Argent

Argent (bundled skills, CLI-only)

This parent bundles all official Argent skills from @swmansion/argent (see UPSTREAM.md for source, version, and inventory). The skills in the direct argent-* child directories describe MCP tool calls; this parent is the CLI translation layer — no MCP client is required.

CLI-only guard (applies to every nested skill)

• Everything an upstream skill phrases as "call tool X" runs as argent run X --args '<json>' (single-quoted JSON, matching the upstream examples verbatim), argent run X --flag value for simple flags, or --args - with JSON on stdin for large payloads. Discover exact flags with argent tools describe <name> — never guess from an MCP schema. (Note: argent tools describe --help treats --help as a tool name — use argent tools --help for help.)
• Discover tools with argent tools and argent tools describe <name>.
• Do not run argent init/install/update/uninstall, register MCP servers, or copy any argent config, skills, or rules into the repo — flows run without init; the tool server auto-starts.
• Keep telemetry disabled (set it with argent telemetry disable --scope global if not already); never read telemetry status (it may list personal identifiers).
• Never put secrets, tokens, or private user data in flow YAML, run reports, or screenshots. Never type plaintext credentials — use the {{secret:<NAME>}} placeholder mechanism the nested skills document.
• Nested skills may name side-effectful setup (permission grants/resets, service teardown, feature flags such as argent enable argent-lens). Never run an unapproved keychain change (e.g. the errSecInternalComponent keychain fix in argent-ios-device-setup asks for the login password — ask the user first), device wipe/-wipe-data, unrelated process termination, or feature-config enablement — those need explicit user approval first. Prefer in-app and on-dialog paths (e.g. argent-settings-permissions is a last resort, exactly as that skill states).
• If the argent CLI is missing or no usable device exists, stop and report BLOCKED with the missing prerequisite.
• Device pick: argent run list-devices --json. Simulator = iOS entry without kind: "device" (boot with argent run boot-device --udid <UDID>). Physical iPhone = iOS entry with kind: "device"; only state: "connected" (cabled, USB usable) can run — paired means not reachable, and there is no boot-device for phones. The first flow run on a phone builds and signs an on-device runner (minutes cold).

CLI invocations for the tool calls in nested skills

• Tool call X { "udid": "...", ... } → argent run X --args '{"udid":"..."}'. Nested run-sequence steps arrive as --steps-json '[...]'. Booleans: --flag, --flag true, --flag=false, --no-flag. Arrays are repeatable flags.
• Tool output images (screenshot, screenshot-diff): the CLI writes the PNG to a temp path and prints it instead of attaching an image; pass --includeImageInContext false + --scale 1.0 for full-resolution baselines (per argent-screenshot-diff) and open the printed path when pixels matter.
• Returned file paths (screenshots, recordings, profiler traces, log files) are real host paths — follow them with your file tools as the nested skill says.
• Tools listed as unavailable in a context (e.g. on a physical iPhone) fail the same way through the CLI; the nested skill's limits still hold.
• Flows: argent flow run <path.yaml> --device <id> --platform ios --json — always a file path, never a bare flow name (see the flow-policy override below).

Flow storage override (overrides upstream path instructions)

Upstream skills default flows to .argent/flows/<name>.yaml. In this repo, flow files live ONLY at _ai/task/{SLUG}/reproduction/flows/<safe-name>.yaml (safe name: letters, numbers, _, -) so reproduce-bug and verification-gate can replay the exact file on the candidate; never save flows under .argent/flows (that includes recorder defaults, run: targets, and snapshot baselines — the latter live beside the flow when needed). Pass the file path to argent flow run, never a bare flow name. Where an upstream text says .argent/flows/<name>.yaml, read it as the file path form: argent flow run <path.yaml>.

Recorder coverage: the flow-recorder tools (flow-start-recording, flow-add-step, flow-add-echo, flow-finish-recording) are keyed by name + absolute project_root and always write under <project_root>/.argent/flows/ — no flag redirects them. Under this repo's storage contract flow YAML lives ONLY at the task path, so the recorder is unavailable: author flow YAML directly with file tools at the task path against the upstream schema (argent-create-flow/references/flow-yaml.md), executing steps live as you build them. No temporary .argent/flows workaround is sanctioned — never create .argent/ anywhere in this repo, and do not invent recorder flags.

Candidate provenance before any replay (mandatory)

Before replaying a flow on an installed app, establish that the installed app is the exact candidate being verified (same commit/build). If the app is missing or stale (built from an earlier commit, or provenance cannot be established), build and install the candidate via xcodebuildmcp-cli first; if provenance cannot be established at all, report BLOCKED. When the candidate is already proven installed on the device, do not rebuild routinely. A stale build can produce a false pass; this check is what makes flow proof candidate-bound.

Division of labor

• xcodebuildmcp-cli owns builds and mechanical proof; Argent owns user-flow proof. Do not run both routinely.
• This bundle's sanctioned replay route is argent flow run <path.yaml> — never a bare name. flow-execute also supports an explicit flow_path; only its name form resolves .argent/flows/. Choosing the standalone task-path route is repo policy, not a tool limitation. The recorder tools, unlike flow-execute, have no output-path override and are unavailable under the flow-storage policy above; author YAML with file tools instead.

iOS reproduction contract (with reproduce-bug)

For bug reproduction the flow is the shortest faithful replay of the reported entry point, and it must be discriminating — not just "the actions ran":

1. Initial-state assertion: start with a check on the screen the reporter started from (e.g. await: { visible: { id: ... } }), so a wrong-screen run fails loudly instead of passing vacuously.
2. Reported entry point: replay the shortest sequence from that entry point, no extra setup.
3. Expected-behavior assertion: end with a condition that holds only when the bug is fixed (assert:/await: on the expected state), so the same flow proves the fix and fails on the bug. Never weaken a requested check to obtain a pass.
4. Unexpected replay failure → only the diagnosis the failure needs (e.g. a snapshot: step or one-off argent run describe); do not expand into broad QA.

Router — load only the nested skill the task needs

Each nested skill is self-contained; load only what the current task needs, and use its exact local path below.

For iOS bug reproduction and user-flow verification, start with the reproduce-bug / verification-gate routing outside this bundle, then use argent-create-flow (authoring) and argent-ios-simulator-setup / argent-ios-device-setup (device pick) from here.

License notice

The nested skills are imported from @swmansion/argent (Software Mansion), licensed under the Apache License, Version 2.0, copyright 2026 Software Mansion S.A. The complete upstream license text is bundled verbatim at LICENSE in this directory and must travel with any redistribution of this bundle; it also carries the upstream package's additional terms for its proprietary binary components, none of which are included in this skills-only bundle. See UPSTREAM.md for exact provenance.