Enforce session-wide low-cost routing for bounded reconnaissance #342

Closed
opened 2026-07-13 21:35:44 +02:00 by timothy · 3 comments
Owner

Problem

The standing kickoff requires a cheap selector for queue ranking and broadly says to route bounded recon to fast/small models, but the latter is not operationally explicit enough. During #330 the expensive orchestrator performed a multi-query GitHub duplicate search inline instead of dispatching it to the available GPT-5.4 Mini route.

Proposed

Strengthen docs/handoffs/chicorytv-issue-queue.md so the rule is unambiguous for every future session:

  • cost routing applies throughout the session, not only queue selection;
  • before a batch of bounded searches, inventories, log triage, URL/status sampling, or report drafting, use the cheapest suitable fast/small model at low/minimal effort;
  • the orchestrator receives only a compact result packet and performs the minimum focused live recheck;
  • same-tier subagents provide parallelism, not cost savings;
  • direct tool-only mechanical checks remain inline;
  • if the cheap worker lacks a required tool, feed it already-collected bounded evidence for synthesis or explicitly report the limitation instead of silently expanding orchestrator recon.

Track mechanical hook enforcement/Claude-hook porting in server-management, where the global client configuration is owned.

Done-when

  • standing kickoff states that low-cost routing is session-wide and gives concrete bounded-recon examples
  • fallback and same-tier-agent behavior are explicit
  • hook-enforcement follow-up is linked
  • docs-only review passed

Triage

Backlog / Medium — workflow cost-control hardening; no product correctness impact.

## Problem The standing kickoff requires a cheap selector for queue ranking and broadly says to route bounded recon to fast/small models, but the latter is not operationally explicit enough. During #330 the expensive orchestrator performed a multi-query GitHub duplicate search inline instead of dispatching it to the available GPT-5.4 Mini route. ## Proposed Strengthen `docs/handoffs/chicorytv-issue-queue.md` so the rule is unambiguous for every future session: - cost routing applies throughout the session, not only queue selection; - before a batch of bounded searches, inventories, log triage, URL/status sampling, or report drafting, use the cheapest suitable fast/small model at low/minimal effort; - the orchestrator receives only a compact result packet and performs the minimum focused live recheck; - same-tier subagents provide parallelism, not cost savings; - direct tool-only mechanical checks remain inline; - if the cheap worker lacks a required tool, feed it already-collected bounded evidence for synthesis or explicitly report the limitation instead of silently expanding orchestrator recon. Track mechanical hook enforcement/Claude-hook porting in server-management, where the global client configuration is owned. ## Done-when - [x] standing kickoff states that low-cost routing is session-wide and gives concrete bounded-recon examples - [x] fallback and same-tier-agent behavior are explicit - [x] hook-enforcement follow-up is linked - [x] docs-only review passed ## Triage **Backlog / Medium** — workflow cost-control hardening; no product correctness impact.
timothy added the priority: mediumin-progress labels 2026-07-13 21:35:51 +02:00
Author
Owner

claiming — session started 2026-07-13T19:35:51.445Z

claiming — session started 2026-07-13T19:35:51.445Z
Author
Owner

Hook enforcement is now tracked in timothy/server-management#592; the broader Claude/internal/git-hook port is tracked in timothy/server-management#593. The standing kickoff update links both.

Hook enforcement is now tracked in `timothy/server-management#592`; the broader Claude/internal/git-hook port is tracked in `timothy/server-management#593`. The standing kickoff update links both.
timothy removed the in-progress label 2026-07-13 21:50:52 +02:00
Author
Owner

Done

  • Root cause: session-wide low-cost routing was phrased too loosely, so bounded GitHub reconnaissance stayed in the expensive primary session.
  • What changed: the standing kickoff now makes cheap routing mandatory across the whole session, names concrete bounded tasks, requires hard input/output caps and compact handoffs, distinguishes same-tier parallelism, and defines the missing-tool fallback.
  • Hook follow-up: timothy/server-management#592 implemented and installed the managed SessionStart policy; timothy/server-management#593 tracks the broader Claude/internal/git-hook port.
  • Verification: git diff --check; pre-push API drift, ESLint, TypeScript, and Vite checks; all seven Gitea PR statuses green/skipped-as-designed.
  • Review: no findings from a one-file GPT-5.4 Mini/low cold review; Review-verdict: MERGEABLE @ 8da1256ffaa455975d6f2e50023a08dbea12163c.
  • Commit/PR: 8da1256ffaa455975d6f2e50023a08dbea12163c, PR #343 fast-forward merged.
  • Docs updated: docs/handoffs/chicorytv-issue-queue.md.
## Done - **Root cause:** session-wide low-cost routing was phrased too loosely, so bounded GitHub reconnaissance stayed in the expensive primary session. - **What changed:** the standing kickoff now makes cheap routing mandatory across the whole session, names concrete bounded tasks, requires hard input/output caps and compact handoffs, distinguishes same-tier parallelism, and defines the missing-tool fallback. - **Hook follow-up:** `timothy/server-management#592` implemented and installed the managed SessionStart policy; `timothy/server-management#593` tracks the broader Claude/internal/git-hook port. - **Verification:** `git diff --check`; pre-push API drift, ESLint, TypeScript, and Vite checks; all seven Gitea PR statuses green/skipped-as-designed. - **Review:** no findings from a one-file GPT-5.4 Mini/low cold review; `Review-verdict: MERGEABLE @ 8da1256ffaa455975d6f2e50023a08dbea12163c`. - **Commit/PR:** `8da1256ffaa455975d6f2e50023a08dbea12163c`, PR #343 fast-forward merged. - **Docs updated:** `docs/handoffs/chicorytv-issue-queue.md`.
Sign in to join this conversation.
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: timothy/ersatztv#342