[meta] Roadmap discussion — what's the next year of chimera-run?

[0bserver07]

Open discussion

Chimera 0.7.0 is on PyPI (2026-05-10). The post-launch trajectory is open. This issue is a discussion thread for what the next 6–12 months should look like.

The thesis we shipped

An open-source Python framework for building coding agents. Pick your provider, pick your tools, pick your loop. Chimera wires them together. Recreate SWE-Agent, Aider, Cline, or Codex in a few lines, or design something entirely new.

7 reference CLIs ship on one substrate as proof. The standard surface is in docs/architecture.md.

Where this could go next

These aren't promises, just directions. Reactions / votes / counter-proposals welcome.

Direction A — Deepen the substrate (research-y)

• Real agent learning / RL fine-tuning over recorded sessions (the chimera/training/ layer is scaffolded; nothing trains on real data yet)
• Formal verification integration — Z3, Lean. The docs/specs/formal-verification-integration.md spec exists; not implemented.
• Neural-guided synthesis — docs/specs/neural-guided-search.md describes a learned policy for synthesis tree search. Nobody's built it.

Direction B — Productize (engineering-y)

• Hosted variants: Modal / Daytona / e2b cloud sandboxes (see [env] Cloud sandbox providers — Modal, Daytona, Northflank/e2b #144)
• Better metric collection: cost-per-task, time-to-merge, success rate dashboards
• A CI mode: drop chimera into .github/workflows/ and have it auto-fix flaky tests, write missing docs, propose refactors
• Bridge to Anthropic Skills (the SKILL.md format we already implement)

Direction C — More benchmarks (eval-y)

• ProgramBench live (issue [bench] Run ProgramBench live — first 10 instances, GLM-5.1 + qwen3-coder #141)
• MultiSWE-bench live in all 5 languages
• Custom benchmark generator: take a real repo's git history, replay PRs as agent tasks
• "Skill challenge" benchmarks: 50 specific coding skills, eval which agent posture wins each

Direction D — More postures (composition-y)

• 8th CLI for a specific niche. Ideas surfaced in audits:
  • Notebook-first agent (Jupyter integration)
  • Pair-programming agent that suggests during keystrokes
  • Code-review-only agent (no edit tools)
  • Architecture / planning agent (no code tools)

Direction E — Community + ecosystem

• Plugin marketplace as a real hosted thing (today the placeholder is at examples/plugin-index.json)
• An official Discord / discussion forum
• A "use Chimera in your tool" guide for IDE integrations (Zed, VS Code, JetBrains)

What I'd vote for

• Now: [bench] Run ProgramBench live — first 10 instances, GLM-5.1 + qwen3-coder #141 (ProgramBench live) + [ci] Migrate PyPI publish from API token to Trusted Publisher #143 (Trusted Publisher) + Direction E plugin marketplace (it's 80% built, just needs hosting)
• Next month: [env] Cloud sandbox providers — Modal, Daytona, Northflank/e2b #144 (cloud sandboxes) + Direction C custom benchmark generator
• Q3: Direction A real training over recorded sessions

How to chime in

Comment here with your direction preferences, missing items, or "this is the wrong roadmap, here's a better one." Voting via reactions is fine.