converge

name	converge
description	Cartographer-scoped forward-projection synthesis that ranks the dev corpus into a "what's next" menu. Applies vision×readiness scoring, memorial-tier safeguards, and search-bias compensation to convert the memkit reports (cleanup-backlog-refresh, dedupe-clusters, sprint-digest, path-update-proposals) into per-theme plan edits and a session-start handoff. Runs after librarian hygiene has produced clean inputs. Use when the operator asks "what's next?", before a major /shift kickoff, or when planning a roadmap pass. The output is a ranked menu, not a maintenance pass.
metadata	{"sourceRuntime":"claude","master":"package","governance":"epr:elohim-agent/skills/converge"}

Converge — Trajectory Synthesis + "What's Next" Menu

Extracted from memory-kit (May 2026). The hygiene tools (cleanup, path-update, dedupe-memory, skill-audit, memory-review) handle maintenance. This skill handles synthesis: turning the reports those tools produce into actionable plan edits and a ranked next-actions handoff.

Spec: genesis/docs/superpowers/specs/2026-05-10-converge-skill-design.md.

When to invoke

• After a memory-kit weekly sweep, before a major /shift or /deliver
• When an operator types "what's next?" at session start (the menu surface this skill produces is the canonical answer)
• Never in isolation — convergence-themes.md is computed from memory-kit reports; running converge without fresh reports produces a stale menu

Four phases

python3 .claude/scripts/converge/converge-scan.py     # Phase 1: deterministic theme detection
# (Phase 2 synthesis subagent dispatched — see prompt below)
python3 .claude/scripts/converge/converge-apply.py    # Phase 3: operator-approved plan edits
# Phase 4: session-start UX reads the produced next-actions.md

1. Theme detection (deterministic, converge-scan.py) — clusters BACKLOG items, dedupe pairs, plan-status cooling items, sprint-digest themes, and path-rename clusters by shared theme keywords (extracted from active plan filenames + filtered through stopwords + multi-source-type requirement). Outputs .claude/memory-kit/<TODAY>/convergence-themes.md with each theme's contributing items and a candidate canonical plan.

2. Synthesis (LLM judgment, subagent dispatch) — when this skill is invoked after a fresh scan, dispatch a general-purpose subagent (opus) with the prompt below. Produces:

   • Per-theme proposals at .claude/memory-kit/<TODAY>/converge/<theme>-proposal.md with structured edit blocks (mark-done, add-as-outstanding, merge-redundant, remove-obsolete, surface-question) — each operator-reviewable via - [x] Accept checkboxes.
   • The next-actions menu at .claude/memory-kit/<TODAY>/next-actions.md — top 3-5 ranked recommendations with vision × readiness scoring and pre-authored Objectives.

3. Apply (deterministic, converge-apply.py) — for each accepted edit, performs the plan modification. v1 supports mark-done (- [ ] → - [x] with evidence comment) and add-as-outstanding (insert new - [ ] after named anchor); other kinds are operator-manual edits in v1.

4. Session-start UX — when a human says "what's next?", read the latest next-actions.md, present conversationally, hand off to /shift or /deliver with the chosen pre-authored Objective. No new skill needed; this is a convention.

Phase 0 — ROADMAP + MAP currency (runs first, before theme detection)

/converge is the cartographer's seat, and the cartographer owns the standing prioritization home — genesis/data/timeline/roadmap/vision-readiness-sprint-roadmap.md. Before producing a "what's next" menu, regenerate that roadmap from the three live inputs (placement-audit.py --ledger × --focus × the re-mined vision axis) per the cartographer's ROADMAP-CURRENCY mandate. A stale roadmap poisons the menu it feeds, so this runs first:

1. Roadmap regeneration (cartographer, standing duty). Re-rank the sprints against today's ledger + cluster-state + re-mined gospel-tier #1 axis; re-stamp the dated regeneration header even if the ranking is unchanged; move newly-AVAILABLE work into a sprint and newly-degraded work out to §3 (never rank BLOCKED-BY-ENV). The roadmap's §4 single-highest-leverage-move becomes the top of the next-actions menu — they must agree.
2. MAP-currency glance (librarian + historian, the LEGIBILITY/PATH co-owners). Check the map-drift accumulator (.claude/memory-kit/map-currency-drift.json, written by map-drift-signal.py, surfaced at the SessionStart budget headline): did architecture seeds or pillar guides move without architecture/MAP.md following? If so, that is a LEGIBILITY drift — surface it in the menu's "Quiet but load-bearing" section as a structural-hygiene candidate (the owning agents apply the fix per their authority split). MAP currency is not the cartographer's to fix, but a drifted walk-path is a legibility cost the menu should name so onboarding doesn't silently rot.

This Phase 0 is the in-flight maintenance half of /converge; the four phases below are the synthesis half. Both run each pass.

Synthesis subagent prompt (Phase 2 dispatch)

You are the synthesis phase of /converge. Read .claude/memory-kit/<TODAY>/convergence-themes.md
to see the themes detected from the latest memory-kit dreaming pass.

For each theme above noise (skip themes with low total signal or that look like
generic vocabulary like "execution", "peer" alone — focus on substantive ones
like "iroh", "epr", "recovery", "doorway", "topology", "storage"):

PER-THEME WORK
==============

1. Read the theme's canonical plan candidate fully.
2. Read each linked BACKLOG item's source spec/plan.
3. `git log --since="60 days ago" --oneline` for the theme's repo paths.
4. Search recent sprint-results in .claude/shifts/ for theme mentions.
5. Read .claude/data/dev-intent.jsonl entries (if exists) mentioning the theme.

Produce a per-theme proposal at .claude/memory-kit/<TODAY>/converge/<theme>-proposal.md
with structured edit blocks. Each edit block has form:

  ### N. `<kind>` — <one-line description>
  **Plan**: `<canonical-plan-path>`
  **Edit**:
  ```edit
  target: <text of the existing checkbox to mark done>
  evidence: <citation: file paths, commit hashes, scenarios>

• Accept

Edit kinds: mark-done — checkbox in plan should be checked (target = checkbox text; evidence = citation) add-as-outstanding — new outstanding item should be added (anchor = section heading; new_item = text) merge-redundant — sections to collapse (manual review in v1; describe in proposal) remove-obsolete — task no longer relevant (manual; describe + cite) surface-question — open question to add to plan (manual; cite source)

Be conservative on mark-done: only mark when the deliverable is unambiguous (file exists at expected path; scenario passing; commit message references the task).

NEXT-ACTIONS MENU

After per-theme proposals are written, produce .claude/memory-kit//next-actions.md. This is the "what's next" surface.

For each refreshed plan that's ready-to-execute, score:

VISION-ALIGNMENT (0-10) — does completing this plan advance the manifesto's design principles?

Read genesis/docs/content/elohim-protocol/manifesto.md (Part II Design Principles 1-6: Distributed Architecture / Graduated Intimacy / Values Alignment / Community Governance / Wealth as Circulation / Living Memory). Walk each principle, mark YES/SOMEWHAT/NO/TANGENTIAL for whether the plan's deliverables advance it.

Score guide: 9-10: directly advances 3+ principles substantively (substrate-level work) 7-8: directly advances 1-2 principles substantively (cross-cutting feature) 5-6: enables Principle 1 (infrastructure / tooling) 3-4: hygiene / operational 1-2: internal-tooling-only

Substrate work multiplies. Tactical work is linear. Cite which principle(s) the plan advances in the reasoning.

READINESS (0-10) — how close is this plan to "an agent ships in 1-2 cycles"?

Signals (additive): Worktree exists with partial work: +3 All blockers explicitly resolved or absent: +3 Open items well-scoped + ordered (mechanical work, not design): +2 Recent commit activity in adjacent area: +1 No pending operator decisions: +1 Plan has clear next step at top of unfinished list: +1

Subtractive: Has unresolved design questions in Open Questions section: -2 Requires external dependency (review, vendor, governance): -2 Plan has no checkbox structure (narrative-only): -3

RECOMMENDATION = vision × readiness, sorted descending.

When picking the TOP recommendation, prefer concrete gates over master plans — master plans are too coarse to be a single Objective; recommend the next gate the master plan unblocks. When scores tie, prefer the plan closer to "ships in 1-2 cycles" over "ships in 5 cycles, more vision impact" — closes the loop faster.

ALIGN WITH THE MAINTAINED ROADMAP. The menu is the transient, session-shaped projection of the persistent prioritization home at genesis/data/timeline/roadmap/vision-readiness-sprint-roadmap.md (regenerated in Phase 0 above). The menu's TOP recommendation MUST agree with that roadmap's §4 single-highest-leverage-move; if your fresh scoring disagrees, that is signal the roadmap regeneration is stale — reconcile (re-rank the roadmap) before emitting the menu, do not silently diverge. The roadmap is the durable ranked surface; the menu is its session read-out with pre-authored Objectives. They are one ranking, two readouts (the same "one loop, two readouts" shape as the unified memory loop).

Output format (top 3-5 entries):

Top recommendation:

• Plan: <path>
• Vision-alignment: N/10 —
• Readiness: N/10 —
• Pre-authored Objective: <1-2 sentences, action-oriented, ready to drop into /shift>
• Estimated cycles: 1-2 / 2-3 / 4+
• Recommended skill: /shift or /deliver
• Blockers (if readiness < 8):

Other ready plans

... (rank 2-5)

The menu is what the human reads at session start. Make every line load-bearing. Total length: ≤80 lines for top recommendation, ≤30 lines per other entry.

When done, print: counts of themes synthesized, proposals written, plans surfaced in the menu, and the path of next-actions.md.

MEMORY-DAMAGE SAFEGUARDS (read before proposing any mark-done, merge, or remove)

The biggest failure mode of this skill is aggressive consolidation that loses gold. The opposite failure (hoarding) is also bad, but it's a slow leak; aggressive consolidation is irreversible damage. When uncertain, default to PRESERVATION, not action.

Concrete safeguards:

1. Memorial-tier check before any mark-done or remove. Before proposing to mark a checkbox done OR remove an item as obsolete, search for citations of that item or its theme in:

   • genesis/docs/content/elohim-protocol/ (manifesto, epics, constitution)
   • The latest two manifesto-tier MEMORY entries If cited there, the item is foundational-adjacent. Use OPERATOR-CALL with reasoning instead of auto-proposing the change. Note: "high citation in vision documents — operator may want to memorialize before any cleanup."

2. Convergent-insight respect. When dedupe-clusters surfaces same-concept entries from different conversations or sprints, do NOT default to "merge." The rediscovery itself is signal — the insight is robust enough to surface multiple independent times. Default action: leave both, propose only as OPERATOR-CALL with the framing "convergent insight — multiple independent reach for the same point; operator may prefer to memorialize as a principle rather than merge as duplication."

3. Quiet-but-foundational surface (additive section in next-actions.md). In addition to the top 3-5 ranked recommendations, include a separate section titled "## Quiet but load-bearing." This lists items that the search ranks LOW but that you (the synthesis subagent) judge as important:

   • Plans untouched in >60 days but referenced in manifesto/epics
   • Memory entries with low DF but high vision-alignment
   • Specs whose deliverable hasn't shipped but whose principle is cited elsewhere This counters the search's recency-and-frequency bias. The operator may want to refresh attention on these even if they don't dominate the menu.

4. High-DF promotion warning. If you notice during investigation that a theme has hit or is approaching the DF auto-stopword cap (>30% of docs), flag it explicitly: "{theme} is approaching DF saturation — convergence will lose visibility soon. Recommend memorialization to manifesto/epic tier before that happens." This addresses the reflexive bias where successful themes become invisible.

5. Self-aware uncertainty disclosure. For any proposal where you are not confident (mixed evidence, ambiguous deliverable status, principle-level work that doesn't fit checkboxes), use OPERATOR-CALL with explicit reasoning. Do not force a verdict to fill the proposal slot. Empty proposals are better than wrong proposals; the operator can always run another cycle.

6. The lifetime-memory principle. Humans keep certain memories for life regardless of recency or frequency — formative experiences, foundational relationships, things that shaped who they are. The protocol equivalents are manifesto principles, vision statements, and explicitly memorialized work. These should NEVER be merged, archived, or marked-done by automatic logic — only by explicit operator action with reviewer pass. If you find yourself proposing changes to manifesto-tier content, STOP and reframe as OPERATOR-CALL with maximum caution.

7. Default toward preservation when uncertain. Cleanup is reversible in principle (archived items can be restored) but expensive in practice (operator attention is finite; once attention moves on, archived gold stays buried). Merge is structurally more dangerous (lossy by design; per the lifecycle spec, c ≈ 0.6-0.95 of inputs). When the evidence is ambiguous, propose nothing rather than propose merge. Empty per-theme proposals are valid output.

These safeguards apply across all five edit kinds (mark-done, add-as-outstanding, merge-redundant, remove-obsolete, surface-question). The skill exists to make the corpus more legible, not to compress it for compression's sake. Be conservative; the operator can always tell you to be more aggressive next cycle.

## Biases of the search approach

`converge-scan.py` uses classical IR (TF-IDF + bigrams + DF auto-stopwords + TextRank co-occurrence centrality) over a corpus of plan filenames + frontmatter + section headings + memory titles + sprint titles. **Consuming skills (especially the Phase 2 synthesis subagent) must understand its biases** — both to compensate when investigating, and to author future content that stays legible to the search.

**Surfacing biases:**

| Bias | What it means | What to compensate |
|---|---|---|
| **Filename + heading weighting** | A theme must appear in plan filenames or section headings to surface. Themes living only in body text are invisible. | When reading a canonical plan, check whether topical concepts in the *body* match the filename/heading themes; flag mismatches as "search-invisible work" needing operator review. |
| **Bigram boost (1.6×)** | Multi-word phrasal themes get a score boost; but their constituent unigrams often dominate edges. | Don't assume bigrams = most important. Investigate sub-themes inside top unigrams. |
| **TF-IDF + DF auto-stopwords** | Pervasive vocabulary (>40% of docs) drops out as noise. Once a concept becomes universally adopted, it stops surfacing as a theme. | Treat the DF cap as a signal: high-DF terms are *successfully integrated*, not *unimportant*. Cross-reference with manifesto/epics to confirm. |
| **TextRank centrality** | Themes in dense co-occurrence clusters score higher; isolated themes lose. | Important-but-isolated work may rank low. Do not equate "low convergence rank" with "unimportant." |
| **Multi-source-type requirement (≥2)** | Themes must appear in 2+ memory-kit report types. New themes in only one report don't surface. | First-touch themes are invisible until they propagate. Explicitly check sprint-digest's recent-themes section for things that haven't yet surfaced in convergence-themes. |
| **No semantic similarity (no embeddings)** | Synonyms (`recover` ≠ `recovery`) don't cluster. | Manually fold related themes if they share a canonical plan. |

**Reflexive biases** (today's writes shape tomorrow's search):

1. **Filename choice = future theme visibility.** Be deliberate; the filename is the strongest signal converge sees.
2. **Heading discipline = topical legibility.** Use specific section headings, not narrative-only docs.
3. **Successful delivery = silent disappearance.** Correct for tactical work; *wrong* for principles. Counter via memorialization — promote completed-but-foundational work to manifesto/epic content.
4. **Convergence = DF inflation.** As more docs cite a theme, its DF rises; eventually it hits the cap and disappears. Counter: when a theme starts appearing in >30% of new docs, promote it upstream into manifesto-tier permanence.

## Boundary

- Only modifies plans (after operator approval). Never edits specs, memory entries, sprint-results, or skills.
- Never invents tasks — every add-as-outstanding cites a source.
- Reads from `.claude/memory-kit/<date>/` (shared report dropoff with memory-kit) and writes its own outputs there. The output directory naming is legacy from when converge lived inside memory-kit; rename is follow-up scope.
- Hands off via the `next-actions.md` convention to `/shift` or `/deliver`. Not autonomous.

## Related

- `.claude/skills/memory-kit/SKILL.md` — produces the reports converge consumes
- `genesis/docs/superpowers/specs/2026-05-10-converge-skill-design.md` — the design rationale and end-state vision
- `genesis/docs/content/elohim-protocol/architecture/2026-05-10-memory-lifecycle-design.md` — the lifecycle primitives this skill operates within
- `genesis/data/timeline/roadmap/vision-readiness-sprint-roadmap.md` — the standing PRIORITIZATION home the cartographer regenerates in Phase 0; the next-actions menu is its session read-out
- `genesis/docs/content/elohim-protocol/architecture/MAP.md` — the LEGIBILITY/PATH walk; its map-drift accumulator surfaces in the menu's "Quiet but load-bearing" section when seeds move without the map following