever its own value says: decisions while approvals use Model decision, which needs its answers. Renderers show the row on and inert with this note. undefined when the row's value applies.

Toggle entry: absent → appended, present → removed. Pure.

Opens the panel when closed and closes it when open.

seed selects the comparison deterministically (day-of-year keeps it stable within a day but rotating across days). null when the total is too small for the multiple to land (< 2× the work).

A machine name from a free-form one (a server's display name): the valid characters kept, the rest dashed.

The source code a tool call runs, for renderers that show it highlighted instead of the raw JSON input: Codemode's script as JavaScript, a shell / bash command as bash (the command alone: the label carries the description, the rest is plumbing). null for every other tool and for calls without the code field. language is a fence-style name both highlighters accept (Shiki in the GUI, OpenTUI's tree-sitter in the TUI).

How to install name on the host capabilities describe: without root when it can (bun's script, ripgrep's release into ~/.local/bin, Homebrew), else a root command to copy, else the tool's page. undefined for a tool that is not in HOST_TOOLS.

The overview line for one tool call.

Returns the lowercase verb phrase for a tool in a batch summary, such as read a file.

A preset by id.

The line that the palette shows once a preset applies.

The preset a set of settings matches, or custom when it matches none. Keys left out take their default value.

The preset that a row of transcriptPresetPalettePicker applies.

The palette entry for the conversation preset. It names the preset in effect, and running it opens the presets (transcriptPresetPalettePicker).

The list that the palette entry opens: the three presets, the one in effect checked.

The problems of a trigger.

Ellipsize a title to max characters; titles are user text and can be a paragraph.

The words of a tutorial on a server's settings page: the lines that name this computer name the server instead.

Returns the unique ids of the chats referenced in a reference list. References from other providers are ignored. Order of first appearance is kept.

Walk a reference list and return the deduplicated set of files in first-mention order — input to "attach these files to the prompt" downstream logic.

Walk a parsed prompt for skill references and return the deduplicated list of skill names — input to agent.activateSkill(name) calls on submit.

updated 12s ago / updated 3m ago for the newest snapshot in a list, so a refresh that returns identical numbers is still visibly a refresh. null when the list is empty.

events with event, a tool-progress state, in place of the one it replaces (same call, same subagent), or appended. Other events are appended. A script keeps one progress event, however long it runs.

The usage block's content for a summary: pure, so every host paints the same block.

The usage block's view: its tab and range, from the control values (unknown ones fall back).

The usage block.

Maps a usage percent to a UsageTone.

Every problem in a grouping: { field, labels?, order? }.

Validate + normalize the raw field values into a config entry. Returns the first problem in row order so the dialog can surface it.

Every problem in a query. Empty when value is a valid SessionQuery.

Checks a schedules.create input. Throws a ScheduleInputError that names every problem.

Checks a schedules.update patch. Throws a ScheduleInputError that names every problem.

Every problem in a sort: a list of { field, dir? }.

Every problem in a template: too long, a lone brace, or a placeholder that names no field.

The query a sidebar view stands for, to save it as a space: the space's membership (see SessionViewSpace), then the project, agent, archive, activity and extension filters, all holding, flattened. Returns the problems instead when the result is not a valid query (too deep, too many nodes or values).

Time spent waiting on the user up to now, in ms.

The saved runs with run in, without runs whose reply is gone, at most BLOCK_RUNS_MAX (the latest kept).

sort with a last { field: 'id', dir: 'asc' } key, unless it already ends on id: a total order, so a page cut by paginate resumes at one place.

Working time of a run up to now: its time since start, less its waits on the user.

Several hosts list the id (two hosts can hold the same session id: a server sharing this machine's data dir, a copied store) and the lookup did not say which: refused, never guessed.

Error thrown when a browser operation fails. code says why.

A saved host's URL answered with another host id (reinstalled, or the name now points elsewhere). The connection is closed before anything is read from it, and nothing is recorded under the saved id.

A refused add / rename; code says why.

The id lived on a host that was removed from this window.

One terminal's output by stream offset, capped, replayable into a fresh terminal with its modes.

Error thrown for a host URL a client refuses.

A schedule input that failed the checks. issues names every problem.

Thrown by the query functions for an invalid query, sort, grouping, page or cursor.

Error thrown by terminal operations. code says why.

No host told this client about the id: nothing to route it to.

The most enabled schedules the agent can have on one session.

The shortest interval for a recurring schedule that the agent makes.

How long a recurring schedule that the agent made lives, unless you pin it.

Name of a provider's API-key credential when listed next to its OAuth accounts.

The approvalMode options, with what each one does: the one list every settings UI shows.

What each approval mode gives you and guards against.

Id of the artifacts panel.

How much of a file's start planAttachment reads to tell text from binary.

callId prefix minted by the TUI's ! bash mode (see runBashMode in src/tui/commands/run-bash-mode.ts). The prefix is the durable signal — it survives persistence and replay — that a shell tool exchange was USER-initiated, so renderers can keep it visible even when the tool-display settings hide model-driven plumbing: the user typed the command and expects to see it and what it printed.

Output kept of a run (the end of it), in characters.

How long a Run here may take.

Runs kept per session; the oldest go first.

Session metadata key of the saved runs.

Default for an action's own wait (element to become actionable, text to appear). A request can raise it up to BROWSER_MAX_WAIT_MS.

Largest serialized evaluate result in bytes.

Longest wait in ms an operation may request.

How long the host waits for a page to mount and register before open gives up.

Every browser operation name.

Most elements listed in a snapshot outline.

Id of the browser panel.

Transport budget on top of the action's own wait.

Most page text characters in a snapshot.

Default wait in ms for wait_for.

Built-in panel tabs in display order.

Below this many re-read tokens a switch is cheap: short sessions show no hint.

notice.label the host stamps on prompt-cache break/expiry rows (host/notices/index.ts).

Largest calibration factor: estimates are at most tripled.

Most recent finished runs calibration reads, per project.

Smallest calibration factor: estimates are at most halved.

Fewest finished runs before calibration applies.

Trigger character that starts a chat reference.

States a plan never leaves. stopped is not one: Resume reopens it.

Trigger character — shared with skills.

The step's sections, in display order.

Every connector, in display order within each category.

What the "continue" action sends after a run that had done work.

Glyph drawn for the daemon health dot.

Tone and label to use for each DaemonHealth state.

The shared daemon runtime. ZIDANE_NO_DAEMON=1 is set by zidane --no-daemon.

The default source of extensions: the zidane repository, whose extensions/ folder holds the native extensions.

What a listing of the default source says when it cannot be read.

The model a chat runs on when nothing picks one (no session pick, no descriptor default).

The TTL a picker selects when the question suggests none.

Default value of every setting.

Id of the Default space.

Name of the Default space.

Id of the diffs panel.

The durability options, with what each one does.

Tools whose tool-result event is suppressed when showEditDiffs is on.

The modes of a fresh terminal: nothing set or reset yet.

Initial rate state, before any sample.

Most bytes one import brings in, all files together.

Most entries one import brings in.

Most bytes of one imported file (one call over a connection).

Past this, a run is in the open-ended last step: "more than 24 hours".

Without level probabilities, the over point is the estimate times this.

Percentile of the answer's distribution past which a run counts as over its estimate.

Folders of extensions/ left out of the wizard on purpose, with the reason. Every folder is either listed above or here.

An extension id, as facet paths name it. A copy of EXTENSION_ID_PATTERN (src/extensions/runtime/validate-definition.ts), kept here so this module stays a pure leaf; a test keeps the two equal.

A facet key: a letter or _, then up to 63 letters, digits, _ or -. No dots.

Every gate, so a client can report which kill switches are engaged.

Id of the files panel.

Trigger character — @ is the conventional file-mention prefix in chat UIs.

Fires kept per schedule.

Owner of the built-in blocks.

The tools probed, in the order the UIs list them. versionArgs: what prints the version, when not --version.

Id of the inbox panel.

Panel state at startup: closed, no active tab, no user picks.

Upper bound on preview lines rendered for one expanded segment. Same idea as the Logs tab's LOGS_RENDER_CAP: the scrollbox can hold arbitrary height, but pasting a runaway 64KB instruction file into the layout pass is a waste of a frame. File-backed segments have $EDITOR as the escape hatch for the remainder.

Failure reasons the next attempt would hit again: the request or the setup is wrong. A timeout, a rate limit or a network error is retried at the next trigger instead, and shows nothing.

This computer's key in every host directory (the in-process host, the embedded one, or the daemon).

This computer's label in lists and pickers.

The logo of each local model server the wizard looks for.

The name of each local model server that the wizard finds.

The cycle order the l key walks through — coarsest last so it wraps back to "everything".

Display text for each MachineLinkState.

The label of the notices a host emits about a session's machine.

Default per-chunk target for chunkMarkdownForRender. Chosen from measurement: marked's lexer (which OpenTUI's <markdown> runs internally) is O(n²) over paragraph-heavy input — ~500ms at 64KiB, ~2s at 128KiB, ~6s at 256KiB on an M-series laptop — while staying comfortably in the linear zone below ~32KiB (≤ ~200ms). Chunking a large document into ≤24KiB pieces keeps the total render linear in document size with a small constant.

The longest cursor paginate accepts. Bounds every cursor it makes: up to 9 keys of at most MAX_SORT_TEXT characters.

Cap on a PDF's or a text file's bytes.

Cap on an image's bytes. Base64 inflates ~33%, so this keeps the encoded payload under Anthropic's 5 MB image ceiling while covering essentially every screenshot. A composer scales a bigger image down to fit.

Longest name.

The largest page paginate hands out.

Faster than any served model today (the fastest wafer-scale ones run ~3k): a reading above is a glitch.

Longest prompt text.

How deep all / any / not may nest.

How many values one in / nin may list.

How many nodes (conditions and combinators) one query may hold.

How long a string value may be.

How many in / nin values one query may list in total.

Most schedules one host stores.

The most JSON one session's facet values may take together, in bytes of their JSON text.

How many keys a sort may have.

How many characters of a text value sort (and travel in a cursor).

How long a template may be.

What to do about a refused key, under the "Key refused" headline.

Shown with the MCP category: OAuth tokens stay where they were made.

The status words both apps show for a server, so the desktop app's badge and the TUI's read the same. idle says nothing.

Every McpTransport in the order a form lists them.

The media types by lowercase file extension.

The messageDelivery options, with what each one does: the one list that settings and the setup wizard show.

What each "Messages during a run" choice of Conversation gives you and guards against.

JSON-RPC "method not found": the host predates an additive method.

The shortest interval for every, and between two cron times, for a schedule that a person or an extension makes.

The name from the id alone, for a surface without the provider catalogs.

The step's sections, in display order.

Every native extension, in display order within each group.

The GitHub repository native extensions install from.

Leading glyph per severity. info is deliberately quiet: a plain i, single-width in every terminal.

The logo of each agent whose setup the wizard reads. Pi has none.

The name of each agent whose setup the wizard reads.

The step's sections, in display order.

Every feature the step can offer, in display order.

The files attached to the preview prompt: a chart of the traffic and the team's notes. The events carry their names. An app that draws them supplies the bytes, matched by name.

The preview conversation, as the transcript renders it.

Every setting the wizard can change.

The settings each step edits. Skipping a step puts them back.

Every step in flow order, with what it writes and its copy.

Version of the wizard, stored in each record it writes. A record from an older version still counts as done: existing users open new steps from Settings or the palette, never on their own.

The sentences that both apps say word for word, in English.

session.metadata.orchestratedBy — the creating orchestrator session's id, on a worker.

Caps applied by normalizePaletteItems: item count, text length and keyword count.

Group name the hosts use for the most-recently-used rows shown on an empty query.

Maximum number of recent item ids kept.

Ranked rows returned for a query; the host virtualises anyway, this bounds the work.

The logo of each built-in provider, by provider key.

Id of the queue panel: messages sent during a run, waiting for it to end (the desktop app only).

The recent sessions block's panel id.

Reminders: prompts set to reach the session at a set time (only while Settings.schedules is on).

Remote hosts (zidane-server). The setting ("Remote hosts") gates the saved servers: the Hosts settings, the per-host session groups and picking a host for a new session. Connecting through an explicit --host / ZIDANE_HOST_URL is its own opt-in, but ZIDANE_NO_REMOTE=1 still wins over it (src/tui/host-target.ts). ZIDANE_NO_REMOTE=1 is set by zidane --no-remote.

The glyph that leads the line on every surface.

Leads the agent when the run switched to it from the previous run's.

Leads the turn count: one cycle = one model round trip.

From when a status shows how long it has lasted. Shorter waits are routine (a model's first token often takes a few seconds), so they show no clock.

How long the tail must read waiting before a surface says so, so the gap between a tool call and the next turn never flashes it. The TUI and the desktop app use the same value.

The header text while the run waits on the model. Both apps show these words.

Metadata key on a session that a schedule created, and on the queued prompt of a fire.

TextOrigin.source of a fire's prompt.

Characters kept per output line.

Output lines a progress state keeps.

A secret name: an env-var style identifier, also the file name in the store.

Every SecretTtl, in the order a picker shows them.

The saved secrets of the session's project (only while Settings.secrets is on).

Image types every provider takes as they are.

Session metadata key of the pinned account (next to providerKey / modelId).

Status → how it reads. needs-attention is deliberately the loudest: it is the one status where the human, not the main, has to act.

The metadata key holding a session's facet values. Namespaced: the flat session metadata also holds keys like cwd and archived.

Every row field a query can read, with its type.

Character that separates the host key from the session id in a session key (U+001F).

session.metadata.role — present and 'main' on the one orchestrating session.

The scope of top-level runs. A sub-agent run's scope is its run id.

What the safety settings of Autonomy and Bring your setup give you and guard against.

Every option-valued setting the settings UI lists, in display order.

Every boolean setting the settings UI lists, in display order.

"Run setup", the palette entry both apps list to open the wizard again.

Longest block that runs, in characters.

Fence languages that run.

A /side chat about the session (the desktop app only).

Trigger character — slash-commands convention.

Smooth reveal rate (chars a second) while almost nothing waits.

Backlog (chars) at and past which the smooth reveal runs at SMOOTH_MAX_CPS.

Smooth reveal rate (chars a second) once the backlog reaches SMOOTH_MAX_BACKLOG_CHARS.

Session metadata key holding the space id.

Default segment bound for a growing streaming markdown event. Coalescing copies the ENTIRE accumulated string into fresh state on every drain tick, so one long streamed message costs O(L²) aggregate — past this threshold the event is split and the copy re-bounded. The split reuses chunkMarkdownForRender (blank-line boundaries outside fences, unsafe-continuation guard), so the finalized segments are byte-identical to the chunks the renderer would have produced inside a single event, and the one-row markdown/markdown margin stands in for the consumed blank line exactly like the renderer's inter-chunk gap. When no safe boundary exists (one giant fence/list) the event keeps growing until one arrives.

In the order the UIs list them.

What each category is called and carries, for both UIs.

Display label for each sync mode.

Every sync mode, in the order the UIs list them.

Sync field name of the user-scope SYSTEM.md file.

History kept per terminal for new views (characters).

Terminals a session, a project or a draft may have open at once (user terminals).

Terminals a session, a project or a draft may have open at once.

Characters a client may be behind before the host stops sending to it.

Id of the terminal panel.

Most the agent gets from one read (characters).

Plain-text lines kept per terminal for the agent.

Lines a desktop view can show (its scrollback, plus a screen of up to 1000 rows): what it asks terminal.attach for (maxLines) and what it keeps of the output it drew.

Scrollback lines a desktop terminal view keeps (each mounted view holds its own).

Glyph shown for each todo status.

Display metadata for the built-in tools, keyed by tool name.

Every setting a conversation preset sets, in the order the fine-tuning lists them.

The id of the palette entry for the conversation preset. Both apps use it.

The three presets, from the quietest to the most complete. Built by a pure call, so a bundle that never reads them drops them and the settings they read.

Post-trim size — the hysteresis gap keeps re-trims O(rare), not per-append.

Every tutorial, in the order the docs list them.

The words of each tutorial.

The usage block's panel id.

System-prompt fragment appended to every WORKER's agent. A worker's last message becomes the main's summary, so it must be a report, not a sign-off — the main should be able to act on it without reading the transcript.

Identified active trigger span. Returned by findActiveTrigger so callers can show the popover, query the provider, and on commit replace the span with the selected item's insertText.

Reads terminal output in chunks, as lines of styled spans.

Text in one style. style is absent for plain text.

How a span of text is drawn. Absent fields are the default.

A file a composer was handed, as planAttachment needs it.

One credential source detected for a provider.

The last run of a block, as the session keeps it.

What a window asks the host to run.

What changed on the page because of an action; the agent reads this instead of re-snapshotting every step.

One console message or failed request.

Input of each browser operation, keyed by operation name.

Result of each browser operation, keyed by operation name.

A browser operation sent from the agent to the process that owns the page.

A captured image of the viewport or of one element.

A readable view of the page: status, element outline and visible text.

Where the page is: its URL, title and loading state.

A target on the page: a snapshot ref (preferred), a CSS selector / text= / role= locator, or a point.

Options for buildVisibleInstructionRows.

A tab for a panel the host renders itself.

Content of the prompt-cache chip. Built by cacheBreakChip.

A model entry as displayed in the cross-provider picker.

What Settings shows about one catalog item.

Provider payload for a chat reference.

A chat that can be referenced from a prompt.

An extension-contributed / command. Provide EITHER run (one step) or steps.

One argument suggestion returned by Command.completeArgs.

Parsed /<name> [args] invocation.

Read-only view of the active prompt buffer + cursor passed to provider callbacks. Kept minimal so providers stay portable across renderers.

A discoverable, selectable thing — one row in the autocomplete popover. TItem is provider-specific; consumers can inspect data when they need the originating payload (e.g. the full SkillConfig for tooltips).

Provider contract. Implementations decide their own ranking, fuzzy-match rules, async loading behavior, and reference grammar.

The words that connectorAuthText uses. Each app writes them its own way.

A section of the Connectors step.

A service the Connectors step adds as an MCP server.

What the user typed for a connector.

Token sign-in of a connector: what to ask for and where it goes.

The shaped, render-ready breakdown.

Data exposed to extension context-breakdown contributors.

Host-supplied options for agent.getContextBreakdown.

One category row in the breakdown.

Exact per-segment counts from a provider's countTokens, when available. Each field is the count of the cumulative payload up to and including that segment, so categories are derived by differencing (see buildContextBreakdown). Any subset may be present; missing fields fall back to the heuristic for that category.

One display row contributed by an extension to the context panel.

Extension-owned context panel section. Core treats rows as opaque display data.

One expandable child row under a category (e.g. a single MCP tool / rules file).

Tools grouped by MCP server, for the expandable per-server rows.

One chunk of text present in the session's default context.

A host-discovered instruction file. Structurally compatible with AgentsMdFile from ./agents-md (whose source field is the SCOPE, not the text — contents is the text), so callers can pass those straight in.