Rationale: Mature, actively maintained plugin (v0.1.13; recent commits added Lorenzen move validation, Sawyer flow envelope, stemma tracing) sitting squarely on the dominant production frontier. Anthropic’s own context-engineering playbook names sub-agent isolation returning 1-2k-token condensed summaries to a lead as a core pattern — that is exactly intersynth’s contract (compact-by-contract, <=15 lines, full reports to disk). Its “disagreement preserved not resolved” design (severity_conflict, contested verdicts) plus hearsay/sycophancy/conformity scoring directly anticipates the 2026 reliability-science shift from pass@1 to variance/robustness and the documented LLM-judge failure modes (self-preference, reward hacking) — it is ahead of most competitors on discourse quality. The mission and the interflux delegation contract are sound and current; nothing is obsolete. It earns keep, not modernize, because the architecture is correct — the gaps are additive refinements, not a redesign.
Multi-agent synthesis engine: 3 Haiku subagents (synthesize-review, synthesize-research, synthesize-documents), one SessionStart nudge hook, and lib-verdict.sh shell utility — no commands, no skills. It is the synthesis stage of the orchestrator-worker pattern: parallel review/research agents dispatched by interflux/Clavain write .md files to OUTPUT_DIR; intersynth reads them, validates structure, deduplicates findings (file:line + same-issue rules), scores cross-agent convergence, and writes structured verdict JSONs to .clavain/verdicts/ plus findings.json/summary.md (review) or synthesis.md (research) to disk, returning only a <=15-line compact summary so the host context stays clean. It implements unusually advanced discourse mechanics: reactions/conductor scoring, hearsay classification, Lorenzen dialogue-game move legality, sycophancy/conformity scoring, stemma analysis (Jaccard over shared evidence sources), QDAIF diverse-perspective scoring, and a Sawyer flow-envelope health gate. Tightly bound to interflux via docs/spec/contracts/synthesis-delegation.md (confirmed present + current). Opt-in auto-bead creation from P0/P1 findings.
Two frontier ideas it lacks. (1) Grammar-constrained / deterministic-skeleton generation: findings.json and all the arithmetic (dedup, Jaccard stemma, Gini/novelty/relevance for Sawyer, convergence counts, DWSQ) are described as steps a Haiku LLM performs by reading prose and emitting JSON — there is no schema-enforced structured output (Pre3/IterGen/StructuredOutput-style) and no deterministic script doing the set arithmetic. The interleave ‘deterministic skeleton + LLM islands’ pattern and constrained-decoding research would make output machine-parseable without retries and push pure computation off the LLM (cheaper, faster, more reliable). (2) Embedding-based semantic dedup: finding matching is purely lexical (file:line + ‘same issue’), so two agents describing the same defect in different words are not merged; RAG-MCP / Select-then-Solve embedding gating (and the in-ecosystem intersearch infra) would catch semantic duplicates the current rules miss.
Low and healthy. Cleanly separated from interflux (dispatch/triage), interspect (agent calibration/routing), intertrust (reputation scoring), and intercheck (quality guards) — intersynth uniquely owns the aggregate/dedup/verdict-write stage. The one intentional overlap is lib-verdict.sh, which Clavain keeps a documented backward-compat copy of; this is a managed seam (intersynth is the canonical home), not accidental redundancy. The sycophancy/conformity scoring lightly touches intertrust’s reputation territory but operates per-synthesis-run rather than as durable reputation, so they compose rather than collide.
[“Extract the deterministic computation (dedup grouping, Jaccard stemma transitive closure, participation Gini, novelty/relevance ratios, convergence counts, DWSQ) into a lib-synth.sh / Python helper that emits findings.json against a JSON Schema, so Haiku only writes the free-text island fields (titles, summaries, mini-narratives) — adopt the interleave deterministic-skeleton+LLM-island pattern for reliability and lower token cost.”, “Add optional embedding-based semantic dedup as a pre-pass before the lexical rules in step 6: query intersearch (intersearch__embedding_query) on finding titles+evidence to merge semantically-equivalent findings phrased differently (RAG-MCP / Select-then-Solve gating), behind a flag so it degrades gracefully when intersearch is absent.”, “Validate findings.json against a committed JSON Schema in tests/structural/ and assert the <=15-line / ‘Protocol:’ first-line return contract, so synthesis output reliability is regression-guarded rather than prose-checked.”, “Expose the existing conformity/sycophancy/Sawyer signals as an explicit per-run ‘synthesis confidence’ field in findings.json so downstream consumers (beads, quality-gates) can weight verdicts by discourse health — connecting the plugin to the pass@1->variance reliability-science frontier it already half-implements.”]