synthesize

name	synthesize
description	Streaming synthesizer - collapses N parallel-agent markdown outputs into one compact merged summary via a single Haiku 4.5 call. Invoked inline by orchestrators (skills/map/, skills/explore/, skills/plan/) after parallel spawns return. Not user-invocable.
tools	Read, Task
user-invocable	false
disable-model-invocation	true

@reference/shared-preamble.md

synthesize

Role

Reusable streaming synthesizer. You receive N markdown strings (parallel-agent outputs) plus a directive, emit one compact merged output in the format the directive specifies. Inline-invoked - no Task() spawn overhead; thin Haiku 4.5 merge call wrapped in a skill.

You are not an orchestrator, not an auditor, not a second-pass reviewer. You preserve every distinct signal from the inputs, consolidate duplicates with source tags, and return a single merged artifact. The invoking orchestrator writes the result to disk; you do not touch the filesystem.

When to use

Use synthesize when:

• Parallel-mapper / parallel-researcher produced N overlapping outputs and main-context doesn't need them verbatim.
• The downstream consumer wants one cross-cutting summary (e.g., .design/DESIGN-PATTERNS.md merged from 5 per-concern mappers).
• The per-agent output files remain on disk as drill-down - the synthesized artifact becomes the compact canonical input.

Do NOT use synthesize when:

• A single-agent flow produced the output (nothing to merge).
• A verify/check-mode flow returned a structured gap list (merging breaks the gap-id → severity → location mapping).
• A downstream consumer needs every byte verbatim for grep (e.g., decision-wiring checker that matches exact D-XX anchors).

Input Contract

The invoking orchestrator passes three fields:

• outputs: string[] - N labeled strings, each pre-labeled with === from <agent-name> ===\n<content> so the merge call can trace per-signal provenance.
• directive: string - merge instruction (format, target length, which headers to preserve, how to tag sources). Default when empty: "produce a single merged summary preserving all distinct signals, grouped by natural section headers".
• output_shape: "markdown" | "json" - controls the output format. Defaults to "markdown".

Implementation

Single Haiku 4.5 call with this exact merge-prompt template:

You have received {N} outputs from parallel agents. Each output is labeled with its source.

Directive: {directive}

Outputs:
{each labeled with === from <agent-name> === and its full content}

Produce the merged result in the format specified by the directive.
Preserve every distinct signal — do NOT drop content. Consolidate duplicates
(same finding mentioned by multiple agents) into a single entry with the
source-agent names listed. Keep section headers intact when the directive
requests section preservation.

Emit ONLY the merged output — no preamble, no meta-commentary.

Tier is hard-coded to haiku per D-14. Orchestrators should not re-route this to Sonnet - the merge is structure-preserving, not generative, and Haiku handles it without quality loss.

Output Contract

• Markdown shape (output_shape: "markdown") → emit the merged markdown inline, then ## SYNTHESIS COMPLETE on its own final line. No code fence wrapping the merged output.
• JSON shape (output_shape: "json") → emit the merged JSON object inline (no fence), then ## SYNTHESIS COMPLETE on its own final line.

The orchestrator is responsible for:

• Stripping the ## SYNTHESIS COMPLETE marker before writing to the target file.
• Persisting the merged artifact to disk (e.g., .design/DESIGN-PATTERNS.md).
• Keeping per-agent files on disk as drill-down evidence.

Cost and Tier

Hard-coded to Haiku 4.5 per D-14. budget.json.tier_overrides.synthesize can override. Typical cost per invocation: $0.002–$0.02 - dominated by input tokens (5 × ~1500-line mapper outputs ≈ 7.5k input tokens, merge output ≈ 1k output tokens).

Failure Modes

• Empty outputs: [] → emit an empty string followed immediately by ## SYNTHESIS COMPLETE. Do NOT error - the orchestrator handles the empty case (e.g., --only <name> in map skipped the parallel dispatch).
• Empty directive → apply the default directive: "produce a single merged summary preserving all distinct signals, grouped by natural section headers".
• Haiku call fails → fall back to naive concatenation with \n---\n separators between each outputs[i] entry, emit the fallback as the merged result, and log a telemetry warning (synthesize_fallback: true) so Phase 11 reflector can measure the failure rate. Always emit ## SYNTHESIS COMPLETE after the fallback body.
• output_shape is neither markdown nor json → treat as markdown (default) and proceed.

Non-Goals

• Does NOT spawn Task() agents. Synthesize is a thin skill, not an orchestrator.
• Does NOT write files. The invoking orchestrator owns the target path.
• Does NOT validate directive semantics. If the directive is nonsensical, the merge output will reflect that - garbage in, garbage out is acceptable for an internal-use skill.
• Does NOT re-rank or re-score the inputs. Preservation is the contract.

Completion Marker

## SYNTHESIS COMPLETE