Skip to main content
On this page

shared/workflows/phase-0-context-gather

Source: shared/workflows/phase-0-context-gather.md

workflow: Phase 0 — context gather

The fan-out step. Every polymorphic skill runs this first. Reads everything referenced in the user’s input.

Inputs

  • Raw user prompt + any URLs / paths / IDs in it.
  • Current working directory (resolves <repo>/.adk/ + <repo>/ai-guidelines/).
  • $ADK_CONFIG_HOME/core.yaml (+ repos.md + connectors/*.md) (workspaces, repos, data sources).

Steps

  1. Run scripts/url_classifier.py on the prompt. Get back JSON: {urls: [...], local_paths: [...], freeform: "..."}.
  2. Pick the working repo: match cwd against repos.md frontmatter repos[*].path. If no match: ask the user to confirm or specify a repo.
  3. Pick the active workspace: derive from the matched repo’s workspace field. If multi-workspace user is ambiguous: ask.
  4. Resolve task-slug + task folder via scripts/adk_task_slug.py --skill <stem> --input <prompt> --create --json. The script picks the right root per shared/paths.md (repo-bound vs global), derives the discriminator from the input (Jira key, PR number, repo+date, etc.), and returns the absolute path. Skills that are always global (pr-review, investigate, setup, improve, explain) pass --scope global; skills that are always repo-bound (implement, document) pass --scope repo; hybrid skills (review, sync) pass --scope auto and let the script decide.
  5. Use the returned task_dir as the working dir for the rest of this skill run. Every artifact path below is relative to it.
  6. Fan-out fetch each URL in parallel (cap 4 concurrent) via the matching MCP. See shared/input-classifiers/<type>.md for each URL type’s fetcher.
  7. Merge results into <task_dir>/context.md. One section per source. Quote ≤15 words per claim; link out for the rest.
  8. Optional RAG enrichment: if rag.enabled: true AND prompt matches rag.trigger_keywords (or user explicitly says “check our internal docs”), query the RAG MCP with the prompt + key entities. Merge results tagged [source: rag].
  9. Load relevant guidelines: based on detected task category (frontend / api / data / observability / security / …), pre-load matching shared/guidelines/*.md into the working context.

Output

<task_dir>/context.md layout:

Markdown
# context for <task-slug>## working repo- name: storefront-bff (from repos.md repos[0])- path: /Users/sujeet/code/acme/storefront-bff- workspace: personal-work## sources fetched### [jira] SF-1234 — coupon enginefetched: 2026-05-18T10:00Zsummary: 2-sentence summarykey fields: status=In Progress, assignee=sujeet, sprint=Sprint 47acceptance criteria:  - …link: https://acme.atlassian.net/browse/SF-1234### [confluence] storefront-bff design docfetched: 2026-05-18T10:00Zsummary: 2-sentence summaryrelevant section: "Coupon engine" — at https://acme.atlassian.net/wiki/...### [github-pr] #456 — prior coupon workfetched: 2026-05-18T10:00Zsummary: …relevant diff: services/coupon.py +120/-30link: https://github.com/...### [rag] internal coupon policy docfetched: 2026-05-18T10:00Zchunks: 3relevance: highquoted: "…" (link)### [skipped] [slack] — credentials file missing tokengap: SLACK_BOT_TOKEN not exported by $SLACK_CREDENTIALS_FILE

Anti-patterns

  • Fetching transitively. Follow links the user gave you. Don’t follow links inside fetched content. One hop.
  • Embedding fetched bodies. Quote tightly, link for the rest. Save tokens.
  • Failing the whole phase on one source. If 4 of 5 sources fetched and 1 failed, report the failure and proceed.
  • Auto-running expensive queries (full DD log dump, Snowflake table scan). The user did not ask for that — only fetch what each URL points at.