Configuration
All project configuration lives in one file: .flow/config.json. Read and write it through flowctl so writes stay validated and atomic:
flowctl config get <key> --json # one key (dotted path)flowctl config get --json # the whole treeflowctl config get <key> --raw --json # the file's own value: null when the key is absentflowctl config set <key> <value> # validated writeconfig get merges built-in defaults into what it returns, so an absent key reads as its default. --raw reads .flow/config.json without the defaults, which tells an unset key apart from one set to false; setup uses it to find questions it has never asked.
A missing file means defaults. An unreadable, malformed, or non-object file is reported and never reset (6.1.0+): readers warn once per process naming the file, with line and column for a JSON syntax error, config set refuses to overwrite it, and flowctl validate reports a root error. Fix the file by hand, then rerun.
Every key below comes from the published JSON Schema at plugins/flow-next/schema/flow-config.schema.json - the same schema /flow-next:setup stamps into the file’s $schema field, and the one the repo’s drift test holds flowctl to. A key flowctl accepts that the schema does not document fails the suite. This page is the human rendering of that contract.
Core pipeline
Section titled “Core pipeline”Always-relevant knobs for the default spec-to-PR flow. Every optional layer here is off or conservative by default; this page states what each key does.
| Key | Type | What it does |
|---|---|---|
review.backend | enum | Default review backend (codex, copilot, cursor, claude, host, none) or spec form backend[:model[:effort]], e.g. codex:<model>:high, or command:<label> for the argv in this machine’s FLOW_REVIEW_COMMAND_<LABEL> variable (label: lowercase letters, digits, - and _; the argv itself is never a config key - see Review command). cursor folds effort into the model name (no :effort rung); claude takes the full grammar with the CLI’s efforts (low, medium, high, xhigh, max); host and none are bare-only. copilot accepts no none/minimal effort. If unset, review commands require —review or FLOW_REVIEW_BACKEND. Default: null. |
review.fanoutExecution | enum | First-round CLI review draw scheduling: concurrent (default) or sequential. Sequential finishes each draw and its sidecars before starting the next in supplied axis order. Re-review stays single. Invalid writes and persisted values are rejected before dispatch. Default: concurrent. |
review.maxIterations | integer | Cumulative review-round cap per scope (default 8, minimum 1 - the cap can never be disabled). The env var MAX_REVIEW_ITERATIONS takes precedence over this key. In an autonomous run (flow —auto) this key may only lower the cap, so an autonomous agent cannot extend its own review gate. Default: 8. |
planSync.enabled | boolean | Enable plan-sync after task completion. Off by default; run /flow-next:sync manually when a task invalidates a downstream assumption. Default: false. |
planSync.crossSpec | boolean | Cross-spec plan-sync: scan other open specs for stale references after each task (opt-in; increases sync time). Default: false. |
memory.enabled | boolean | Enable the memory system: skills capture and search categorized learnings under .flow/memory/. Default: true. |
judge.enabled | boolean | Enable the optional judge when TYPESAFE_API_KEY is present in the environment. Default true; false disables requests. Invalid non-boolean values warn and behave as true. See Optional Jev judgments. Default: true. |
scouts.github | boolean | Enable github-scout during planning (requires the gh CLI). Default: false. |
makePr.derivedPaths | object | Optional derived-file classification rules for the make-pr export: bucket names (dualCopy, mirror, state) mapped to arrays of rules ({path|prefix, source}). A configured value fully replaces flow-next’s built-in default shapes; never required and not part of the seeded defaults. |
pipeline.qa | enum | Optional live QA stage. String-enum off | on | auto, NOT a bool; any other value, including bool true, is OFF. off: QA runs only when you invoke /flow-next:qa. on: one live /flow-next:qa pass at the all-tasks-done juncture before make-pr on every spec. auto: attended /flow-next:flow and /flow-next:flow —auto both read the key through the flow skill’s gate-selection reference and run the pass only when the spec’s acceptance describes UI behaviour on a drivable surface and a target can be started; otherwise the stage records skipped(config: pipeline.qa=auto: <reason>) and the route advances to make-pr. QA never hard-blocks a run: NEEDS_WORK and BLOCKED advance to the PR, and their findings become open items on a draft PR. flowctl stores the value and never interprets it; drivability is the skill’s judgment. Optional: flow-next runs fully without it. It costs a live-app drive pass per spec plus a running deploy and a configured driver. Values: off, on, auto. Default: "off". |
Autonomy: flow —auto & land
Section titled “Autonomy: flow —auto & land”The build loop (/flow-next:flow --auto) and the ship loop (/flow-next:land). All off/conservative by default. The pilot.* keys keep their spelling; they configure flow --auto, which replaced the /flow-next:pilot command (removed in 6.1.0).
| Key | Type | What it does |
|---|---|---|
pilot.autonomy | enum | Backlog mode for /flow-next:flow --auto (the key keeps its pilot. spelling). Scalar string-enum (ready | backlog), NOT a bool. ready = flow --auto selects only already-ready specs. Only the literal backlog widens selection to the whole open backlog; the per-run --backlog flag forces it for one run; any other value stays ready. In long-horizon mode a backlog run drives its one selected item to a terminal and stops, and the next invocation selects the next item. Backlog mode never authors a spec or sets ready. It grants no merge authority by itself; --until=merge separately authorizes land for the selected item. Values: ready, backlog. Default: "ready". |
pilot.gateClasses | enum | Backlog-mode force-gate for flow --auto --backlog (the key keeps its pilot. spelling): class names (e.g. risky, prod-config) that force surfacing before action - a matching item is parked with a question (ASKED) instead of advanced full-auto. Empty = full-auto for every workable item. Default: []. |
land.* - settings for /flow-next:land, which handles one named pull request. Two keys remain.
| Key | Type | What it does |
|---|---|---|
land.patienceMinutes | integer | Minutes since the author’s last push (land’s catch-up does not count) that land waits before merging, so review bots can post, when the calling flow, not a human in the session, authorized the merge. The window is for bots, not people; raise it if your bots are slower. A human’s current merge authorization waives the wait, and a goal’s child pull request into its integration branch skips it. When GitHub reports no push time, land uses the head commit’s earliest check-suite creation time, then its committer date. Default: 10. |
land.mergeVerdictCommand | string | null | Opt-in repo merge gate: a shell command land runs once per invocation, only after every other merge gate passes, with a 600-second bound. Exit 0 allows the merge; any non-zero exit, a missing or unexecutable command, or a timeout blocks it with NEEDS_HUMAN. It runs from the repository where land was invoked without changing the checkout, so it must judge the remote FLOW_HEAD_SHA, not local HEAD. The environment also supplies FLOW_BASE_REF, FLOW_PR_NUMBER, FLOW_SPEC_ID (empty when several specs match) and space-separated FLOW_SPEC_IDS. Never executed under --dry-run. Unset, null, and an empty string all mean off. Default: "". |
Retired land.* keys. A config file that still carries them loads normally. Land prints one notice naming the ignored keys and leaves the file unchanged; delete them when convenient.
| Retired key | Use instead |
|---|---|
land.release | Run releases separately from your project’s release instructions. |
land.reviewSignal, land.automatedReviewers, land.cleanReviewCommentPattern | The default gate is green checks, a nonblocking GitHub review decision, and zero unresolved threads. Tighten it in your instruction file, with branch protection, or with land.mergeVerdictCommand. See the merge gate. |
land.reviewTrigger, land.requestReviewers | Request reviewers outside land. |
land.ciFixBudget | Land makes one focused fix or one flake rerun per failure and reports BLOCKED naming the check. It keeps no ledger and sets no label. |
land.patienceMinutesAfterReview | land.patienceMinutes, anchored to the last push. |
Removed in 8.0.0: the rp review backend
Section titled “Removed in 8.0.0: the rp review backend”rp (RepoPrompt) is gone from the accepted review.backend values. A config that still reads rp keeps loading: review prints a one-line notice and runs as if no reviewer were configured, and config set refuses the value. Set another reviewer with flowctl config set review.backend codex (or host, claude, copilot, cursor). review.backend itself stays; 7.1.0 announced its removal for 8.0.0, and that was reversed. See Review backends.
Removed in 7.0.0: pipeline.chainStages
Section titled “Removed in 7.0.0: pipeline.chainStages”pipeline.chainStages is gone from the schema. A config that still sets it keeps working: flowctl ignores the key and prints a one-line note. Under flow --auto --tick, make-pr now runs on the next tick; a long-horizon flow --auto run already runs QA and then make-pr as consecutive hops. Delete the key from .flow/config.json when convenient.
Removed in 7.0.0: the HTML render lenses
Section titled “Removed in 7.0.0: the HTML render lenses”artifacts.html.enabled is gone from the schema. A config that still sets it keeps working: flowctl ignores the key and prints a one-line note. For a visual view of a spec, a plan or a diff, run /flow-next:visual, or ask the agent for an HTML page.
Removed in 4.0.0: routing moved into prose
Section titled “Removed in 4.0.0: routing moved into prose”The work.delegate* keys (packaged codex delegation) and the models.* block (the role map and its staleness stamps) are gone from the schema and from flowctl. The removed keys are work.delegate, work.delegateConsent, work.delegateDecision, work.delegateEffort, work.delegateModel, work.delegateSandbox, models.roles, models.verifiedAt and models.verifiedWith. Routing lives in the routing block in your CLAUDE.md / AGENTS.md instead, and implementation offload is the implementer tier plus a bridge recipe.
Leftover keys are inert, not dangerous. A .flow/config.json still carrying them keeps working: flowctl ignores them entirely and prints one non-blocking advisory naming what it found, on the config surfaces and the work entry points, to stderr so a --json read stays parseable. Delete them when convenient.
Chart (pre-capture discovery)
Section titled “Chart (pre-capture discovery)”Bounds for the optional decision-map stage.
| Key | Type | What it does |
|---|---|---|
chart.maxDecisions | integer | Charting-time decision ceiling (default 12). chart create with an initial-map refuses past this count without —force-size —reason. Later sharpening may grow past it. Default: 12. |
chart.claimStaleAfter | number | Stale-claim age threshold in hours (default 24). release-claim —break-stale —reason is allowed only after a claim is at least this old; always audited (actor, prior owner, age, reason). Default: 24. |
Feature map
Section titled “Feature map”The due threshold for the optional feature map.
| Key | Type | What it does |
|---|---|---|
features.staleAfterCommits | integer | Due threshold (default 50). flowctl features status reports the feature map due a maintain pass when a feature file’s **Last proven:** commit is at least this many default-branch commits old, counting only commits that change a file outside .flow/ and documentation. Values below 1 or non-integers read as the default. See Keep the feature map current. Default: 50. |
Regulated environments
Section titled “Regulated environments”Opt-in development evidence and spec approval for teams that must show what was built, tested, reviewed and approved. All three keys are absent by default, and absent changes nothing. See Flow-Next in regulated environments.
| Key | Type | What it does |
|---|---|---|
evidence.persist | enum | off | on. Only the literal on activates: flowctl then writes committed copies of every review and QA receipt (one file per round), QA captures, each task’s recorded evidence with its test runner’s report, and a run record into <evidence.dir>/<spec-id>/ and stages them. Any other value, including bool true, is off. Values: off, on. Default: off (absent). |
evidence.dir | string | Repo-relative root of the per-spec evidence folders. A path outside the repository or through a symlink is refused when evidence is written. Default: .flow/evidence (absent). |
spec.approval | enum | local | pr. local keeps spec readiness as it is. pr makes a spec ready only once its spec PR (head branch spec/<spec-id>) merges with at least one approving review; the approval is recorded with the spec. An invalid value stops instead of falling back to local. Values: local, pr. Default: local (absent). |
Tracker bridge
Section titled “Tracker bridge”Projection to Linear / GitHub / GitLab / Jira. Inactive unless enabled.
tracker.* - Tracker-sync bridge settings (Linear / GitHub / GitLab / Jira). See Tracker sync.
| Key | Type | What it does |
|---|---|---|
tracker.version | integer | Tracker config schema version. Default: 1. (machine-written) |
tracker.enabled | boolean | Enable the tracker-sync bridge. The bridge is active iff raw tracker.enabled == true OR raw tracker.type is one of linear/github/gitlab/jira. Optional: flow-next runs fully without it. It costs a bidirectional round-trip per enabled lifecycle event plus a conflict policy and a second place state can be wrong; enable it when other people need to read or edit status where they already work, or run /flow-next:tracker-sync manually and leave the bridge off in between. Spec-only is a first-class mode. See Tracker sync. Default: false. |
tracker.type | enum | Tracker backend: linear, github, gitlab, or jira. Values: linear, github, gitlab, jira (or null). Default: null. |
tracker.provenance | string | null | Free-form provenance written by the discovery ceremony on confirmation (who/when/signals). Default: null. (machine-written) |
tracker.perEvent.capture | enum | Sync op fired when a spec is captured: off | pull | push | reconcile | comment. Values: off, pull, push, reconcile, comment. Default: "off". |
tracker.perEvent.interview | enum | Sync op fired after a refine pass updates a spec (the key keeps its historical name): off | pull | push | reconcile | comment. Values: off, pull, push, reconcile, comment. Default: "off". |
tracker.perEvent.plan | enum | Sync op fired after planning: off | pull | push | reconcile | comment. Values: off, pull, push, reconcile, comment. Default: "off". |
tracker.perEvent.work.firstClaim | enum | Sync op fired on a task’s first claim: off | pull | push | reconcile | comment. Values: off, pull, push, reconcile, comment. Default: "off". |
tracker.perEvent.work.done | enum | Sync op fired when a task completes: off | pull | push | reconcile | comment. Values: off, pull, push, reconcile, comment. Default: "off". |
tracker.perEvent.makePr | enum | Sync op fired when make-pr opens a PR: off | pull | push | reconcile | comment. The PR link + In Review push is unconditional whenever the bridge is active. Values: off, pull, push, reconcile, comment. Default: "off". |
tracker.perEvent.resolvePr | enum | Sync op fired after resolve-pr: off | pull | push | reconcile | comment. Values: off, pull, push, reconcile, comment. Default: "off". |
tracker.perEvent.completionReview | enum | Sync op fired after the spec completion review: off | pull | push | reconcile | comment. The ceremony seeds comment (verdict + R-ID coverage; never terminal Done). Values: off, pull, push, reconcile, comment. Default: "off". |
tracker.perEvent.qa | enum | Post the /flow-next:qa ship verdict as a tracker comment: off | comment only. comment is the only sensible verb for a verdict; the QA skill treats any non-off value as comment. Not switched on by the ceremony’s default-on set - QA-specific opt-in. Values: off, comment. Default: "off". |
tracker.perEvent.land.merged | enum | Post-merge touchpoint for /flow-next:land. After a confirmed merge, land moves each matching spec’s issue to its terminal status through the tracker API whenever the bridge is active; this leaf does not gate that status write. Land writes no local sync receipt, timestamp, or verdict comment. A failed touchpoint keeps MERGED and names the merge commit; running land again on the merged pull request retries only the touchpoint. A closed spec on an open pull request stays In Review. Values: off, pull, push, reconcile, comment. Default: "off". |
tracker.charts | enum | Optional chart lifecycle projection. String-enum off|on, NOT a bool: only the literal on projects charts as parent issues with decision children through the tracker facade. Local chart operations always succeed when off or when the bridge is inactive. Values: off, on. Default: "off". |
tracker.perTracker.teamId | string | null | Linear team id. Default: null. |
tracker.perTracker.projectId | string | null | Linear project id. Default: null. |
tracker.perTracker.labelMap | object | Label linkage map (tracker-specific shape). Default: {}. |
tracker.perTracker.priorityMap | object | Priority linkage map (tracker-specific shape). Default: {}. |
tracker.perTracker.repo | string | null | GitHub repo as owner/name, written by the discovery ceremony (machine-written; not part of the seeded defaults). (machine-written) |
tracker.perTracker.project | string | null | GitLab group/subgroup/project path (URL-encoded once for the API, never double-encoded). Default: null. |
tracker.perTracker.host | string | null | Self-managed GitLab base URL. null resolves from glab config / CI_SERVER_URL; gitlab.com is never assumed. Default: null. |
tracker.perTracker.baseUrl | string | null | Jira site base URL (Cloud or DC/Server). The JIRA_BASE_URL env var overrides it at runtime. Default: null. |
tracker.perTracker.projectKey | string | null | Jira project key (the JQL / listOpenIssues scope). Default: null. |
tracker.perTracker.authScheme | enum | Jira auth shape decided once at the discovery ceremony: cloud-basic (Cloud HTTP-basic email:API_TOKEN) or bearer-pat (DC/Server bearer PAT). Credentials still read from env each run, never stored here. Values: cloud-basic, bearer-pat (or null). Default: null. |
tracker.perTracker.apiVersion | integer | null | Jira REST API version. null until the resolver pins 2; migration converges a legacy 3 to 2. Default: null. |
tracker.perTracker.owner | string | null | GitHub repository owner (discovery-fingerprint input; dynamic per.get read via _FINGERPRINT_KEYS). |
tracker.perTracker.issueType | string | integer | null | Jira issue type (name or id) for created issues; a configured value that does not resolve against the live project is an error. |
tracker.perTracker.blocksLinkType | string | null | Jira issue link type used for blocks relations (e.g. Blocks). When unset, flowctl uses the first link type on the site whose outward phrase contains “block”, and refuses with a capability error when there is none. Probe and mutation use the same resolved name. |
tracker.perTracker.preferredTransport | string | null | Linear transport preference (mcp routes through the MCP continuation; anything else uses HTTP). |
tracker.perTracker.transport | string | null | Legacy alias for preferredTransport (read second). |
tracker.perTracker.sslVerify | boolean | Verify TLS certificates against Jira. false is an explicit opt-out for a self-hosted internal-CA / self-signed cert (JIRA_SSL_VERIFY env overrides). Default: true. |
tracker.perTracker.tokenCommand | array | null | Argv list, run without a shell, that prints the Linear API key, Jira API token or Jira PAT when LINEAR_API_KEY / JIRA_API_TOKEN / JIRA_PAT is unset. Runs at most once per flowctl process with a 10-second timeout; its output is never written anywhere. On one machine, FLOW_TRACKER_TOKEN_COMMAND_LINEAR or FLOW_TRACKER_TOKEN_COMMAND_JIRA holds the same JSON argv and takes its place without committing a local path. See Credentials from a secret store. Default: unset. |
tracker.perTracker.ideaProjectKeys | array | null | Jira Product Discovery project keys whose issues are ideas, e.g. ["IDEAS"]. Absent or empty turns idea intake off; set, attended flow can start from an idea and the tracker idea verbs run. Jira Cloud only. See Idea intake. Default: unset. |
tracker.perTracker.statusMap | object | Legacy normalized-status to Jira status map ({name}/{id}; id preferred - names are project-renamable). Live entries migrate into tracker.resolved.destination.statusIds; dead entries are dropped with a warning. Default: {}. |
tracker.staleAfterHours | integer | Staleness threshold (hours) consumed by sync list-stale. Default: 24. |
tracker.conflictTiebreak | enum | Status who-wins tiebreak: flow-wins | tracker-wins | always-ask. Strict enum: invalid CLI writes are rejected; malformed persisted values fail before status work. In autonomous mode always-ask resolves to queue, not prompt. Values: always-ask, flow-wins, tracker-wins. Default: "always-ask". |
tracker.readyState | string | null | Readiness projection: the tracker workflow state meaning ready for work (a Linear state name, Jira status name, or a GitHub/GitLab label). When set, pull-side sync projects it onto the local spec ready flag - one-way, tracker is authoritative. null = projection off. Default: null. |
tracker.specWins | boolean | true | false. Only the literal true activates: work’s first claim also writes the rendered spec as the issue body, and make-pr’s reconcile sends the rendered spec as the final body, so tracker-side body edits do not survive. Any other value is off. Default: false (absent). |
tracker.specIds | enum | Id scheme for new specs when a tracker bridge is active: flow (native fn-N) or tracker (tracker-keyed KEY-N-slug / synthetic gh-N / gl-N). Strict enum on write; malformed on-disk values fail closed to flow. Not materialized at init so setup can detect never-asked via a raw null read. Values: flow, tracker. Default: "flow". |
tracker.transport.timeoutS | number | null | Per-request timeout in seconds (0-600). |
tracker.transport.maxRetries | integer | null | Retry attempts per call. |
tracker.transport.backoffCapS | number | null | Backoff cap in seconds. |
tracker.transport.concurrency | integer | null | Max concurrent tracker calls. |
tracker.resolved.destination.statusIds | object | Normalized status slots (todo, in_progress, done; optional provider slots) mapped to provider status ids. (machine-written) |
tracker.resolved.destination.stateIds | object | Normalized state slots mapped to provider state ids (Linear). (machine-written) |
tracker.resolved.capabilities._source | object | Machine-written capability provenance (which probe/endpoint established each flag); GitLab’s resolver persists it alongside the boolean capability keys. (machine-written) |
tracker.resolved.scopeResolvedAt.destination | string | ISO timestamp of the last successful destination resolution. (machine-written) |
tracker.resolved.scopeResolvedAt.destination.statusIds | string | ISO timestamp of the last successful destination.statusIds resolution. (machine-written) |
tracker.resolved.scopeResolvedAt.destination.stateIds | string | ISO timestamp of the last successful destination.stateIds resolution. (machine-written) |
tracker.resolved.scopeResolvedAt.capabilities | string | ISO timestamp of the last successful capabilities resolution. (machine-written) |
tracker.resolved.resolvedAt | string | null | Non-null only when all required destination fields, required normalized slots, and capability booleans are present. (machine-written) |
Environment overrides
Section titled “Environment overrides”A few knobs are runtime environment variables rather than config keys:
| Variable | Effect |
|---|---|
FLOW_ACTOR | Actor identity for claims/receipts. Wins over git email, git name, $USER. |
FLOW_REVIEW_BACKEND | Per-invocation review backend override (bare or backend:model:effort spec form). |
MAX_REVIEW_ITERATIONS | Session override for review.maxIterations (the config key is the durable form). |
MAX_REVIEW_TRANSPORT_FAILURES | How many consecutive review attempts may end with no verdict (a backend that crashed, timed out or printed nothing) before the review stops with TRANSPORT_UNHEALTHY (exit 5). Default 2. This is separate from the round cap. |
FLOW_REVIEW_DEEP=1 | Runs implementation review as if --deep were passed: an adversarial pass after the primary review, plus security and performance passes when the changed files call for them. |
FLOW_VALIDATE_REVIEW=1 | Runs implementation review as if --validate were passed: on NEEDS_WORK, a validator re-checks each finding against the code and drops clear false positives. --interactive has no environment form. |
CODEX_SANDBOX | Sandbox for the codex review backend when --sandbox is auto or absent: read-only, workspace-write, danger-full-access or auto. auto resolves to read-only on macOS and Linux and danger-full-access on Windows, where Codex’s read-only sandbox blocks reads too. Widen it only on Windows; a review that needs to write is a prompt bug. |
FLOW_PR_CREATE_CMD | make-pr’s PR-create seam for App/bot-authored PRs. |
GITLAB_HOST | <host>:<port> for an SSH GitLab origin whose web port is not the default; make-pr, resolve-pr and land then reach the instance through glab on that port. |
GH_ENTERPRISE_TOKEN / GITHUB_ENTERPRISE_TOKEN | With either set, gh answers for every host, so forge detection asks GitLab first about an origin host it does not already know, and treats the host as GitHub only when GitLab does not claim it. |
FLOW_AUTONOMOUS=1 | Question-suppression for autonomous drivers. |
TYPESAFE_API_KEY | Turns on the optional Jev judgments; read at call time only, never written to config or receipts. judge.enabled=false forces them off. |
JIRA_BASE_URL / tracker tokens | Tracker transport credentials and endpoint overrides. |
FLOW_TRACKER_TOKEN_COMMAND_LINEAR / FLOW_TRACKER_TOKEN_COMMAND_JIRA | A JSON argv that prints the Linear or Jira token, read in place of tracker.perTracker.tokenCommand so a path that exists on one machine is never committed. The token variable still wins, and a failing command does not fall back to the config key. See Credentials from a secret store. |
FLOW_NO_DEPRECATION=1 | Silence deprecation notices. |
FLOW_STATE_DIR | Directory for runtime task-claim state, in place of flow-state/ under the git common dir that every worktree of a repo shares. Set it per process when two agents drive flowctl in sibling worktrees of one repo and must not see each other’s claims. Keep it outside the repository tree. |
FLOW_NEXT_FLOWCTL | Path to another flowctl (your own build, a port, a fork under test). The plugin’s launchers hand every call to it, before looking for Python, and return its exit code. A path the system cannot run stops with an error; it never falls back to the bundled flowctl. A path to the launcher itself runs it as usual. Point it at the other flowctl directly, not at a wrapper that calls the plugin’s launcher. |
Precedence where both exist: explicit CLI flag → environment → .flow/config.json → built-in default.
Team defaults
Section titled “Team defaults”Set policy where the team can inspect it - the config file is committed, so flowctl config set IS the policy record. The decisions worth making explicitly as a team: the review backend (cross-family beats same-family), review.maxIterations, whether planSync.crossSpec is worth the reconciliation time at your spec volume, the tracker perEvent map (which lifecycle moments project), and the autonomy posture (pilot.autonomy, and land.mergeVerdictCommand on repos with no branch protection). Everything else has a sane default.