Technical reference / Rendering & development
Extension points
Where to change prompts, collaboration and analysis, and how to preserve comparability.
Execution prompts live in prompts/planning.py, with shared writing guidance in prompts/style.py. Analysis prompts live in prompts/analysis.py. A task's prompt supplies the challenge; it is distinct from these stage instructions.
| Change | Affected boundary |
|---|---|
| Initial/refinement instructionsReviewed E 0.1.5 · A 0.1.4 · R 0.1.16 | Changes how models propose and respond to peers; preserve the structured proposal contract or update its schema and readers. |
| Voting criteriaReviewed E 0.1.5 · A 0.1.4 · R 0.1.16 | Changes selection incentives; keep valid-proposal validation and inspect raw voter prompts. |
| Analyst criteriaReviewed E 0.1.5 · A 0.1.4 · R 0.1.16 | Changes the interpretation layer without rerunning the council, provided stored evidence is sufficient. Preserve old analysis records for comparison. |
| Shared writing styleReviewed E 0.1.5 · A 0.1.4 · R 0.1.16 | Can affect multiple stages/components. Record every affected component version. |
For a controlled experiment, retain the code revision and prompts for each variant. Configuration identity alone does not capture a source-template edit.
The current graph is the starting architecture. A critic introduced after round X, a supervisor, adversarial roles, message routing or conditional recurrence requires executor implementation work. It cannot be activated by adding arbitrary keys to today's JSON configuration.
A process extension needs an explicit state/transition design, clear participant inputs, termination rules and trace records for the new interactions. The report would then need to distinguish those mechanisms from the existing all-to-all previous-round sharing. Preserve enough evidence to evaluate whether a change actually improves outcomes rather than simply increasing agreement.
Implement the Provider contract, select its environment key and structured-output mode, construct the client, describe effective settings without credentials and define any non-generic pricing rules. Register the adapter in llm.PROVIDERS. Check the recorded usage shape and raw exchanges with provider-specific validation.
Do not assume OpenAI-compatible transport implies identical reasoning options or JSON-schema support. The current non-OpenAI client guard exists to prevent accidental use of an OpenAI key on another endpoint.
The current deterministic layer is concentrated in report/diff.py and report/contributions.py; the LLM interpretation is in report/analysis.py and its prompts. A conceptual method would need to represent ideas across wording changes, merges and splits, expose supporting evidence and uncertainty, and preserve compatibility with historical traces.
Do not relabel the present title counts as semantic influence. A new method should have its own version and explanation so comparisons can distinguish a change in model behavior from a change in measurement.
Update the relevant component build and root changelog when runtime behavior changes. Review the matching documentation pages and their version scopes, update screenshots only when necessary, rebuild the microsite and check links. docs/reference/manifest.json records the source files reviewed for each page; the builder rejects an unnoticed component-version mismatch.
Documentation-only changes do not require changing executor, analyst or renderer versions. The documentation edition and changelog identify the reference update separately.
Source files used for this reference
Reviewed at 7361ea0b71a5.