v0.8.65: Evaluate the Orchestration disposition in constitution.md — help, deadweight, or harm?

[Hmbown]

We are trying the ## Orchestration disposition added to crates/tui/src/prompts/constitution.md via #3470 (merged for the 0.8.65 rebuild). This issue tracks the empirical evaluation, to be done after the 0.8.65 global rebuild + hands-on testing.

Question

Does instilling the orchestrator disposition ("when work outgrows one context you are an orchestrator; delegate the doing, verify every returned slice, keep the loop alive") actually change agent behavior for the better, or is it deadweight prompt tokens — or does it cause over-orchestration of small tasks?

What to measure (post-rebuild)

• Does the agent delegate appropriately on large/multi-context tasks vs. before?
• Does it over-orchestrate small/obvious one-file tasks (the disposition explicitly warns against this — does the warning hold)?
• Verification discipline: does "never trust a worker's done" translate to more ground-truth checks?
• Token cost of the added constitution section vs. observed benefit.

How

A/B feel-test during real 0.8.65 usage (disposition on vs. a build with the section removed), plus a few representative large tasks. Decide: keep as-is, soften (e.g. the unqualified "Always have work in flight" at constitution.md:390), or remove.

Ref: #3470 (the disposition), #3154 (Fleet EPIC).

[Hmbown]

Pre-fix evaluation map (2026-06-23)

Disposition: keep as a question/evaluation issue. Do not treat this as an automatic prompt edit. The useful output is evidence: keep, trim, move, or remove the Orchestration disposition.

Code/doc seams:

• crates/tui/src/prompts/constitution.md contains the ## Orchestration guidance under evaluation.
• crates/tui/src/prompts.rs:332 includes that markdown as BASE_PROMPT.
• crates/tui/src/prompts.rs:2434 has the skeleton guard test for required constitution structure.
• .codewhale/constitution.json, AGENTS.md, and CLAUDE.md can add repo-local operating policy, but they should not duplicate broad base prompt guidance.
• crates/tui/src/context_report.rs:256 can surface prompt/context structure for diagnostics.

Suggested evaluation plan:

• Compare at least three real task types: small single-file fix, multi-issue release maintenance, and Fleet/sub-agent orchestration.
• For each, record whether the disposition improves delegation/verification or causes extra planning/verbosity/over-delegation.
• Check cache/prompt cost impact if the section is long.
• If it helps only CodeWhale contributors, prefer moving specifics to repo docs/constitution rather than the global base prompt.

Acceptance checks:

• The issue closes with a concrete disposition: keep as-is, trim, move to repo-local policy, or remove.
• If changed, tests that guard constitution structure and key behavior are updated.
• The result should reduce ambiguity for agents, not add another vague instruction block.