docs(queue): start tool-bearing selectors at low effort #352

Merged
timothy merged 1 commits from codex/queue-tool-bearing-low into main 2026-07-14 20:15:07 +02:00
+17 -11
View File
@@ -11,7 +11,7 @@
## Orchestrator launch profile (operator-facing; do not paste)
As of 2026-07-13:
As of 2026-07-14:
- **Codex:** use **GPT-5.6 Sol, `ultra`** for the orchestrator. `ultra` is the Codex
multi-agent orchestration setting; if it is unavailable, use **GPT-5.6 Sol, `max` effort**
@@ -20,8 +20,10 @@ As of 2026-07-13:
orchestration preset built on `xhigh`). Reserve **Fable 5, `high` effort** for the frontier
escalations listed below; use `xhigh` there only when the decision remains genuinely difficult.
- **Queue preflight:** when the user has not named an issue, use the client's cheapest suitable
fast/small model at low or minimal effort. Route it as a subagent when possible; otherwise run a
separate low-cost selector session and pass its compact selection packet to the orchestrator.
fast/small model at **low** effort. Queue selection is tool-bearing, so do not use `minimal`:
Codex validates the full enabled tool profile, and enabled `web_search` can reject `minimal` even
when the selector intends to call only Gitea MCP. Route it as a subagent when possible; otherwise
run a separate low-cost selector session and pass its compact selection packet to the orchestrator.
Never spend the orchestrator tier on queue ranking.
- If those names are unavailable, choose the client's strongest long-horizon coding/agentic model
at its high extended-reasoning setting. Do not weaken the orchestrator to a fast/mini model; use
@@ -50,16 +52,20 @@ available, otherwise use a fresh agent with no implementation role.
**Low-cost routing applies throughout the session, not only to queue selection.** Before any batch of
bounded searches, inventories, log triage, URL/status sampling, or report drafting, dispatch the cheapest
suitable fast/small model at low or minimal effort. Strictly cap its files, queries, log window, and output;
the orchestrator receives a compact evidence packet and performs only the minimum focused live recheck.
Same-tier agents add parallelism, not cost savings, while direct tool-only mechanical checks may stay inline.
If a cheap worker lacks a required tool, give it already-collected bounded evidence for synthesis or report
the limitation explicitly instead of silently expanding orchestrator reconnaissance. Global Codex hook
enforcement is tracked in `timothy/server-management#592`; the wider Claude-hook port is tracked in
`timothy/server-management#593`.
suitable fast/small model. **If any tool is enabled or required, start at `low`; this includes enabled
`web_search` even when the prompt names only Gitea MCP. Reserve `minimal` for explicitly tool-free synthesis
over already-supplied bounded evidence.** Never retry a known tool-bearing profile at `minimal`. Strictly
cap its allowed files, queries, log window, result/evidence count, and output; instruct it to stop and return
partial evidence at a cap instead of broadening scope. The orchestrator receives a compact evidence packet
and performs only the minimum focused live recheck. Same-tier agents add parallelism, not cost savings,
while direct tool-only mechanical checks may stay inline. If a tool-bearing launch is rejected at `minimal`,
retry once with the same model and unchanged scope at `low`. If a cheap worker lacks a required tool, give it
already-collected bounded evidence for explicitly tool-free synthesis or report the limitation instead of
silently expanding orchestrator reconnaissance. Global Codex hook enforcement is tracked in
`timothy/server-management#592`; the wider Claude-hook port is tracked in `timothy/server-management#593`.
**When the user has not named an issue, queue selection is mechanical fast/small work, never orchestrator
work.** Dispatch exactly one selector on the cheapest suitable model at low or minimal effort. Give it only
work.** Dispatch exactly one selector on the cheapest suitable model at `low` effort. Give it only
#237's body, last ~6 comments, every OPEN milestone and its OPEN issues, OPEN `review`-labeled
issues, and the reviewer-repo candidate list plus each candidate's claim/deliverable comments; it
returns a ranked shortlist of at most three issue IDs with one-line rationales and live-state evidence. The