refactor(delegate_task): tasks-only interface + depth-derived delegation (1,201 → 773 tok/call, −36%)

[teknium1]

Campaign entry (#95681). NOTE: the tracker's old 651-tok figure was the static stub — this tool builds its real text via dynamic_schema_overrides; AS-SERVED on a default config is 1,201 tok/call. All numbers measured through model_tools.get_tool_definitions() (the wire path).

Three maintainer-directed interface changes + diet, in commit order:

Commit 1 — depth-derived delegation, session-filtered rules, background unadvertised

role retired. The old model made the CALLER declare role='leaf'|'orchestrator', then silently coerced it when config disagreed. Now the runtime derives it: a child may itself delegate iff delegation.orchestrator_enabled and its depth < max_spawn_depth (_build_child_agent, the existing enforcement point). role/tasks[].role leave the schema; handler accepts+ignores them (old transcripts, kanban dispatcher). Depth-1 installs (default): the word "orchestrator" appears NOWHERE in the served schema. Behavior delta on nesting-enabled installs only: children with depth budget now always get delegate_task (previously only when the caller remembered to ask).

Restrictions rule session-filtered. The "children cannot call X" line now names only tools the SESSION actually has (same model_tools seam as the browser_exec terminal-gate; memo key already includes toolsets): full kit → full list · partial → filtered · none → line dropped. send_message is never named — gateway-internal vocabulary; enforcement unchanged (DELEGATE_BLOCKED_TOOLS still strips all five).

background (86 tok of "DEPRECATED / IGNORED") unadvertised; handler unchanged.

Commit 2 — tasks[] is the only advertised shape (the clarify #95907 pattern)

The runtime already wrapped single-goal calls into a one-entry task list internally; the schema was the only place the two-shape fiction lived — and it bred real confusion (top-level context looked like shared/broadcast context; it never was: each child sees only its own tasks[i].context).

• Top-level goal/context/output_schema leave the schema (handler-accepted legacy, wrapped at dispatch as before; do-not-re-add comments in place).
• tasks[] (minItems 1) is the one shape; item fields now carry the full teaching, including the context-isolation rule models previously had to guess: "Each child sees only its own context — repeat shared background in every task that needs it."
• The min-2 batch gate is gone (a one-entry array is now canonical, matching the runtime's internal shape); the placeholder/template goal checks still run on every entry, and the short-goal check applies only to multi-task fan-outs (single tasks legitimately use short goals — preserves the historical exemption including the gpt-5.4-mini tasks=[]+goal regression case).
• Missing-input error teaches the canonical shape.

Numbers (as served, o200k_base, default config)

	tok/call
main	1,201
commit 1	819
commit 2	773 (−428, −36%)

4 advertised params (tasks, action, subagent_id, message), down from 9.

Verification

• E2E (wire path): depth-1 + depth-3 (temp HERMES_HOME) variants; all four session-filter shapes; role/orchestrator/leaf/send_message asserted absent from serialized schema.
• E2E (dispatch): one-entry array with short goal spawns clean; legacy goal=/context=/role= still flow; bare call gets the teaching error.
• 153/156 delegate tests pass — 3 failures are the pre-existing output_schema trio (reproduced identically on clean origin/main via stash). 5 contract tests updated to pin the new interface, incl. flipping the one that REQUIRED the min-2 rejection into one that requires one-entry acceptance.

[github-actions]

૮ >ﻌ< ა ci review

_{ran on 2d54f00 — refactor(delegate_task): tasks[] is the only advertised shap}

⚠️ Warnings

OSV vulnerability scan · View job

6 known vulnerabilities found in pinned dependencies.

• CVE-2026-67213 in website/package-lock.json
• CVE-2026-71554 in uv.lock
• CVE-2026-70608 in package-lock.json
• CVE-2026-56876 in package-lock.json
• CVE-2026-70606 in package-lock.json
• CVE-2026-71554 in uv.lock

How to fix:

Review the findings in the Security tab. Update the affected dependencies if a patched version is available.

---

debug info

CI timings

CI timings · View report · View job

Wall time 4m8s vs 4m17s (-3.5%). 7 job(s) slower, 5 faster,

• Check contributors / check-attribution: +9.0s
• OSV scan / Scan lockfiles / osv-scan: -8.0s
• Python tests / Run tests: +6.0s
• OS-specific tests / Windows-only tests: -6.0s
• Python lints / ruff enforcement (blocking): +4.0s