Atlassian Tasks

Inputs

Preconditions

• instructions is mandatory; ask if missing.
• A resolvable Atlassian URL is required. Use url if provided; otherwise extract from instructions (any host paired with /jira, /wiki, /confluence, /secure, /browse, or /rest). If none found, ask.
• Never ask for a token when pat_token_file is absent — use Mode B.
• Never assume values for required API fields not supplied by the user; ask explicitly.
• Never read, print, summarize, quote, or copy PAT values, browser-auth cookie values, or credential-bearing headers. Treat pat_token_file and browser-auth state_file values as opaque credential source paths only.

Workflow

1. Resolve URL. Prefer url; otherwise extract from instructions. If multiple candidates, take the most specific. If none, stop and ask.
2. Identify product (jira or confluence) from URL and/or instructions.
3. Resolve endpoint. Call atlassian-endpoint-resolver with the inferred operation (CRUD verb + resource type). Use the top-ranked result.
4. Check for missing fields. If any required path/query/body field is not derivable from the URL or instructions, ask for only those values and wait.
5. Format rich text. Before constructing a body that includes Jira descriptions/comments/environment/textarea fields or Confluence page/blog/comment bodies, use atlassian-rich-text-formatter with the product, endpoint/schema context, field name, and source content. Use the returned ADF, wiki markup, or Confluence storage value exactly.
6. Authenticate. Mode A (PAT) or Mode B (browser) per rules below.
7. Execute. Build only non-secret request metadata, then call scripts/request.py per rules below.
8. Return results per Output spec.

Authentication

Mode A — PAT file provided: Do not open or read pat_token_file; pass only its path to scripts/request.py with --pat-file.

Mode B — No PAT file:

1. Launch atlassian-browser-authentication with the resolved URL. If it fails or returns no state_file, stop and report the error.
2. Do not open or read state_file; pass only its path to scripts/request.py with --state-file.
3. On 401 Unauthorized, rerun atlassian-browser-authentication with FORCE and retry once.

Request Construction

• Base URL: <scheme>://<host><context_path> where context_path is the non-API prefix (e.g. /jira, /confluence).
• Endpoint URL: Append resolved path to base URL, substitute {paramName} tokens, append URL-encoded query params.
• Body: JSON built from endpoint schema and user-supplied values only. Do not fabricate or default any field. For request bodies, write non-secret JSON to a temporary file, pass it with --body-file, then delete the temporary file. Do not delete pat_token_file or browser-auth state_file.
• Execution: Run every API request with: uv run "$SKILL_DIR/scripts/request.py" --method <METHOD> --url <URL> (--pat-file <PAT_FILE> | --state-file <STATE_FILE>) [--query KEY=VALUE ...] [--body-file <BODY_JSON>]. Prefer reusable prefix approval: ["uv", "run", "$SKILL_DIR/scripts/request.py"].
• Rich text fields: Never pass raw Markdown into Jira or Confluence REST rich-text fields. Use atlassian-rich-text-formatter to generate the endpoint's expected representation, then insert that value into the JSON body without reformatting it.
• Pagination: Auto-fetch subsequent pages when a response contains startAt/maxResults, a pagination token, or _links.next, unless instructions imply a limit.

Sub-Agent Rules

• All API requests execute in sub-agents, not the main agent.
• Independent calls run in parallel; dependent calls run sequentially.
• Merge all sub-agent outputs before returning.

Output

• Single call: Return the result directly.
• Multiple calls, aggregation needed: Return the combined result as the primary output. Omit per-call breakdowns unless the user asks or a call failed.
• Multiple calls, independent results: Present each result in execution order.
• Any failure: Include the failed step, HTTP status, and a short response excerpt.