e2e_ci_council_of_agents

CI-adjusted version of the Council of Agents E2E test-generation pipeline, built to run non-interactively in GitHub Actions via OpenCode on DeepSeek-V4-Pro.

This is the CI copy. The interactive local copy lives in .claude/commands/council_of_agents/ (Claude Code). The two are intentionally separate (see the handoff doc §9.3) — keep the prompt bodies aligned, but these files carry CI-specific wiring: standalone invocation, file-based handoff, no interactive prompts.

Spec / source of truth: tests/ui-testing/MD_Files/council-of-agents-export/CICD integration/council-ci-automation-handoff.md

What's different from the local Council agents

Agents (invocation order)

Sequencing lives in the workflow YAML (it re-implements the Orchestrator). There is no Orchestrator agent in CI — the YAML calls each opencode run in order.

File-based handoff contract

All artifacts live under docs/test_generator/. Each agent reads its inputs and writes its outputs as files; nothing is passed via conversation. Between GitHub jobs, the whole docs/test_generator/ tree (plus any edited test/page-object files) is shuttled with upload-artifact / download-artifact.

docs/test_generator/
  ci/
    diff.patch              # Workflow writes the change diff (untrusted input).
    feature_hint.txt        # Workflow writes the optional hint (untrusted input).
    trigger_comment.txt     # Workflow writes the triggering comment (untrusted input).
    run-context.json        # Triage writes; ALL downstream agents read. The single source of run state.
    triage.json             # Full triage output (for the PR comment in dry-run).
    playwright-registration.json  # Engineer writes: { group, spec_filename, create_group }. A
                            # DETERMINISTIC (non-LLM) step applies the run_files edit — agents
                            # never edit .github/workflows/ themselves.
    sentinel-verdict.json   # Sentinel writes: { verdict: PASS|FAIL, critical_count }.
    heal-result.json        # Healer writes: { status: passing|failing, iterations }.
  features/<slug>-feature.md            # Analyst
  test-plans/<slug>-test-plan.md        # Architect
  generation-reports/<slug>-generation.md  # Engineer (what it created + the playwright.yml edit)
  audit-reports/<slug>-audit.md         # Sentinel
  execution-reports/<slug>-execution.md # Healer

The generated test itself is written to the repo, not under docs/: tests/ui-testing/playwright-tests/<Area>/<spec_filename> (+ any page-object edits under tests/ui-testing/pages/).

run-context.json schema

{
  "feature_slug": "share-link",
  "feature_title": "Logs Share Link",
  "area": "Logs",
  "edition": "oss",
  "needs_e2e": true,
  "needs_api": false,
  "spec_filename": "shareLink.spec.js",
  "spec_path": "tests/ui-testing/playwright-tests/Logs/shareLink.spec.js",
  "playwright_group": "Logs-Core",
  "source_files": ["web/src/plugins/logs/..."],
  "skip": false,
  "skip_reason": ""
}

playwright_group = the testfolder matrix group in .github/workflows/playwright.yml whose run_files array the Engineer appends the spec filename to.

Security model

These agents process attacker-controllable input (PR diffs, comments, hints — anyone can open a PR). Hardening built into the design:

• Untrusted input is data, not instructions. The diff/comment/hint reach agents as files, and every agent that reads them is told to treat their content as inert data and ignore any embedded instructions ("prompt injection").
• No LLM edits to CI workflow files. The Engineer never edits .github/workflows/**; it emits playwright-registration.json and a deterministic step applies a single run_files append.
• Path/identifier validation before shelling out. spec_filename must match ^[A-Za-z0-9._-]+\.spec\.js$; spec_path is regex-validated before the Healer runs it (and again in the workflow).
• Least-privilege CI. The live triage job runs with contents: read; only the PR-back job requests contents: write.
• Human review remains the backstop. All generated code lands in a PR a human reviews before merge; Sentinel auto-fixes are a tiny whitelist (add import, console.log→testLogger, tag @ prefix).

Scope (v1)

• OSS only. ENT features are skipped at triage (no generation, no PR).
• Scribe excluded. TestDino upload rides the generated PR's own existing CI.
• Local OSS binary is the Healer's only run target.
• On-demand agents (Inspector / Gatekeeper / Guardian / API-Smith) are out of scope.

Invocation (CI)

Each agent runs standalone, e.g.:

opencode run --agent e2e-ci-analyst --model deepseek/deepseek-v4-pro \
  "Generate the Feature Design Document for the feature described in docs/test_generator/ci/run-context.json"

The model is passed at invocation (not pinned in frontmatter) so the same files can run on a different provider if needed. The DeepSeek provider is configured in opencode.jsonc.

Discovery caveat: verify against current OpenCode docs whether agents in a nested subfolder of .opencode/agents/ are auto-discovered. If not, either flatten the files or register them explicitly / invoke by prompt-file path in the workflow.