| name | atlas-scatter |
| description | Phase L sub-mode (Scatter-Gather Locate) — TRANCE-gated parallel fan-out of read-only Locate probes across topologically-disjoint module clusters, each in a clean-context subagent, merged back into findings.md via a deterministic dedup contract. Never default; serial Locate is the standard path. Use ONLY when map.md surface is large (>5 modules OR >25 files) AND ≥2 disjoint DECISION_TARGET sub-questions exist. Trigger phrases — "audit across all modules", "trace these N independent flows", "scout this large surface in parallel". Do NOT load for a single sub-question, tightly-coupled sub-questions, or a tight budget. |
| allowed-tools | search_symbol graph_query search_text view_file test_dry_run memex_read |
| metadata | {"methodology":"ATLAS","phase":"L"} |
SKILL: Scatter-Gather Locate (Phase L sub-mode)
When to use
Load when: map.md is complete AND the activation trigger below fires.
This is the operationalized G1 TRANCE form of the diffuse scatter primitives
already in skills/locate.md (Operator pattern) and agent.md P0 rule 7.
Unload when: every sub-question has ≥1 merged finding at confidence ≥ M
or a GAP, and the merged findings.md has entered the Phase A fold.
This mode does not add capability — it adds parallelism. It is still
read-only (I-1). Scatter never grants a write tool, never resets the budget,
and never escapes a refusal gate. Branch budgets sum to ≤ the parent budget.
Activation trigger (both-flags rule — TRANCE-gated)
Scatter-Gather is GATED, never default. Standard-tier Locate stays serial.
Both flags MUST hold (cortex C5 + C6). If either is false, the mode is inert
and you run the serial probe ladder in skills/locate.md.
| Flag | Source | Threshold |
|---|
| Surface size (C5) | map.md MAP-MODULES / file count | > 5 modules OR > 25 files in scope |
| Disjoint sub-questions (C6) | mission.md DECISION_TARGET decomposition | ≥ 2 topologically-disjoint sub-questions (different modules / concerns) |
Either flag alone → stay standard tier (serial Locate). Both flags →
escalate to the Scatter-Gather sub-mode.
When NOT to scatter (carry-through from locate.md)
- Only one sub-question exists.
- Sub-questions are tightly coupled (one's answer feeds another's probe).
- Budget is already tight — coordination overhead outweighs the parallel win.
In any of these, do not scatter. Run serial Locate.
Fan-out plan (bounded, deterministic partition)
- Derive the partition deterministically. Run a single parent-side
graph_query("callgraph_slice: <scope>") (or, if unavailable, partition by
MAP-MODULES centrality clusters from map.md). Do not LLM-guess the
clustering — the partition is a structural fact, not an inference
(see skills/locate.md Tier-2 graph-first decomposition).
- One sub-mission per disjoint cluster, mapped 1:1 to a disjoint
sub-question where possible.
- Hard cap: 5 branches (cortex C1; orchestrator-worker sweet spot ~5).
If the partition yields > 5 clusters, merge the lowest-centrality clusters
until N ≤ 5.
- Per-branch budget =
parent_remaining / N. Budgets are partitioned,
not multiplied. The sum of branch budgets MUST NOT exceed the parent's
remaining budget.
Diversity is scope-diversity, not redundancy. Branches probe different
modules. Do not spawn N identical probes of the same module — quality of
the partition dominates redundant diversity.
Sub-mission spec (Operator pattern, clean context)
Each branch is an ephemeral clean-context subagent seeded with ONLY its
scope-slice + one sub-question + its budget. The clean context is the point:
it prevents trajectory self-conditioning (no branch sees another branch's
path). Reuses the locate.md sub_mission YAML, tightened:
sub_mission:
parent_mission_id: <MISSION-ID>
branch_id: <1..N>
question: <one disjoint sub-question>
scope: <path globs for this cluster only>
budget:
max_tool_calls: <parent_remaining / N>
max_tokens_input: <parent_remaining / N>
map_excerpt: <ONLY the MAP-MODULES + MAP-GRAPH slice for this cluster>
Each branch runs the Tier 1–5 probe ladder (skills/locate.md) within its
scope and returns exactly one structured object — no transcript (I-4):
return:
branch_id: <1..N>
findings: [FINDING-XXX, ...]
gaps: [GAP-XXX, ...]
ruled_out: [<negative result notes>, ...]
telemetry: {tokens_used, tool_calls}
The parent never ingests branch transcripts. Branch working state is destroyed
on return.
Merge + dedup contract (two-stage bounded gather)
Stage 1 — Gather & dedup (parent-side, deterministic)
- Renumber to global IDs. Branch-local
FINDING-XXX / GAP-XXX ids are
renumbered to a global namespace to avoid cross-branch collision.
- Dedup on mechanical fields first. Merge two findings into one when they
match on (path, overlapping line-range); use claim-equivalence only
as a tiebreak. The merged finding keeps the highest confidence tier and
unions the evidence anchors and
ruled_out notes.
- Contradictions are NOT silently merged. When two branches make
conflicting claims about the same anchor, emit a single finding marked
[DISPUTED] (D7) carrying both claims, and drop it to confidence L
pending a follow-up probe. Do not pick a winner mechanically.
Stage 2 — Fold (reuse existing machinery)
The merged findings.md flows unchanged into the existing Phase A
clean-context fold (skills/abstract.md). No new aggregator is built —
the AgentFold subagent already aggregates; scatter merely feeds it a merged
list instead of a serial one.
Stop condition (explicit — D5)
Scatter completes when every sub-question has ≥1 merged finding at
confidence ≥ M OR a GAP. Scatter does not re-spawn: max recursion
remains 1 (agent.md P0 rule 9). A branch that returns only L/empty triple
records a GAP; the parent does not re-scatter that cluster.
Telemetry
Record at the gather edge:
scatter | branches: N | fan_out_ratio: <sum(branch_tokens)/serial_estimate>
| dedup_count: <merged-away> | disputed: <count> | tool_calls: <sum>
fan_out_ratio and dedup_count are the feedback signals: a low dedup count
on overlapping scopes means the partition was too coarse; a high one means the
clusters were not actually disjoint (revisit the trigger).
Exit gate
