Setup intake

setup_intake

Captures the irreducible parameters before any workstream begins.

The 9 parameters

The agent walks the user through these 9 questions in order. At most three questions per turn. The user can answer "default" to accept the recommendation for any item.

1. Mission

What is the goal of this workstream? One sentence. The user describes what success looks like.

Example: "Ship CE_RD_OS v0.1.0 to GitHub as a public, presentable, privacy-clean dual-track artifact."

2. Role

What role should the agent take? One sentence.

Example: "Senior research engineer. Compound-engineering practitioner. Conservative on scope, aggressive on quality."

3. Sources

Where does the source material live? List paths and URLs.

Example: "/mnt/project/research_findings.md, /mnt/uploads/handoff.md, arXiv:2507.17746."

4. Rubric mode

How should outputs be evaluated against the Trinity rubric?

• boolean (default): each of the 9 cells is 0 or 1. Sum threshold for pass.
• sparse_jump_3: each cell is -1, 0, or +1.
• scaled_0_1: each cell is a continuous score 0.0 to 1.0.
• ml_assisted: scaled mode plus an ML backend (see parameter 7).

Default: boolean.

5. sub_agent profile

The agent runs sub-tasks under this profile. Three sub-parameters:

• model_type: enum from claude_opus_4_7, claude_sonnet_4_6, claude_haiku_4_5, kimi_k2_6, minimax_m2_7, glm_5_1, qwen_3_6_plus, or other_via_openrouter.
• effort: enum from minimal, low, medium, high.
• loop_bound: union type. One of:
  • {type: count, count: <int 10..100>}, fixed iteration count.
  • {type: confidence, confidence_threshold: <float 0.5..0.99>}, loop until confidence threshold met.
  • {type: until_answer, until_answer: true}, loop until user answers.

Default for routine work: {model_type: claude_sonnet_4_6, effort: medium, loop_bound: {type: count, count: 20}}. See references/agentic_models.md for guidance.

6. Memory option

How should state persist across turns or sessions?

• none (default): stateless. Recommended for first-time users.
• mem0: memory layer (github.com/mem0ai/mem0, Apache-2.0). Bolt-on, framework-agnostic.
• letta: agent runtime with Core / Recall / Archival memory tiers (github.com/letta-ai/letta, Apache-2.0).
• custom: user-supplied via MCP.

See references/memory_options.md.

7. ML backend option

For rubric refinement and prompt evolution.

• none (default): rubrics stay as defined.
• gepa_via_mlflow: arXiv:2507.19457, github.com/gepa-ai/gepa, Apache-2.0. Reflective prompt evolution via MLflow integration.
• openevolve: github.com/algorithmicsuperintelligence/openevolve. Evolutionary search over prompts and code.
• custom: user-supplied.

v0.1.0 ships the hook only. Implementations are deferred to v0.2. See references/ml_backend_options.md and references/gepa_integration.md.

8. Loop bound (workstream-level)

Separate from the sub_agent loop_bound. Bounds the entire workstream.

• {type: count, count: <int>}: stop after N iterations.
• {type: confidence, confidence_threshold: <float>}: stop when the Trinity rubric aggregate score crosses threshold.
• {type: until_answer, until_answer: true}: stop when the architect signals done.

Default: {type: until_answer, until_answer: true}.

9. Output target

Where does the workstream's output go?

• chat_only: outputs render in this conversation, no files written.
• repo: outputs write to the current repo under the appropriate path.
• text_bundle: outputs target the 10-file text zip via publish_bundle.
• dual: both repo and text_bundle.

Default: dual.

Intake procedure

1. Read system_directive (already loaded if you reached here).
2. Greet the user, state the 9 parameters above will be asked.
3. Ask in batches of three (1-3, then 4-6, then 7-9), or one at a time if the user prefers.
4. For each answer, confirm before proceeding.
5. After all 9, summarize the locked configuration and ask the user to approve before any workstream begins.
6. Save the locked configuration as a host-side note titled ce_rd_os_setup.md. The format is YAML matching the sub_agent_default block in registry.yaml.

What the agent does NOT do during intake

• Does not start any workstream before all 9 parameters are locked.
• Does not assume defaults silently. If the user says "default", the agent confirms the default value before locking.
• Does not invent fields. The 9 parameters are the irreducible set for v0.1.0.
• Does not exceed three questions per turn.

Citations

• arXiv:2507.17746, RaR. Coverage / Self-contained / Importance / Reference Guidance design principles inform the rubric mode parameter.
• arXiv:2507.19457, GEPA. Source for the gepa_via_mlflow option.
• github.com/letta-ai/letta, Apache-2.0. Source for the letta memory option.
• github.com/mem0ai/mem0, Apache-2.0. Source for the mem0 memory option.
• The "no proprietary model name in identifier fields" pattern is documented in references/agentic_models.md.

End setup_intake.