Plastic Labs Internal Wiki

Recently Updated

roadmap-q2-2026

5 min read

Honcho Feature Roadmap — Q2 2026

ARCHIVED (2026-07-13) — historical snapshot, do not action. Q2 2026 elapsed. Of the 14 planned specs only the LLM refactor shipped (as done/llm-client-refactor.md; litellm was rejected, contra the litellm.md name below). The plan did not track execution and omits the work actually in flight (scopes DEV-1970, cloudevents-full-tracing DEV-1900, renameable-primitives DEV-1760, conclusion-quality ML-335, k8s migration DEV-1669). Superseded by the operating-model reorg.

Status: Archived | Owner: vineeth | Last updated: 2026-03-25

Overview

14 feature specs organized into 5 implementation waves. Each wave’s features can be implemented in parallel within the wave, but waves must be completed in order due to dependencies.

Target: All features shipped by end of Q2 2026. SDK Target: v2.1 (additive, non-breaking). API Surface: Additive only. No breaking changes.

Dependency Graph

Wave 1 (Foundation)
├── conclusion-tagging ─────────────────┬──→ dreaming-enhancements (Wave 2)
│                                       ├──→ reasoning-traces (Wave 2)
│                                       ├──→ dialectic-enhancements (Wave 3)
│                                       └──→ composable-peers-v2 (Wave 4)
├── search-interleaved-messages ────────→ (independent, no downstream deps)
├── sdk-improvements ───────────────────→ (independent, but all later specs benefit)
└── llm-client-refactor ────────────────┬──→ dialectic-enhancements (Wave 3)
                                        ├──→ multi-modal (Wave 4)
                                        ├──→ cli (Wave 5)
                                        └──→ sqlite-support (Wave 5)

Wave 2 (Memory System)
├── dreaming-enhancements ──────────────┬──→ composable-peers-v2 (Wave 4)
│                                       └──→ dialectic-enhancements (Wave 3)
└── reasoning-traces ───────────────────┬──→ dialectic-enhancements (Wave 3)
                                        └──→ composable-peers-v2 (Wave 4)

Wave 3 (Dialectic)
└── dialectic-enhancements ─────────────→ (completions endpoint is low priority, can defer)

Wave 4 (Major Features)
├── file-system-primitives ─────────────→ composable-peers-v2
├── composable-peers-v2 ────────────────→ (terminal — no downstream deps)
└── multi-modal ────────────────────────→ (terminal)

Wave 5 (Infrastructure)
├── cli ────────────────────────────────→ mocking-server
├── sqlite-support ─────────────────────→ (terminal)
└── mocking-server ─────────────────────→ (terminal)

Wave 1 — Foundation

Timeline: Implement first. No dependencies. Theme: Small, additive changes that unlock everything else.

#SpecMigrationScopeKey Deliverable
1conclusion-tagging.mdYes (add metadata JSONB column to documents)SmallPublic metadata on conclusions, read-only level/times_derived/source_ids
2search-interleaved-messages.mdNoSmallinclude_context param on search, seq_in_session in filters and response
3sdk-improvements.mdNoMediumPydantic TypedDict filters (Python), Zod filter types (TS), cursor pagination opt-in
4litellm.mdNo (internal refactor)LargeProviderBackend protocol, ModelConfig primitive, prefix caching, integration tests

Implementation notes:

  • Specs 1-3 can be implemented by a single developer in parallel
  • Spec 4 (LLM refactor) is the largest in this wave — primarily internal, no API changes
  • Spec 1 unblocks the most downstream work — prioritize it

Wave 2 — Memory System

Timeline: After Wave 1 conclusion-tagging lands. Theme: Expanding Honcho’s reasoning and memory capabilities.

#SpecMigrationScopeKey Deliverable
5dreaming-enhancements.mdYes (dream_definitions table)LargeTaggingSpecialist, materialized dreams, consolidation improvements
6reasoning-traces.mdYes (reasoning_traces table)LargeTrace storage in PostgreSQL, message→trace→conclusion provenance, agent tools

Implementation notes:

  • Specs 5 and 6 can be implemented in parallel
  • Both require the metadata column from conclusion-tagging (Wave 1)
  • Spec 6 requires careful integration with honcho_llm_call() in clients.py

Wave 3 — Dialectic Enhancements

Timeline: After Waves 1-2 land. Theme: Making the dialectic more powerful and interoperable.

#SpecMigrationScopeKey Deliverable
7dialectic-enhancements.mdNoLargeStructured outputs (JSON Schema), evidence/citations field, OpenAI-compatible endpoint

Implementation notes:

  • Three features in one spec, but can be implemented incrementally:
    1. Structured outputs (depends on LLM refactor for reliable cross-provider structured output)
    2. Evidence/citations (depends on conclusion-tagging for metadata, reasoning-traces for provenance)
    3. OpenAI completions endpoint (lower priority — defer if needed)
  • Structured outputs + evidence can ship together as they touch the same response schema

Wave 4 — Major Features

Timeline: After Waves 1-2 land. Can overlap with Wave 3. Theme: The big architectural additions.

#SpecMigrationScopeKey Deliverable
8file-system-primitives.mdYes (peer_files, file_versions tables)LargeVersioned peer files, auto-embedding, grep/search, agent tools
9rfc-composable-peers.md (v2)Yes (peer_memberships table)LargeSub-peers, aggregation, invalidation, restructuring operations
10multi-modal.mdYes (message attachments column)LargeImage/PDF support, storage backend, deriver vision, SDK upload

Implementation notes:

  • Spec 8 (files) should land before Spec 9 (composable peers) since v2 references peer files
  • Spec 10 (multi-modal) is independent of 8 and 9 — can be done in parallel
  • Spec 9 has the most dependencies (conclusion-tagging, files, dreams, traces) — schedule last in wave

Wave 5 — Infrastructure

Timeline: Can be done in parallel with Wave 4. Theme: Developer experience and deployment flexibility.

#SpecMigrationScopeKey Deliverable
11cli.md (v2)No (packaging)Largeuv workspaces, honcho serve/up/down/status/doctor, docker compose orchestration
12sqlite-support.mdNo (dialect layer)MediumDatabaseDialect protocol, sqlite-vec, FTS5, locking abstraction
13mocking-server.mdNoSmallhoncho mock command, in-memory SQLite, stateful request validation

Implementation notes:

  • Spec 11 (CLI) is the biggest structural change — repo restructuring into uv workspaces
  • Spec 12 (SQLite) depends on the LLM refactor’s dialect-agnostic design
  • Spec 13 (mocking) depends on CLI packaging but is small scope

Migration Summary

WaveSpecNew TablesModified TablesIndex Changes
1conclusion-tagging—documents (+metadata JSONB)+GIN on metadata
2dreaming-enhancementsdream_definitionsqueue (+dream_type enum)—
2reasoning-tracesreasoning_tracesdocuments (+trace_id ref in internal_metadata)+composite on agent_type/created_at
4file-system-primitivespeer_files, file_versions—+HNSW on file chunk embeddings
4composable-peers-v2peer_memberships—+composite on parent/child
4multi-modal—messages (+attachments JSONB)—

Total new tables: 5 Total modified tables: 3 All migrations are additive — no column drops, no type changes, no data loss.

Risk Summary

RiskSeverityMitigation
LLM refactor breaks provider-specific behaviorHighIntegration test matrix across providers with multi-turn tool loops
Reasoning trace storage bloats databaseMediumSize-based tiering (inline < 50KB, S3 for larger), configurable retention
Sub-peer aggregation performance at depth > 2MediumDefault depth=1, query-time CTE with depth limit
Multi-modal storage costs for managed serviceMediumCloudEvents for billing, configurable size limits per tenant
SQLite missing PostgreSQL features at edge casesLowFeature detection at startup, clear error messages for unsupported operations
uv workspace restructuring breaks CI/CDMediumPhased migration: workspace packages first, CI updates second

Success Metrics

  1. All 14 specs implemented with passing tests
  2. SDK v2.1 released with all new features
  3. No breaking changes to existing API consumers
  4. All migrations run cleanly on existing production databases
  5. Integration tests pass across Anthropic, OpenAI, Gemini, and Groq providers
  6. CLI honcho up bootstraps a working local instance in < 60 seconds

Graph View