agent-skills-governance

Agent-Skills Governance

This skill is bundled with AlphaCouncil so every installer gets the same guardrails. It follows the addyosmani/agent-skills pattern: a skill is an executable workflow with gates, anti-rationalizations, and exit criteria, not a vague style guide.

Source inspiration: https://github.com/addyosmani/agent-skills

Mandatory Use

Use this skill whenever alphacouncil-agent is invoked or when editing the plugin's research/reporting behavior. Do not require the user to install addyosmani/agent-skills separately.

Stop Gates

Before giving a final investment answer, all gates below must be satisfied:

1. Selection gate: for every full or quick council run, call begin_council_selection with the intended council_mode, show display_markdown with the complete returned catalog expandable (number, identity, method and best_for), obtain this run's submission, and call confirm_master_selection with display_ack: true. Full accepts 1..N/ranges/IDs/all; quick accepts only 1-4 methods and forbids all. Do not create a run, fetch evidence or spawn a worker until it returns a selection_receipt. Data-only screen, options, news and market calls are exempt.
2. Receipt gate: pass the one-use selection_receipt with the same symbol, prompt, language and council_mode used at selection. The receipt is mode-bound: never use a full receipt for quick or a quick receipt for full, and never override it with masters or masters_roster. Restart selection if it is missing, stale, expired, consumed or mode-mismatched.
3. Scope gate: infer ticker, language, horizon, and user goal, or ask only the minimum blocking question.
4. Evidence gate: full requires every planned evidence role to complete; after the bounded parse repair, any mandatory failure stops the run before masters/debate/PM and produces incomplete. Quick fixes four roles (market_data, earnings_deep_dive, valuation_long_short, news_industry_management) and may terminate degraded only when at least two completed and every failed role has a recorded diagnostic/data gap.
5. Source gate: every material claim maps to scoped source IDs like <task>:S1; never cite bare S1.
6. Master gate: every master named in the confirmed receipt has either reported or produced a deterministic out_of_scope result, every result has a readable final statement, and every stable ID appears in the published Master Bench. An unexecuted or hidden selected seat keeps the run incomplete or needs_revision.
7. Debate gate: full requires all three bull/bear rounds, exact round-3 Q&A and PM. Quick requires one parallel statement round and PM; one degraded side is allowed only if the other side and PM completed, while two degraded sides or PM failure is incomplete.
8. Verification gate: quick still enforces scoped source-ID presence but never claims the adversarial verifier fan-out ran. Full/deep verification claims must match the recorded runtime scope.
9. Report gate: final_report.md, user_response.md, artifact_index.md, report_quality.json, and per-analyst/method/debate Markdown files are written against full_v2 or quick_v1, never a hybrid.
10. Quality gate: if the applicable report_quality.json is not passed, report needs_revision, not complete. A quick passed result never satisfies full_v2.
11. Handoff gate: chat may be concise, but it must state run status, contract, rating, winner, confidence, valuation/position, material gaps and saved file locations. Full also carries forward setup and invalidations. Quick names all four analyst statuses, selected method IDs/stances, dated recent company/industry news, every degraded item, and that it is not equivalent to full council. Full handoff ends with the exact selected seat count and every complete, untruncated per-seat method statement. A failed seat is retained in that final ledger as statement_status=not_produced, never as a fabricated stance. Use returned user_response_markdown rather than an ACK-only or manually authored recap. If a visible barrier fails, call finalize_visible_run first so that handoff exists.
12. Deadline gate: plugin-managed quick uses headless analyze_symbol; plan_visible_run quick is rejected. Never extend its end-to-end deadline above 600000 ms or silently omit work to appear complete.