Imported from jfrog/devin-plugin (skills/jfrog/SKILL.md). Install upstream with npx skills add jfrog/devin-plugin --skill jfrog. Copyright stays with the author.

JFrog Skill

The foundational skill for all JFrog agent interactions. Covers JFrog Platform concepts, jf CLI setup and authentication, and intent routing to workflow skills.

At a glance (always-read core)

Network-facing jf this session. Exempt until <SID>: jf --version, jf config show.

Tier A — always-read floor (before first non-exempt jf):

• UA: Environment check once → on exit 0/1, export its exact stdout line as JFROG_CLI_USER_AGENT atop every bash that runs jf (never invent / rebuild the UA)
• CLI offer: after Environment check exit 0/1 (skip jfrog-init / MCP-only). NEWER_AVAILABLE → stop, Yes/No; SKIP/No → silent
• Server: resolve default once → --server-id <SID> after subcommand (jf api --server-id …, never jf --server-id … api). One request → one server (unless user names servers, e.g. compare <a> and <b>)
• Error (401/403/404/timeout): stop — never retry another server / never infer multi-server. Override only if user names a server
• No prep mutations: missing repo/artifact/build → stop + report; no create/copy/move/upload to fill the gap (workaround ask ≠ permission)
• Never guess tools / jf api paths → tool list / --help / references/. 404 → stop (no guessed retry). jf api needs product prefix (/artifactory, /xray, …)
• Hard-rule signals: Cautious execution, Server selection rules, Gotchas Tier A bullets below — not tips
• Gotcha floor (Tier A): never interactive (jf config add, jf login, template wizards, …); if a call fails with --server-id, do not retry without it; 401/403/404/timeout → stop, never hop servers; --quiet is not global — check --help before adding it

Tier B — path-gated MUST (before jf api / AQL / advanced CLI I/O / MCP-result-via-shell anti-patterns): full references/cli-gotchas.md, references/jf-api.md, references/preserving-command-output.md, references/cli-command-discovery.md. Setup / jf setup / ordinary CLI do not require Tier B.

Tier C — on-demand: references/INDEX.md domain refs; login / CLI install when needed.

Contents (prefer full SKILL.md; At a glance = Tier A floor if you only see the top):

Floor for partial reads: Tier A (this section) before first non-exempt jf. Prefer the full SKILL.md when you can. Load Tier B only when the next action needs jf api / AQL / advanced CLI I/O (checklist). Domain detail → Tier C references/INDEX.md.

Interact with the JFrog Platform through three tool tiers — see Tool selection strategy. In code examples, <skill_path> is this skill's directory, resolved automatically by the agent. If unresolved, locate this SKILL.md file and use its parent directory.

Out of scope: MCP server management. Installing, listing, removing, or configuring MCP servers (e.g. "install an MCP", "what MCPs can I install", "list my MCPs", "which MCP servers can I use", "what's in my approved MCP catalog") is handled by the JFrog Agent Guard workflow (@jfrog/agent-guard) — a separate workflow, not this skill.

Tool selection strategy

Try the tiers in order; move to the next only when the current does not cover the operation or fails:

1. JFrog MCP tools (preferred): CallMcpTool against the JFrog MCP server. Discover available tools from the server's tool list; never guess tool names.
2. jf CLI subcommands (fallback): dedicated commands such as jf rt upload, jf rt dl, jf build-publish.
3. jf api (last resort): REST/GraphQL endpoints with no dedicated subcommand. Validate the path first — see rule 6 in Cautious execution.

MCP and CLI may use different token scopes. One tier returns 403 → try the other tier before reporting the operation blocked.

Prerequisites

The following tools must be available on PATH:

All JFrog HTTP traffic from Tiers 2 and 3 goes through the jf CLI itself (jf api, see Invoking platform APIs with jf api below) — no standalone curl is required for any JFrog interaction.

Runtime permission for JFrog calls. All jf calls that touch the network need an outbound-HTTPS escalation from the agent runtime. The ~/.jfrog/ credential save (jf config add during login) additionally needs a filesystem-write escalation.

If jf exits 1 with empty output, the runtime's network gate is the first thing to check — re-run with the appropriate escalation above.

Environment check

MCP (Tier 1) skips this check — proceed immediately. Before your first Tier 2 or Tier 3 (jf) operation this session, run the environment check. On exit 0/1, remember its stdout line verbatim as <UA> for the rest of the session. Skip the CLI offer during jfrog-init.

bash <skill_path>/scripts/check-environment.sh <model-slug>
# exit 0/1 stdout: exactly one opaque line — that line IS <UA>. Copy it byte-for-byte.
# Do not parse, rebuild, or approximate the export value from this comment.
# stderr: JSON state (cached 24h at ${JFROG_CLI_HOME_DIR:-$HOME/.jfrog}/skills-cache/jfrog-skill-state.json)

Then, on exit 0/1 only:

bash <skill_path>/scripts/cli-newer-version-offer.sh
# SKIP → continue the original task; do not mention the offer.
# NEWER_AVAILABLE <cli_version> <latest> (suggest_upgrade) → stop; Yes / No. After:
#   bash <skill_path>/scripts/cli-newer-version-offer.sh --clear
# Yes → references/jfrog-cli-install-upgrade.md, then check-environment.sh --force
# (Tier 2/3 only on exit 0/1). No → silent. Next offer = next new latest.

Exit 2/3 produces no <UA>; follow the exit table below and do not proceed to Tier 2 or 3.

Pass your own model slug, lowercased, with version (e.g. opus-4.7, gpt-5.6-sol, gemini-2.5-pro, composer-2-fast). Examples, not an allowlist — emit a new/unlisted name verbatim, not unknown. Not harness/role (subagent, agent) or bare family (claude, gpt); subagents inherit the parent's slug. unknown only if truly unidentifiable.

Never invent JFROG_CLI_USER_AGENT

On exit 0/1, the script's stdout line is <UA> — export it verbatim (never invent, rebuild, or edit it). On exit 2/3 there is no <UA> — do not synthesize one. Current stdout starts with jfrog-skills/, never with model/.

• Parent session: if <UA> is missing or starts with model/ (legacy), discard it and re-run check-environment.sh. Export the new exit 0/1 stdout line only when it does not start with model/; otherwise stop (do not invent).
• Subagents: use only the parent-passed exact <UA> — never re-run the script or construct a replacement. If that value is missing or starts with model/, stop (do not export / do not invent); do not re-run.

Export JFROG_CLI_USER_AGENT once per bash invocation

At the top of every bash invocation that runs jf, export <UA> once; all jf calls in that invocation pick it up:

export JFROG_CLI_USER_AGENT='<UA>'
export JFROG_CLI_AI_MODEL='<model-slug>'   # jf >= 2.120.0 emits ai-model/<slug> from this
jf config show
jf api /artifactory/api/system/version

JFROG_CLI_AI_MODEL carries the model the CLI cannot infer from the environment; export it alongside <UA> (same <model-slug> you passed the script). Older CLIs ignore it; the remembered <UA> already carries the slug when the script recorded one.

Do not repeat the assignment per jf call (JFROG_CLI_USER_AGENT='<UA>' jf … on every line). This is a session-global invariant: it applies to every jf invocation in the session, including jf calls you make while following any workflow skill that builds on this base skill. Examples elsewhere in this skill and in references/*.md omit the export for readability — the rule is global. When launching a subagent, pass <UA> in its prompt (see Never invent) and whether the CLI offer already ran. Subagents do not re-ask.

Exit 2 or 3 prints no <UA> on stdout. Do not invent or hand-assemble one from this file or from jf --version.

Exit 2 or 3 is not a fatal error. Attempt to install or upgrade the CLI (see references/jfrog-cli-install-upgrade.md). If installation succeeds, re-run the environment check. If installation is not possible (no permissions, restricted environment), proceed with MCP (Tier 1) only. Both jf CLI commands (Tier 2) and jf api (Tier 3) require a working jf installation.

JSON parsing (jq)

Use jq for all JSON parsing of CLI and API output (pipes, -r, filters).

~/.jfrog/skills-cache/ — allowed files only

${JFROG_CLI_HOME_DIR:-$HOME/.jfrog}/skills-cache/ is not a general scratch or temp directory. Use it only for these two artifacts:

1. jfrog-skill-state.json — written by scripts/check-environment.sh (24-hour CLI check cache).
2. onemodel-schema-${JFROG_SERVER_ID}.graphql — cached OneModel supergraph schema (see references/onemodel-graphql.md).

Do not save HTTP response bodies, GraphQL query results, ad-hoc JSON, reports, or any other temporary files under skills-cache/. Write those to a host temp path instead (for example /tmp/<name>-$$.json or mktemp -d), echo the path when a follow-up Shell step must read the file — same pattern as Preserving command output below.

Cautious execution

HARD RULES — never skip. Speculative / preparatory / guessed ops are forbidden. Before any JFrog CLI command, MCP tool call, or API call:

1. Confirm the operation is needed to fulfill the user's request. If the request is ambiguous or could refer to multiple systems (e.g. "builds" could mean Artifactory build-info or CI/CD pipeline runs), ask the user for clarification instead of guessing. Never fetch data from the wrong system — a wrong answer is worse than asking a question.
2. Resolve the target server using the Server selection rules below — there must be no ambiguity about which server is used
3. For mutating operations (create, update, delete, upload), confirm with the user unless the intent is clearly implied. This applies to all tiers (MCP tools, CLI commands, and jf api with POST/PUT/DELETE).
4. Prefer read operations first to understand current state before making changes
5. Never invent preparatory mutations. If the requested operation fails because a precondition is not met (artifact missing from the specified repo, repository does not exist, package not at the expected location, build not found), stop and report the gap to the user. Do not perform copy, move, upload, create-repo, or any other mutating operation to satisfy the precondition. "Put it there so the download succeeds", "make it work", or "do whatever you need" is still a workaround — not permission to invent the missing artifact. Only perform that mutation when it is the user's requested work (publish this file, create this repo, move this artifact), not a helper to make a different operation succeed. These "helper" mutations can have cascading effects the user has not considered — virtual repository resolution changes, storage quota consumption, replication triggers, Xray re-indexing, or permission propagation.
6. Never guess tool names or API paths. For MCP tools, confirm the tool exists in the server's tool list. For jf api paths, validate against <skill_path>/references/ (or JFrog OpenAPI specifications if you have web access). On a 404, stop and report — never retry with a guessed alternative path.

Server selection rules (mandatory)

HARD RULES — never skip or soften. Wrong-server answers and silent server-switching are worse than stopping to ask.

Single-server invariant. After <SID> is resolved, every subsequent network-facing jf call MUST pass --server-id <SID> (default resolved below); bootstrap jf --version / jf config show stay exempt until then. For one user request, all network jf calls use exactly one server-id — unless the user names servers to compare (e.g. compare <a> and <b>), where each call passes its own target's --server-id.

JFrog MCP and CLI use independent auth. MCP tools authenticate through the MCP server session (not jf config); CLI commands authenticate through jf config. If you switch the CLI target server via jf config use, the MCP connection still points to its original server. Do not mix MCP and CLI calls targeting different servers in the same session. If the user asks to switch servers, warn that MCP tools will continue to target the original server until the MCP connection is re-established.

MUST NOT retry on a second configured server after 401/403/404, empty, or partial results; MUST NOT infer multi-server intent from "my"/"our" or from seeing extra entries in jf config show. Override: only when the user explicitly names another id ("on <id>, …", "use <id>", "compare <a> and <b>") — inferred intent is not an override.

Resolve the default once per session

Before your first jf call, resolve the default server-id and remember it as <SID> for the rest of the session, same pattern as <UA>:

jf config show 2>/dev/null \
  | awk '/^Server ID:/{id=$NF} /^Default:[[:space:]]*true/{print id; exit}'
# stdout: the default server-id; if empty, stop and ask which to use

Pass --server-id <SID> to every subsequent jf call. The flag goes after the subcommand name, not after jf itself:

• ✅ jf api --server-id <SID> /artifactory/api/system/version
• ✅ jf rt ping --server-id <SID>
• ❌ jf --server-id <SID> api /… — fails with flag provided but not defined

When launching a subagent, pass <SID> in its prompt — subagents do not re-resolve. Examples elsewhere in this skill and in references/*.md omit --server-id for readability; the rule is global, same as JFROG_CLI_USER_AGENT. To add a new server, read references/jfrog-login-flow.md.

On any error, stop — never switch

If a jf call returns 401/403, 404, network error, timeout, or any other failure, stop with no further jf calls and respond:

<server-id> returned <code> for <endpoint>: <short reason>. Other configured server(s): <list> — I won't query them without your explicit instruction. How would you like to proceed?

Path-gated base references (MUST before jf api / advanced CLI)

These four files are Tier B of the base skill — content that used to live in this SKILL.md. They are not optional INDEX domain lookups, and they are not required before every CLI / setup path.

MUST read every one in full before jf api, AQL via jf api, advanced CLI I/O (temp-file / stdout-stderr patterns), or acting on MCP results via shell/jq. Ordinary jf (e.g. jf setup, jf rt … with known flags) needs Tier A only (At a glance).

The short Gotchas Tier A bullets are the session floor — they do not replace full references/cli-gotchas.md when you enter Tier B.

Skipping any of these on a Tier B path = incomplete base-skill load / hard-rule violation. Skipping them on a Tier A-only path (setup / simple CLI) is not a violation.

When to read reference files

Prefer reading this SKILL.md in full. At a glance is the Tier A floor for partial readers. Path-gated base references above are Tier B (mandatory on those paths, not every session). Everything else under references/INDEX.md is Tier C domain detail — load ≤2–3 most specific files for the task; skip unused domains.

references/INDEX.md lists every references/*.md file (Tier B + Tier C). Add/rename/remove a file → update INDEX in the same change — CI (tests/jfrog/test_reference_index_contract.py) fails if they diverge.

Command discovery

Run --help to verify options — do not rely on memorized commands.

1. jf --help → 2. jf <namespace> --help → 3. jf <command> --help

Tier B — MUST read in full before relying on discovery beyond --help: references/cli-command-discovery.md (namespaces, top-level lifecycle/security commands, Pipelines sunset).

Invoking platform APIs with jf api

Tier 3 for Platform REST/GraphQL, auto-authenticated. Do not use jf rt curl / jf xr curl. Always include the product prefix (/artifactory, /xray, /access, …) — omit → 404.

Tier B — MUST read in full before jf api: references/jf-api.md (prefixes, flags, examples, OneModel GraphQL payload). Body on stdout / status on stderr — see Gotchas + full cli-gotchas.md (Tier B).

Structured inputs

Interactive wizards (jf rt rpt / ptt / rplt) are unusable for agents. Fetch an existing config via REST and edit:

jf api /artifactory/api/repositories/<repo-key>

More REST/template patterns → references/artifactory-api-gaps.md or references/platform-admin-api-gaps.md via references/INDEX.md.

Gotchas — hard rules (never skip)

Not tips. Tier A bullets below are the always-read floor. Full references/cli-gotchas.md is Tier B — MUST before jf api / AQL / advanced CLI I/O / MCP-via-shell; not required before every CLI or jf setup. Short bullets do not replace the full file on Tier B paths.

Tier A floor (every non-exempt jf session):

• --quiet is not global — check --help before adding it
• --server-id: if a call fails with it, do not retry without it (silent default-server switch). See Server selection rules
• Non-interactive only — avoid jf config add, jf login, *template wizards; use references/jfrog-login-flow.md / REST
• Auth errors: 401 → re-login same server; 403 → permissions; 404 → path/prefix/version. Never switch configured servers as a workaround

Tier B reminders (load full cli-gotchas.md + sibling Tier B refs before these paths):

• MCP: read structured tool results directly — do not pipe through shell/jq
• jf api I/O: body → stdout, status → stderr; pipe stdout to jq; never 2>&1 | jq. No -L / -o — redirect: jf api … > /tmp/out-$$.json
• Product prefix required on every jf api path (see Tier B jf-api.md)
• Never re-fetch to retry jq — save output first (Preserving command output + Tier B preserving-command-output.md)

Batch and parallel execution

Independent ops → lightest parallelism: (1) loops/& in one Shell, (2) parallel Shell calls, (3) subagents for large fan-out. Details → references/general-parallel-execution.md.

Preserving command output

Save network responses to a temp file; echo the path; re-read for jq — never re-run the same network call to fix parsing.

Tier B — MUST read in full before advanced I/O / re-parse patterns: references/preserving-command-output.md ($$ + echo, session id, no re-fetch / no cross-context reuse).

Before you run jf — quick checklist

At a glance Tier A floor; add Tier B only when the next action needs jf api / advanced CLI:

• export JFROG_CLI_USER_AGENT='<UA>' in this bash — <UA> is the exact stdout line from check-environment.sh exit 0/1 (never invent / rebuild)
• CLI offer (cli-newer-version-offer.sh) done or N/A (jfrog-init / MCP-only)
• network jf: --server-id <SID> after subcommand (not jf --version / jf config show pre-SID)
• one server; error → stop, don't switch (multi only if user names / compare)
• no prep create/copy/move/upload to fill a gap (workaround ask ≠ permission)
• never guess tools/paths → list / --help / references/; 404 → stop; jf api product prefix (/artifactory, /xray, …)
• Tier A hard rules: Cautious execution + Server selection + Gotchas Tier A floor (interactive / --server-id retry / stop-on-error / --quiet)
• Tier B (only if next action is jf api / AQL / advanced CLI I/O): full cli-gotchas.md, jf-api.md, preserving-command-output.md, cli-command-discovery.md