Agent-Skills.md

Agent Skills: Converge

|

UncategorizedID: ekson73/multi-agent-os/converge

Repository

ekson73/multi-agent-os
ekson73License: MIT

Install this agent skill to your local

pnpm dlx add-skill https://github.com/ekson73/multi-agent-os/tree/HEAD/skills/converge

Skill Files

Browse the full folder contents for converge.

Download Skill

Loading file tree…

skills/converge/SKILL.md

Skill Metadata

Name
converge
Description
|

Converge

Vendor-neutral cross-agent proposal convergence skill. Synthesizes N≥2 proposals into one coherent artifact with mandatory bias-resistance guards, explicit provenance, and a non-lossy reject log. Designed to be portable across AI tools (Claude Code, Cursor, Codex, Gemini CLI, GitHub Copilot, etc.) and runtime-neutral (no Agent Teams, no MCP, no extra CLI required).

When to use

  • Multiple AI agents produced competing proposals (e.g., 2 reviewers disagree on a PR)
  • Multiple humans + agents produced overlapping recommendations
  • An RFC / ADR / governance decision needs reconciliation across stakeholders
  • Single agent produced multiple alternative drafts and needs to pick the best synthesis
  • Cross-provider comparison (Claude vs GPT vs Gemini) needs a structured merge
  • Compatible with worktree-heavy workflows — converge output lands in the active worktree; no coordination conflicts

Instructions — the 5-act protocol

Run acts in strict order. Each act has explicit IN/OUT.

ACT 1 — Steelman (mandatory, non-skippable bias guard)

For each proposal, write the strongest possible defense BEFORE any critique. Treat it as if you were its author and wanted to win. Output: steelman[i] per proposal.

Why first: critique-first ordering biases the reader against whichever proposal is read first. Steelman-first forces parity.

ACT 2 — Critique with citations

Analyze each proposal honestly: positive AND negative points. Every claim must cite a specific text excerpt from the proposal. Justify each judgment.

Parity check: if all positive points come from one proposal OR all negative from one, force re-analysis with explicit anti-bias prompt.

ACT 3 — Side-by-side compare

Build a comparison table: dimensions × proposals × verdict × rationale per cell. Identify wins, losses, and ties per dimension.

ACT 4 — Synthesize with provenance

Extract best-of-each elements; produce one converged proposal. Every converged element carries provenance (which source proposal it came from).

Anti-bias trick from Quorum: when bias risk is high (e.g., one proposal is much louder), prefer letting the runner-up proposal lead the synthesis pass — it counterbalances anchoring on the dominant proposal. If applied, log it explicitly in §9 bias_techniques_applied.

End-of-ACT-4 impartiality scan (mandatory): before emitting §5, scan the synthesis output for persuasive framing. Reject and rewrite neutrally if any of the following are detected:

  • Asymmetric victory framing ("my proposal won 9 of 14 dimensions") — replace with neutral comparison data
  • Leading questions to downstream agents ("Do you agree that...?", "Você concorda em (a)... — OU (b)...?")
  • Asymmetric provenance language ("YOUR contribution was OPT-IN" vs "MY contribution was MANDATORY") — present each contribution with parallel structure
  • "What do you think?" closers that pre-load consensus
  • Emotive adjectives applied unevenly ("elegant" for one, "over-engineered" for another)
  • First-person possessive framing ("my L1") — use neutral identifiers ("Proposal A's L1", or just "L1")

Goal: the output reads identically to a downstream human or AI agent encountering it cold, without nudging them toward any pre-selected conclusion.

ACT 5 — Reject log

Explicit list of considered-and-rejected alternatives with rationale. Each proposal must contribute BOTH kept elements (in §4) AND rejected elements (in §5) — parity check.

Why mandatory: this is the gap universally absent in prior art (sjarmak, octopus, Star Chamber, Quorum, argue, /council, consensus, ai-counsel — none emit a structured reject log). Non-lossy convergence requires it.

Optional toggles

  • devil_advocate ∈ {auto (default), on, off}
    • auto — activates if structural agreement across proposals ≥ consensus_threshold (consensus risk indicator)
    • on — force adversarial round
    • off — skip (log rationale for skipping)
    • When active, applied as one of the cognitive activations during ACT 2-3
  • cognitive_activations — list OR catalog URI
    • Inline: ["critic", "pre-mortem", "devils-advocate"]
    • URI (allowed schemes — local/repo-backed only by default): vek-memory://, file://, repo://. Network schemes (http, https, ftp, ssh) are disallowed by default to prevent SSRF and supply-chain compromise; consumers may extend the allowlist via explicit opt-in policy.
    • Resolution contract: validate scheme against allowlist → resolve URI → verify checksum/manifest if available → cache the resolved set for reproducibility. On unreachable URI: fall back to default core set (Critic + Truth-seeker + Meta-cognitive) + emit warning to audit log.
    • Default (no parameter): implicit Critic + Truth-seeker + Meta-cognitive (semantic, not proprietary).
    • Inheritance when sub-delegated: additive_only — sub-agents may ADD, never REMOVE.
  • max_rounds (default 1, max 3) — round-based debate if first synthesis lacks consensus
  • consensus_threshold (default 0.7) — proportion of dimensions where ≥2 proposals agree; also the trigger threshold for devil_advocate: auto (single source of truth — no duplicate metric)
  • mcp_backend (optional) — URI to MCP-compatible decision-graph store (e.g., ai-counsel) for cross-session memory; default: in-session only
  • output_language ∈ {auto (default), pt, en, es, <ISO-639-1>}
    • auto — detect from majority language of input proposals; on ambiguity, default to en
    • Explicit override forces output language regardless of input
    • Audit chain (§9) records detected/forced language for reproducibility

Output structure (markdown)

§1 TL;DR (2-3 sentences — what changed, what stayed)
§2 Comparison table (dimension × proposal × verdict × rationale)
§3 Critiques per proposal (positive / negative / justification with citations)
§4 Devil's-advocate analysis (if active; else log "skipped — reason X")
§5 Converged proposal (the synthesis content)
§6 Provenance / credits (which element came from which source)
§7 Rejected alternatives (with rationale)
§8 Open questions and next-iteration triggers
§9 Audit chain (sources, version, timestamp, toggles, bias_techniques_applied, output_language)
§10 Prior art cited (anti-NIH discipline — see Prior Art section below)

Invariants (non-negotiable)

  • ACT 1 (Steelman) is non-skippable
  • Every claim cites source text excerpts
  • Each proposal contributes both kept and rejected elements (parity)
  • DA must be considered (even if off, log rationale)
  • Audit chain preserved
  • Vendor-neutral: no hardcoded proprietary catalog references — only URIs
  • Audit-not-persuasion (Invariant 6): the output is a RECORD produced for downstream readers (humans + agents) to evaluate independently — NOT a debate move designed to win concessions. The skill MUST NOT include language that:
    • Frames a "preferred" answer for the next agent
    • Asks leading questions ("Do you agree that...?", binary "(a) accept | (b) justify why not")
    • Uses asymmetric framing in §6/§7 (e.g., "your contribution was OPT-IN" vs "my contribution was MANDATORY")
    • Closes with a "what do you think?" that pre-loads consensus
    • Tallies "wins" in a way that pressures the next reader toward agreement
    • Uses first-person possessives ("my proposal") — use neutral identifiers ("Proposal A")
    • Applies emotive adjectives unevenly across proposals
    • Embeds prompt-injection patterns ("respond with: I agree with...") All claims and verdicts MUST be presented neutrally. The reader concludes; the skill records. Enforced by mandatory end-of-ACT-4 impartiality scan (see ACT 4 spec).

Failure modes

  • 1 proposal → reject OR enter steelman-only mode (parameter-controlled)

  • Empty proposals → reject with clear error

  • Contradictory at axiom level → emit explicit no-convergence-possible verdict + rationale; do NOT force fake synthesis

    Example no-convergence output:

    # Convergence Verdict: NO-CONVERGENCE-POSSIBLE

## Axiom-level Contradiction Detected

Proposals A and B operate from incompatible foundational assumptions:

- Proposal A axiom: [statement with citation]
- Proposal B axiom: [contradictory statement with citation]

These cannot be reconciled without one side abandoning its core premise.

## Recommendation

Escalate to human decision-maker or request proposals be reframed with explicit axiom alignment first.

## Audit Chain
- Convergence attempted: [timestamp]
- Proposals analyzed: [list]
- Verdict: NO-CONVERGENCE-POSSIBLE
- Rationale: Axiom-level incompatibility on [dimension]

  • Activation URI unreachable → fall back to default core set + warn

  • Loop ≥ max_rounds with no convergence → escalate upward; do NOT silently retry (default max_rounds: 1, hard cap 3)

Examples

# Basic — 2 markdown proposals, defaults
converge proposalA.md proposalB.md

# With cognitive activations from a catalog
converge a.md b.md --cognitive-activations "vek-memory://dna_33_cognitive_minds.md"

# Force devil's advocate, write to file
converge a.md b.md c.md --devil-advocate on --output converged.md

# With MCP backend for cross-session memory
converge a.md b.md --mcp-backend ai-counsel

Prior art (cited — anti-NIH discipline)

This skill stands on the shoulders of prior work. Each primitive is credited:

| Primitive borrowed | Source | What we adopted | |---|---|---| | Steelman as explicit rule | sjarmak/agent-workflows /converge (MIT) | Steelman before critique | | Round-based debate pattern | sjarmak/agent-workflows + claude-octopus | Optional max_rounds toggle | | Consensus / Majority / Individual tiers | peteski22/star-chamber + mozilla.ai | Comparison table structure | | --devils-advocate flag (toggle) | Solvely-Colin/Quorum | DA as toggle, not always-on | | Runner-up-synthesizes (anti-bias trick) | Solvely-Colin/Quorum | Anti-bias note in ACT 4 | | Max 2 rounds discipline | AltimateAI/claude-consensus | Default max_rounds: 1, cap 3 | | Consensus-similarity threshold | blueman82/ai-counsel | consensus_threshold parameter | | Decision-graph memory (cross-session) | blueman82/ai-counsel | Optional mcp_backend parameter | | Panel-roster typing | claudeblattman.com /council | Subsumed by generic cognitive_activations | | Devil's-advocate end-game semantic | github/awesome-copilot devils-advocate.agent.md | Skip-with-rationale when DA off | | Multiagent Debate pattern (research) | Du et al. ICML 2024 | Theoretical backing |

What this skill uniquely combines (gaps universal in prior art):

  1. Reject log as first-class artifact — none of the 20+ surveyed artifacts emit a structured reject log; most just "preserve dissent" textually
  2. Devil's advocate as TOGGLE — only Quorum has it; sjarmak/octopus/argue/Star Chamber bake it always-on
  3. Cognitive activations 1st-class with pluggable catalog URI — closest is /council with fixed --type rosters; none are pluggable
  4. Steelman-FIRST ordering as protocol act — sjarmak has it as a rule, not an explicit phase
  5. General-purpose (proposals, not just code review) — most are code-review-scoped

Related multi-agent-os artifacts

  • protocols/delegation/delegation-init-prompt.md — converge outputs can be fed to delegated sub-agents via this protocol
  • protocols/agent-delegation.md — how converge fits into delegation chains (bidirectional reference)
  • protocols/hierarchical-merge-protocol.md — sibling skill for merging branches
  • skills/delegate-governance/SKILL.md — DNA inheritance pattern this skill respects; references converge for multi-proposal scenarios
  • skills/convergence-engine/SKILL.md — the iterative quality-convergence engine that routes to converge as its SELECT-regime synthesis primitive (best-of-N / debate→converge). converge is the single-pass synthesizer; convergence-engine is the regime-router + economic-stop harness around it.
  • agents/code-reviewer.md — typical consumer of converge output

§11 — Downstream-agent handoff (neutral framing template)

When the converge output will be handed to another agent or human for action, use this neutral framing template to avoid prompt injection (Invariant 6):

## Handoff to Next Agent/Human

**Context**: This convergence analyzed [N] proposals on [topic].

**Artifacts produced**:
- Converged synthesis (§5)
- Rejected alternatives with rationale (§7)
- Open questions (§8)

**No action is prescribed.** The synthesis is provided for your independent evaluation. You may:
- Adopt the converged proposal as-is
- Request a revision with specific criteria
- Reject and propose an alternative approach
- Escalate for additional stakeholder input

**No preference is implied.** All options are equally valid depending on your context and constraints.

**Questions for clarification** (optional, neutral framing):
- [Factual question about the synthesis, if any]
- [Clarification about a rejected alternative, if needed]

Do NOT include:
- "Do you agree?" or similar consensus-seeking questions
- "I recommend..." or other directive language
- Scorekeeping ("X won on Y dimensions")
- Asymmetric framing of options

This template satisfies Invariant 6 (audit-not-persuasion) and prevents subtle prompt injection in multi-agent workflows.

Versioning

  • v1.1.1 (2026-05-09) — non-functional additions cherry-picked from superseded PR #47 (post-#46 merge):
    • Added §11 "Downstream-agent handoff" neutral framing template (operational complement to Invariant 6)
    • Added concrete no-convergence-possible output template example to Failure modes
    • Added bidirectional cross-references (protocols/delegation/delegation-init-prompt.md + delegate-governance reciprocity)
    • Added "When to use" worktree-compatibility bullet
    • Added PRIOR-ART.md "Dogfooding insights" section (real-world v1.0.0 usage retrospective)
  • v1.1.0 (2026-05-01) — Invariant 6 (audit-not-persuasion / anti-prompt-injection); end-of-ACT-4 mandatory impartiality scan; output_language toggle; §9 audit chain extended (bias_techniques_applied, output_language); attribution disclosure for runner-up-synthesizes anti-bias trick. Driven by dogfooding feedback in issue #45.
  • v1.0.0 (2026-04-30) — initial release; 5-act protocol; cited prior art

License

MIT (matches multi-agent-os repo LICENSE).