Potpie CLI Flow & Command Reference

July 28, 2026 · View on GitHub

Status: reflects code on main @ 49e12528, last reviewed 2026-07-02.

This is the command reference for the potpie CLI — the full, grouped surface with flags, the shared plumbing every command goes through, and the canonical journey. It is the one doc that may restate flags in full; conceptual depth lives in the sibling docs linked under each group and in See also.

This file lives at repo-root docs/context-graph/cli-flow.md. The code it describes lives under repo-root potpie/context-engine/ (paths below are relative to that root).

One CLI for humans and agents

There is no separate human-vs-agent API. Both people and coding harnesses drive the same potpie CLI. potpie/cli/main.py build_app() is the single console entrypoint ([project.scripts]): one Typer root whose @app.callback exposes three global options, with the rest of the surface assembled from top-level registrars and add_typer sub-apps. Every command routes CLI → HostShell → service(s) → ports.

flowchart LR
  cf_user["user / agent"]
  cf_cli["potpie CLI<br/>main.py build_app()"]
  cf_common["commands/_common.py<br/>get_host • contract() • resolve_pot_id"]
  cf_shell["HostShell facade<br/>(daemon-backed RemoteHostShell by default)"]
  cf_svc["services: pots • graph • graph_workbench<br/>skills • daemon • ledger • nudge • backend"]
  cf_ports["GraphBackend + capability ports"]

  cf_user --> cf_cli --> cf_common --> cf_shell --> cf_svc --> cf_ports

Global options (root `@app.callback`)

Option	Effect
`--json`	machine-readable output for scripts/agents (stable, additive fields)
`--verbose` / `-v`	verbose diagnostics
`--version`	print version and exit

Shared plumbing (`commands/_common.py`)

Every command is wrapped by the same three helpers, so the contract below holds uniformly across the surface.

get_host() — returns a daemon-backed RemoteHostShell by default (the shipped CLI default host mode is a detached daemon). Only CONTEXT_ENGINE_HOST_MODE=in_process builds an in-process shell via bootstrap/host_wiring.py build_host_shell(). A handful of commands bypass the HostShell entirely (see notes on login and cloud).
contract() — the error boundary that maps outcomes to exit codes and emits structured JSON errors (code, message, detail, recommended_next_action):

Exit Meaning
0 success
1 command / validation failure
2 daemon / API / dependency unavailable (incl. CapabilityNotImplemented, ContextEngineDisabled)
3 partial / degraded result
4 auth / permission failure
resolve_pot_id(...) — pot-scope resolver, precedence: explicit --pot > repo-default binding > registered-repo match (active pot wins ties, else ambiguous_pot) > active pot, else no_active_pot. source add passes infer_from_repo=False (registration never infers a pot).

Exit	Meaning
`0`	success
`1`	command / validation failure
`2`	daemon / API / dependency unavailable (incl. `CapabilityNotImplemented`, `ContextEngineDisabled`)
`3`	partial / degraded result
`4`	auth / permission failure

emit()/fail() render the human and --json shapes. All commands support human output by default and --json for scripts/agents.

Command groups & code slots

Surface assembled via top-level registrars (query, bootstrap, auth, ui) and add_typer sub-apps. Note the corrections vs older docs: there is no commands/backend.py — the backend and timeline apps are defined in commands/graph.py; and there is no graph admin command group.

Group / commands	Code slot	Routes to
`resolve` `search` `record` `status`	`commands/query.py`	`HostShell.agent_context` (V1 4-tool wrappers; `status` = host/pot readiness)
`setup` `doctor` `whoami` `use` `config`	`commands/bootstrap.py`	`HostShell` (setup bootstraps daemon → `SetupOrchestrator`)
`login` `logout` + provider groups (`github`/`git`/`linear`/`jira`/`confluence`/`auth`)	`commands/auth.py`	local Firebase/API-key auth + integration read clients (not via HostShell)
`ui`	`commands/ui.py`	ensures daemon, opens read-only graph explorer
`pot` `source`	`commands/pots.py`	`HostShell.pots` (`PotManagementService`)
`daemon`	`commands/daemon.py`	`HostShell.daemon` (`Daemon`)
`ledger`	`commands/ledger.py`	`HostShell.ledger` (clients are stubs — roadmap)
`graph` (+ nested `inbox`, `quality`, `bulk`)	`commands/graph.py`	`HostShell.graph` / `graph_workbench` / `backend` / `nudge`
`timeline`	`commands/graph.py`	`HostShell.graph` (alias of a recent-changes read)
`backend`	`commands/graph.py`	`HostShell.backend` (`GraphBackend`)
`skills`	`commands/skills.py`	`HostShell.skills` (`SkillManager`)
`cloud`	`commands/cloud.py`	managed sync — all raise `CapabilityNotImplemented` (roadmap)

The async ingestion pipeline behind the HTTP API keeps a separate composition root (bootstrap/ingestion_server.py, default backend neo4j) — see architecture.md and ingestion-nudge.md.

Top-level commands

The V1 agent wrappers (resolve/search/record) and status ride the same graph internals as the workbench; they are not a "legacy V1 surface waiting on V2."

potpie resolve <task> [--intent feature] [--include <csv>] [--mode fast|balanced|verify|deep] [--pot <ref>]
potpie search  <query> [--include <csv>] [--pot <ref>]
potpie record  --type <kind> --summary <text> [--scope <k:v>] [--pot <ref>]
potpie status  [--intent <name>] [--harness claude] [--pot <ref>]

potpie setup   [--repo .] [--pot default] [--agent claude] [--backend <profile>] \
               [--scan] [--dry-run] [--yes/-y] [--daemon | --in-process]
potpie doctor
potpie whoami
potpie use     <ref> [--local | --managed]
potpie config  get <key>
potpie config  set <key> <value>
potpie login   [--api-key/-k <key>] [--url/-u <url>]
potpie logout
potpie ui      [--open/--no-open] [--pot <ref>]

resolve / search / record → host.agent_context.{resolve,search,record}. record --type accepts the structured record types (preference/policy/bug_pattern/ fix/verification/decision) plus free-form; it goes through semantic validation and the record→semantic bridge (writing.md). --mode/--include ride in metadata only — they do not change the read path in V1.5 (querying.md).
status — host/pot readiness (agent_context.status): daemon, backend, pot, and skill state. --host is a deprecated no-op (readiness is the default). --verify is rejected here — it moved to potpie auth status --verify, the explicit integration-auth report.
setup — idempotent first-run that builds a SetupPlan (config/storage/daemon/active default pot/source registration/skills). --backend picks the GraphBackend profile (default falkordb_lite); --scan is an opt-in working-tree scan (default off); --daemon/--in-process selects host mode (daemon mode calls host.daemon.ensure() first); --dry-run returns a preview without executing. --pot only overrides the initial pot name.
doctor — local diagnostics composed from backend.capabilities() + backend.mutation.readiness() + daemon.status() + ledger.status(); also reports effective_current_repo_pot and repo_default_pot (the repo→pot routing resolution for the current directory).
whoami — local OSS reports a none identity.
use <ref> — alias for pot use. --managed raises CapabilityNotImplemented (see Roadmap below).
login / logout — Firebase browser session or API-key store; not routed through HostShell. Flags are --api-key/-k and --url/-u.
ui — ensures the daemon, discovers its base_url, opens <base>/ui, the read-only graph explorer.

Roadmap (not yet wired): use --managed raises CapabilityNotImplemented. Managed-backend routing is designed and wired-for but not functional today.

Provider auth groups (root, `commands/auth.py`)

These manage credentials for the agent's own integration reads (jira/linear are a CLI + agent flow, not Potpie connectors — see ingestion-nudge.md).

potpie github  login | logout | repos
potpie linear  login | logout | ls | select
potpie jira     login | logout | ls | select
potpie confluence login | logout | ls | select
potpie auth     status [--verify] | logout # integration auth report + logout

git is a hidden alias group. auth status [--verify] is the explicit local integration-auth report (--verify runs a lightweight API check); the rest of the auth group is deprecated (logout + hidden revoke + provider mirrors).

Pots & sources (`commands/pots.py` → `host.pots`)

A Pot is the unit of tenancy/isolation; the pot id is the storage group_id. Local setup creates and activates a default pot.

potpie pot list [--local | --managed | --all]
potpie pot info
potpie pot create <name> [--repo .] [--use] [--also-default-for-current-repo]
potpie pot use    <ref> [--also-default-for-current-repo]
potpie pot rename <ref> <new-name>
potpie pot reset  [<ref>] [--confirm]
potpie pot archive <ref>

potpie pot linked  [--repo .] [--summary]
potpie pot default show | set | clear [--repo .]

potpie source add    <kind> <location> [--name <n>] [--pot <ref>] [--default/--no-default]
potpie source list   [--pot <ref>]
potpie source status [<id>] [--pot <ref>]
potpie source remove <id> [--pot <ref>]

pot reset is the destructive per-pot wipe — note there is no graph reset command.
pot linked / pot default manage the repo→pot binding consumed by resolve_pot_id. pot linked --summary skips per-pot graph counts for a faster repo-routing summary. pot create/pot use --also-default-for-current-repo set the repo's default binding in the same step (otherwise the CLI warns when the repo default and the selected pot diverge).
source status with no ID prints a per-pot summary of all sources; with an ID it reports that single source.
source add <kind> <location> is generic registration only (no scan/ingest); registering a repo also sets the repo default. Repo-baseline ingestion is harness-led via skills (skills.md), not a scanner.

Daemon (local infra)

potpie daemon start | status | logs [--follow] | restart | stop

daemon (commands/daemon.py → host.daemon) — local recovery tooling, not onboarding steps. DaemonStartError → exit 2.
Supporting-service admin CLI (potpie service …) is not part of the OSS surface; the detached daemon does not expose a compatible /admin/services discovery contract for those commands.

Event Ledger (`commands/ledger.py` → `host.ledger`)

potpie ledger status
potpie ledger sources list [--pot <ref>]
potpie ledger query  [--source <id>] [--type <kind>] [--since <time>] [--until <time>] [--limit 100] [--pot <ref>]
potpie ledger use    managed [--org <id>] | self-hosted <url> [--org <id>]
potpie ledger pull   --source <id> [--filter <expr>] [--pot <ref>]
potpie ledger disconnect

The external Event Ledger is a separate managed-or-self-hostable source-event service that the graph pulls from (it is never the graph's source of truth). ledger query is read-only history (no cursor advance); ledger pull advances the per-(pot,source) cursor. --filter is reserved/unused; ledger use writes config (runtime rebinding is roadmap).

Roadmap (not yet wired): the external ledger clients (adapters/outbound/ledger/managed_client.py, self_hosted_client.py) are TODO stubs — ledger pull/query/status exist but are non-functional against any real provider today. Do not confuse this with the internal Postgres event store (the live "ledger", lifecycle queued/processing/done/error) described in ingestion-nudge.md.

Cloud (`commands/cloud.py`)

potpie cloud login | status | push [--pot <ref>] | pull [--pot <ref>]
potpie cloud skills sync [--agent <id>]

Roadmap (not yet wired): every cloud command (and pot list --managed, use --managed) raises CapabilityNotImplemented. The managed profile shares the same service modules and command language; only the routing is unbuilt.

Skills (`commands/skills.py` → `host.skills`)

potpie skills list
potpie skills install [<id>]   [--agent claude|codex|cursor|opencode] [--scope global|project] [--path .]
potpie skills update  [--all]  [--agent ...] [--scope global|project] [--path .]
potpie skills remove  [<id> | --all] [--agent ...] [--scope global|project] [--path .]
potpie skills status           [--agent ...] [--scope global|project] [--path .]
potpie skills add     <source>            # TODO stub

Skills are CLI-managed instruction bundles that teach the harness how to drive the workbench. There is no top-level potpie install — skills install via potpie skills install. Scope flips to project automatically when --path is given with global. Agents only ever see an advisory install nudge in context_status. The full catalog, per-harness install paths, the correctness gate, and the (separate) server-side reconciliation skill surface are documented in skills.md.

Graph workbench (`commands/graph.py`) — the core surface

The potpie graph … workbench is shipped today as V1.5. Three Typer apps are defined here and mounted at root: graph (with nested inbox, quality, bulk), plus top-level timeline and backend. Each graph command runs inside _graph_command(name), wrapping contract() with the richer workbench envelope (graph_success/error/not_implemented_envelope carrying request_id, subgraph_versions, warnings, unsupported) and emitting OTLP spans + metrics + usage events.

doctor is local-profile diagnostics: daemon/backend readiness, CLI install facts (uv tool env, PATH, python shebang), and recommended follow-up commands. Do not use python -m pip show potpie-context-engine for local dev installs — the package lives in the uv tool environment. Prefer uv tool list, which -a potpie, make cli-status, and make cli-install for repo-local reinstalls (UI build + daemon stop + editable install).

uv tool list
which -a potpie
head -n 1 "$(command -v potpie)"
make cli-status
make cli-install   # repo-local reinstall only
potpie doctor
potpie --json doctor

Read / contract (route `host.graph`)

potpie graph status [--pot <ref>]

potpie graph catalog [--task <text>] [--subgraph <s>] [--profile full|read] [--format auto|table] [--pot <ref>]

potpie graph describe [<subgraph>] [--view <v>] [--examples] [--pot <ref>]

potpie graph read --subgraph <s> --view <v> \
  [--query <text>] [--query-threshold 0.70] [--scope <k:v,...>] [--repo <r>] \
  [--since <t>] [--until <t>] [--time-window/--window <dur>] \
  [--environment <env>] [--source-ref <ref> ...] \
  [--depth <n>] [--direction out|in|both] [--limit 12] \
  [--sort auto|score|occurred_at] [--dedupe auto|none|source_ref|activity] \
  [--format auto|raw|events|table|jsonl] [--detail compact|full] [--relations summary|full] \
  [--current] [--pot <ref>]

potpie timeline recent \
  [--query <text>] [--query-threshold 0.70] [--since <t>] [--until <t>] [--time-window/--window <dur>] \
  [--service <svc>] [--limit 12] [--format ...] [--detail ...] [--relations ...] [--pot <ref>]

potpie graph search-entities [<query> | --query <text>] \
  [--type <label>] [--predicate <p>] [--subgraph <s>] [--scope <k:v>] [--truth <class>] \
  [--source-system <sys>] [--source-family <fam>] [--since <t>] [--until <t>] \
  [--environment <env>] [--external-id <id>] [--source-ref <ref> ...] \
  [--limit 10] [--supporting-claims 0] [--pot <ref>]

graph catalog returns the live contract (versions, commands, 7 truth classes, the 10 mutation ops — all APPLICABLE, 6 source authorities, the 9 views, the public 24 entity types and 25 predicates). --task <text> reorders views by task relevance (ranked_catalog_views) and adds task_ranking metadata to the output (including --profile read); --subgraph filters, --profile full|read and --format auto|table shape output. See ontology.md for the catalog itself.
graph describe routes through GraphService.describe like every other workbench command, so the ontology it reports is the serving host's build (the daemon's, in the default host mode), not the CLI binary's.

graph read is the Retrieve axis — resolves a named <subgraph>.<view> (one of the 9 views), validates required scope/filters, then routes through the one read trunk to an AgentEnvelope of ranked evidence. There is no server-side answer synthesis. timeline recent is the same path as graph read --subgraph recent_changes --view timeline. Reader/ranking/view detail lives in querying.md.

Text-mode presentation (default human output, no --json): --format, --detail, and --relations shape the layout and depth independently.

Flag	Text effect
`--format events` (timeline default)	Bullet list of deduped timeline events
`--format table`	Markdown pipe table (`occurred_at \| source_ref \| …`)
`--format raw` (non-timeline default)	Bullet list of ranked items
`--detail compact`	Core fields only (fact, score, summary)
`--detail full`	Adds truth, coverage, claim/breakdown metadata
`--relations summary`	Inline relation counts and predicate names
`--relations full`	Indented relation sub-lines (or secondary table with `--format table`)

# Human markdown table for recent changes
potpie graph read --subgraph recent_changes --view timeline --format table --limit 10

# Deeper relation detail in text mode
potpie graph read --subgraph recent_changes --view timeline \
  --format events --detail full --relations full --limit 5

# Non-timeline entity table
potpie graph read --subgraph decisions --view preferences_for_scope \
  --scope language:python --format table --limit 5

--json uses the same item shaping (detail / relations) but emits structured JSON instead of human tables/bullets.

graph search-entities is the Filter axis (identity resolution before a write) — structured per-entity lookup, not through the read trunk.

Write (route `host.graph_workbench`)

The canonical write door is graph propose → graph commit --verify (Spine A; application/services/graph_workbench.py). graph mutate is a legacy wrapper that internally calls propose+commit.

potpie graph propose [--file <path> | (stdin)] [--ttl 1h] [--pot <ref>]
potpie graph commit  <plan_id> [--approved-by <who>] [--verify] [--pot <ref>]

potpie graph mutate  [--file <path> | (stdin)] [--dry-run] [--allow-review-required] [--approved-by <who>] [--pot <ref>]

potpie graph bulk apply [--file <path> | (stdin: NDJSON/JSON)] \
  [--chunk-size 100] [--start-chunk 1] [--dry-run] [--continue-on-error] \
  [--verify] [--manifest <path>] [--idempotency-key <k>] [--ttl 1h] [--approved-by <who>] [--pot <ref>]

potpie graph mutation-template \
  [--kind repo-baseline|feature|preference|preference-policy|infra-snapshot|bug-fix|decision|timeline-event|timeline-change]

potpie graph history [--entity <key>] [--claim <key>] [--subgraph <s>] [--plan <id>] [--mutation <id>] \
  [--since <t>] [--until <t>] [--limit 50] [--pot <ref>]

potpie graph nudge --event <e> --session <id> [--path <p>] [--scope <k:v>] [--query <text>] [--limit 5] [--pot <ref>]

graph propose validates + lowers a semantic-DSL batch and persists a plan record (no graph write). The batch payload uses flat op fields (op/subject/predicate/object/value/truth/confidence/evidence[]/description/…; append_event uses verb/occurred_at/actor/targets[]/mentions[]); nested {"event":{…}}/{"claim":{…}} shapes will not parse. The DSL, validation/risk, and the diff shape are owned by writing.md.
graph commit <plan_id> applies a stored plan by id; the agent does not resend mutations. --verify reads the committed claims back and downgrades on missing readback / quality regressions. Medium/high-risk plans need --approved-by.
graph mutate — legacy wrapper (emits a warning steering to propose/commit). --dry-run previews; --allow-review-required + --approved-by auto-applies medium/high-risk ops.
graph bulk apply — chunked NDJSON/JSON application with resumability (--start-chunk, --manifest), --continue-on-error, and idempotency.
graph mutation-template — emits a static schema-only skeleton (no host call).
graph nudge — the zero-token in-session trigger (host.nudge.nudge); the trigger model and the Claude Code hook are in ingestion-nudge.md. NudgeEvent values: session_start, pre_edit, pre_deploy, test_failed, test_passed, stop.

Inbox (`graph inbox …`) — capture uncertain work

Inbox items are pending graph work that never become facts until a harness processes them through propose/commit.

potpie graph inbox add
potpie graph inbox list
potpie graph inbox show          <id>
potpie graph inbox claim         <id>
potpie graph inbox mark-applied  <id> [--plan <id>] [--mutation <id>]
potpie graph inbox mark-rejected <id> [--reason <text>]
potpie graph inbox close         <id>

mark-applied requires a linked --plan or --mutation. States: pending → claimed → applied/rejected/closed.

Quality (`graph quality …`) — read-only diagnostics

potpie graph quality summary             [--subgraph <s>] [--limit 50] [--pot <ref>]
potpie graph quality duplicate-candidates [--subgraph <s>] [--limit 50] [--pot <ref>]
potpie graph quality stale-facts         [--subgraph <s>] [--limit 50] [--pot <ref>]
potpie graph quality conflicting-claims  [--subgraph <s>] [--limit 50] [--pot <ref>]
potpie graph quality orphan-entities     [--subgraph <s>] [--limit 50] [--pot <ref>]
potpie graph quality low-confidence      [--threshold 0.5] [--subgraph <s>] [--limit 50] [--pot <ref>]
potpie graph quality projection-drift    [--subgraph <s>] [--limit 50] [--pot <ref>]

Quality never writes — it recommends repairs through propose/commit or the inbox (writing.md).

Backend-capability commands (route `host.backend`)

potpie graph neighborhood --entity <key> [--predicate <p>] [--depth 2] [--direction out|in|both] \
                          [--limit 50] [--detail summary|full] [--pot <ref>]
potpie graph inspect <entity_key> [--depth 2] [--pot <ref>]      # legacy alias of neighborhood

potpie graph export <file> [--pot <ref>]
potpie graph import <file> [--pot <ref>]
potpie graph repair [--semantic-index] [--entity-summaries] [--all] [--pot <ref>]

graph neighborhood is the Traverse axis (first-class), backed by backend.inspection.neighborhood. graph inspect is a legacy alias that warns toward neighborhood.
Unbuilt capabilities surface as the structured not-implemented contract via _require_backend_capability. Per-profile coverage is in architecture.md.

Roadmap (not yet wired): snapshot (graph export/import) is real only on the in_memory/embedded backends; graph inspect/neighborhood is unavailable on neo4j. On the OSS default falkordb_lite, export/import are unavailable.

Backend group (`backend` → `host.backend`)

potpie backend list                 # KNOWN_PROFILES + active marker
potpie backend status               # capability report
potpie backend use <profile>        # advisory only — NOT persisted
potpie backend doctor               # backend.mutation.readiness()

KNOWN_PROFILES = in_memory, embedded, neo4j, falkordb, falkordb_lite, postgres, chroma, hosted. backend use is advisory only (persisted: False) — it suggests setting CONTEXT_ENGINE_BACKEND; it does not switch the live backend. CLI code never queries SQLite/Neo4j/vector indexes/state tables directly — it routes through services and capability ports.

Environment switches (consolidated)

Variable	Purpose / default
`CONTEXT_ENGINE_BACKEND`	preferred backend selector (host default `falkordb_lite`)
`GRAPH_DB_BACKEND`	legacy fallback selector (ingestion server default `neo4j`)
`CONTEXT_ENGINE_HOST_MODE`	`daemon` (default) \| `in_process`
`CONTEXT_ENGINE_EMBEDDER`	`none` disables the bundled local embedder
`CONTEXT_ENGINE_ONTOLOGY_SOFT_FAIL`	downgrade-instead-of-fail validation
`CONTEXT_ENGINE_AGENT_PLANNER_ENABLED`	service-side LLM reconciliation (default off)
`CONTEXT_ENGINE_MAX_CHUNK_EVENTS`	batch chunk size (default 20)
`CONTEXT_ENGINE_RECONCILIATION_ENABLED` / `_INFER_LABELS` / `_CONFLICT_DETECT` / `_AUTO_SUPERSEDE`	reconciliation feature flags
`CONTEXT_ENGINE_ALLOW_UNSIGNED_WEBHOOKS`, `GITHUB_WEBHOOK_SECRET`, `CONTEXT_ENGINE_INGEST_422`	webhook/ingest controls
`POTPIE_HOOK_DEBUG`, `POTPIE_HOOK_TIMEOUT`, `POTPIE_BIN`, `POTPIE_POT`	nudge-hook env

Backend precedence: CONTEXT_ENGINE_BACKEND > GRAPH_DB_BACKEND > falkordb_lite. There is no NotImplementedError gate on falkordb anywhere.

Canonical journey

flowchart LR
  cf_setup["setup --repo . --agent claude"]
  cf_status["status"]
  cf_read["graph catalog → graph read / search-entities"]
  cf_write["graph propose → graph commit --verify"]
  cf_nudge["graph nudge (zero-token hook)"]

  cf_setup --> cf_status --> cf_read --> cf_write
  cf_nudge -.-> cf_read
  cf_write -.-> cf_read

Local first run (OSS default — falkordb_lite, detached daemon, skills installed during setup):

Published package:

uv tool install potpie   # or: pip install potpie
potpie setup --repo . --agent claude
potpie status

This repo (local development): prefer make cli-install so the graph-explorer UI is built and any old daemon is stopped before the editable install.

make cli-install
potpie setup --repo . --agent claude
potpie status

# read the contract, then the graph
potpie graph catalog --profile read
potpie graph read --subgraph debugging --view prior_occurrences --scope service:refunds-api

# resolve identity, then write through the canonical door
potpie graph search-entities "refund timeout" --type BugPattern
potpie graph propose --file mutation.json
potpie graph commit <plan_id> --verify

Roadmap (not yet wired): the managed-backend journey (potpie login → potpie use <pot> --managed → the same potpie graph … commands) is documented but raises CapabilityNotImplemented today.

Output contract

Human output: an action-oriented summary plus a suggested next command.
--json: stable fields for agents/scripts (additive changes are OK); errors carry code, message, detail, recommended_next_action.
setup --dry-run: returns a preview document; no mutation, dependency setup, source registration, or skill install occurs.
Destructive commands (pot reset, pot archive) require --confirm or interactive confirmation.
Exit codes follow the contract() table above (0/1/2/3/4).