doc-gather Skill | Agent Skills

doc-gather

Gather context from Jira tickets, source repos, and documentation repos to produce a structured context package for downstream skills.

Parse arguments

$ARGUMENTS contains either:

A Jira ticket key (e.g., RHOAIENG-55490)
A PR/MR URL (e.g., https://github.com/org/repo/pull/123)

Detect the input type:

If it starts with http → PR URL
Otherwise → Jira ticket key

Step 1: Resolve ticket metadata

For Jira tickets

Use the resolve script to fetch and structure ticket metadata:

bash ${CLAUDE_SKILL_DIR}/scripts/resolve-jira.sh <JIRA-KEY>

This returns JSON with: key, summary, description, acceptance_criteria, fix_versions, components, linked_tickets, epic_key, status, issue_type.

If the ticket has linked tickets that provide useful context (e.g., parent epic, related stories), resolve those too (up to 5 linked tickets):

bash ${CLAUDE_SKILL_DIR}/scripts/resolve-jira.sh <LINKED-KEY>

For PR URLs

Use gh CLI to extract PR metadata:

gh pr view <PR-URL> --json title,body,labels,files,headRefName

Extract the Jira key from the PR title or body (pattern: [A-Z]+-\d+), then resolve the Jira ticket using the resolve script as above.

Step 2: Parse product configuration

Read the product configuration bundled with this skill:

python3 ${CLAUDE_SKILL_DIR}/scripts/parse-product-config.py ${CLAUDE_SKILL_DIR}/references/rhoai.yaml

From the config, determine:

Context sources — which repos to clone and where to find files
Version mappings — how to map Jira fixVersion to git branches
Component resolver — map Jira component names to repo slugs
Docs conventions — module prefixes, framework, attribute files

Step 3: Resolve version to branches

For each fixVersion from the ticket:

python3 ${CLAUDE_SKILL_DIR}/scripts/parse-product-config.py ${CLAUDE_SKILL_DIR}/references/rhoai.yaml --resolve-version "<fixVersion>"

This determines which branch to checkout for each repo.

Step 4: Resolve components to repos

For each component from the ticket:

python3 ${CLAUDE_SKILL_DIR}/scripts/parse-product-config.py ${CLAUDE_SKILL_DIR}/references/rhoai.yaml --resolve-component "<component-name>"

This identifies which source repos are relevant.

Step 5: Clone and collect candidate files

For each context source in the product config, plus each resolved component repo:

bash ${CLAUDE_SKILL_DIR}/scripts/gather-context.sh <repo-slug> <branch> <path-patterns...>

The branch is determined by:

Version resolution from Step 3 (if applicable)
The branch_hint from the context source declaration
Fallback to main

Step 6: Run filtering pipeline

Assemble a JSON input for the filtering pipeline combining:

All candidate files from Step 5
Task context (ticket metadata as keywords, components, version)
Source declarations from product config

python3 ${CLAUDE_SKILL_DIR}/scripts/context-filter.py <<< '<assembled-json>'

The pipeline runs six stages:

Static inclusion — mark always_include sources
Version branch — (already resolved in Step 3)
Component affinity — filter by component overlap
Path relevance — score by path patterns and keywords
Keyword relevance — BM25 scoring against ticket text
Budget enforcement — select top candidates within 100K token budget

Step 7: Read file contents

Source the git utilities (which internally uses scripts/load-env.sh to load credentials from the environment) and read each selected candidate's content from the cloned repo:

source ${CLAUDE_SKILL_DIR}/scripts/git-utils.sh
git_file_content "workspace/repos/<repo-slug>" "<file-path>"

Attach content to each candidate entry.

Step 8: Write context package

Write workspace/context-package.json with this structure:

{
    "ticket": {
        "key": "RHOAIENG-55490",
        "summary": "...",
        "description": "...",
        "fix_versions": ["rhoai-2.18"],
        "components": ["Dashboard"],
        "linked_tickets": [],
        "epic_key": "",
        "status": "In Progress",
        "issue_type": "Story"
    },
    "product": {
        "product_id": "rhoai",
        "display_name": "Red Hat OpenShift AI",
        "docs_repo": "opendatahub-io/opendatahub-documentation",
        "docs_branch": "main",
        "framework": "asciidoc-modular",
        "module_prefixes": {
            "concept": "con_",
            "procedure": "proc_",
            "reference": "ref_",
            "assembly": "assembly_",
            "snippet": "snip_"
        }
    },
    "context_files": [
        {
            "source_type": "documentation",
            "repo": "opendatahub-io/opendatahub-documentation",
            "file_path": "modules/serving/pages/con_model-serving.adoc",
            "content": "...",
            "relevance_score": 0.85,
            "size_bytes": 4521,
            "estimated_tokens": 1130,
            "signals": [...]
        }
    ],
    "metadata": {
        "gathered_at": "2026-04-14T10:30:00Z",
        "total_candidates": 500,
        "selected_candidates": 42,
        "total_tokens": 85000,
        "version_resolved": "2.18",
        "branch_resolved": "release-2.18"
    }
}

Gotchas

GITHUB_TOKEN must be set in the environment; without it, private repo clones and GitHub API calls will fail silently or return incomplete data.
The 100K token budget in the filtering pipeline is a hard cap; if too many repos are relevant, lower-relevance files will be dropped without warning.
Linked ticket resolution is capped at 5; large epics with many linked issues will only get partial context.

Stop conditions

Halt: Jira ticket not found or resolve-jira.sh fails (missing credentials)
Halt: No context sources configured in product config
Warn and continue: Individual repo clone fails (skip that repo)
Warn and continue: Individual file unreadable (skip that file)

Output

Primary: workspace/context-package.json Report to caller: number of files gathered, total tokens, repos cloned.

Agent Skills: doc-gather

Install this agent skill to your local

Skill Files