Backend identifies which terminal app a SpawnTab call targets.

Name is the short identifier persisted on tasks.harness and used to look up an implementation.

BackgroundAgent is one entry from a harness's background-agent registry. For the claude adapter it is populated from `claude agents --json`. SessionID

BackgroundLauncher is an OPTIONAL harness capability: hosting terminal-free background sessions (claude's Agent View, via `claude --bg`). flow selects

BuildOpts are the injectable inputs to BuildStats (paths injected so tests can point at temp dirs).

Cache maps a jsonl path to its last-scanned rollup, keyed by file identity (mod time + size) so unchanged files are not rescanned.

Constants are the user-editable counterfactual multipliers. They live in ~/.flow/stats.json (JSON, not TOML — the repo has no TOML dependency).

Counts are the raw inputs to the savings model.

FileRollup is the per-jsonl scan result. It is JSON-serializable so the cache can persist it.

Format selects how a list result is serialized to the output stream.

Harness is the contract every agent-CLI adapter implements.

LaunchOpts are options forwarded into the spawn command builder. Harness adapters translate to per-CLI flags (Claude: --dangerously-skip-permissions,

Lookup is one retrieval event mined from a session jsonl.

LookupKind classifies a retrieval — a moment flow served stored context.

Owner mirrors the owners table. An owner is a durable, named, repo-scoped self-prompting controller: each tick runs a headless run and self-paces its

OwnerFilter holds optional filters for ListOwners.

Painter applies ANSI color when Enabled, otherwise returns text unchanged.

---------- playbook models ---------- Playbook mirrors the playbooks table.

PlaybookFilter holds optional filters for ListPlaybooks.

(idx_tasks_session_id is a partial UNIQUE index that requires a dedupe pass against existing data — a flat CREATE UNIQUE INDEX would fail on any DB th

ProjectFilter is the equivalent for ListProjects.

Savings are the counterfactual estimates. AddressableCount is a count, NOT time/$ — it must never be summed into TotalHours/TotalDollars. ContextToken

Stats is the fully-aggregated analytics result for one window/project.

Table is a column-aligned renderer. Columns auto-size to the widest *visible* cell — ANSI escape codes are ignored for width purposes, so a row with c

TagCount is the (tag, task-count) pair returned by ListAllTags.

Task mirrors the tasks table. ProjectSlug is nullable for floating tasks.

TaskFilter holds optional filters for ListTasks.

Usage is a token tally summed across a session's assistant turns.

WeeklyPoint is one bucket of the lookups/tokens sparkline.

Workdir mirrors the workdirs convenience registry.

bgAgentJSON mirrors one element of `claude agents --json`.

bgCap records what a stubbed BGCommandRunner saw on the last spawn/resume (non-`agents`) call.

stubClaudeRunner replaces claude.SkipPermissionsRunner with a capturing stub that returns the supplied error. The old runner had a (slug, prompt) sign

cardData is the view model handed to the HTML template.

cardFonts holds the parsed TrueType fonts, loaded once.

contentBlock represents one block in the content array.

focusFlags bundles per-backend "was called" flags so the focus routing tests can assert which backend FocusSession dispatched to without an awkward mu

jsonlMessage is the message body inside user/assistant records.

---------- jsonl record types ---------- jsonlRecord is the top-level structure of each line in a Claude session jsonl.

kbSeed describes one knowledge-base file to create on `flow init`.

kittyOSWindow / kittyTab / kittyWindow / kittyFGProc are the minimal subset of the `kitty @ ls` JSON schema we care about. The full schema has many mo

listOpts are the format/color/truncation flags every list subcommand shares.

nonBGHarness satisfies harness.Harness (via embedded interface) but NOT harness.BackgroundLauncher, so it exercises the capability gate.

paneInfo is the subset of zellij's list-panes JSON we care about. Pane records have many more fields (geometry, plugin metadata, etc.) but only id, is

---- raw jsonl shapes (only the fields we need) ----

runnerFlags bundles per-backend "was called" flags so routing tests can assert on which backend SpawnTab dispatched to without an awkward multi-return

stringSliceFlag adapts a *[]string into a flag.Value so the same flag name can be passed multiple times and accumulate. Used for repeatable `--tag` /

---------- helpers ----------

taskListRow is the row shape that feeds table, JSON, and TSV rendering for `flow list tasks`. Field order matters for JSON output stability.

versionCache is the on-disk shape of the cache file at ~/.flow/.version-cache.json.

warpStubs captures every interaction with the four mockable vars (Runner, OpenURL, WriteScript, removeScript) so tests can assert on exactly what Spaw