@georesearch/dsh-delegation-tools

Registers four Coordinator tools, delegate_literature, delegate_experiment, delegate_review, and delegate_writing, plus the specialist-only delegation_bootstrap tool.

Every call fixes the label, persona, geoResearchRole, one-shot spawn provider, depth cap, structured output schema, and role allowlist. The first child request is byte-identical for different task types within the same role. The Host stores the dynamic task type, required Skill set, completion criteria, allowed output kinds, authority, question, constraints, and task until the child calls delegation_bootstrap. Harness rc.5 itself pins delegated approval policy to never when the approval service is present. Missing future domain tools are not stubbed. The shipped configuration uses strictRoleCapabilities: true with capabilityStage: phase6, so the complete implemented specialist catalog is mandatory. It also caps each specialist request at specialistMaxTokens: 16384; a lower parent request cap remains authoritative.

Managed children must call delegation_bootstrap exactly once, then successfully call the skill tool for every core and selected supporting Skill before any role tool or structured_output. The Policy service records both protocol stages and the delegation result exposes the Host-observed requiredSkills and loadedSkills lists. Completed children must return a role- and task-compatible outputKind plus a candidate.

The role contract includes the scope-local structured_output tool guaranteed by the output schema. The global toolFilter.allow list deliberately omits it because Harness rc.5 rejects scope-local names in global restrictions.

Harness rc.5 requires one-shot structured output to have an explicit object root. Every task receives a compact { result: } schema containing only that task's allowed completed output kinds plus the bounded decision branch. Generic reports carry exact item and text limits in the tool schema, and the bootstrap result repeats the four-finding hard limit as the last Host constraint. The Host unwraps the transport envelope and applies the authoritative task/output-kind check, role-specific candidate parser, reviewer subject check, and commit validation before reporting success or committing any record.

The cache-sensitive first request keeps stable role policy and the fixed bootstrap protocol. Dynamic task JSON appears in the bootstrap tool result on the second turn. The structured-output schema varies by task so the model does not receive unrelated large candidate contracts; role policy and bootstrap prefixes remain reusable across tasks.

Child provider/model routing is taken from the parent's latest persisted request/header, not only from the Agent's creation-time options. This keeps a session-level model switch authoritative for delegated work and prevents a Flash parent from silently spawning Pro children. The configured specialist output cap is applied after routing so a larger parent allowance cannot reopen an unbounded child response.