Tier the standing reading for schema-bound workflow stages #31

New Issue

2026-07-17T15:20:06+02:00

claude commented

2026-07-17 15:20:06 +02:00

Problem

The anonymous schema-bound stages in implement/workflows/*.js (plan-index, plan-extract, snapshot, mini-verify, the standard end-verify, compiler-driven's build/suite verify, mini preflight) all receive the full standing-reading instruction — read the project CLAUDE.md plus git log -10 --format=full — via the shared STANDING prompt preamble, although a tree-footprint or text-extraction stage consumes none of it and a verify stage needs only the build/test commands. Each such read re-executes on every dispatch, including tight repair rounds; at ~40k+ real tokens per primed context this is pure priming tax on the highest-volume dispatches.

Change

Two STANDING variants in the workflow scripts: verify-type stages read only the ## Skills plugin: project facts build/test commands from CLAUDE.md; tree/extraction stages read nothing beyond their own instructions. git log drops out of both. Judgment roles (implementer, spec-reviewer, quality-reviewer, tester and all named agents) keep the full convention unchanged.
docs/conventions.md § Standing reading gains the tiering sentence; docs/agent-template.md § Standing reading list mirrors it.

Precedent

Narrowed standing reading already exists per role and is the established pattern, not a new convention: tester reads git log -3, plan-recon and grounding-check read -5, fieldtester -8, docwriter -10 --oneline scoped to code roots. The change extends the same principle to the script-defined stages, which have no agent file of their own.

## Problem The anonymous schema-bound stages in `implement/workflows/*.js` (plan-index, plan-extract, snapshot, mini-verify, the standard end-verify, compiler-driven's build/suite verify, mini preflight) all receive the full standing-reading instruction — read the project `CLAUDE.md` plus `git log -10 --format=full` — via the shared `STANDING` prompt preamble, although a tree-footprint or text-extraction stage consumes none of it and a verify stage needs only the build/test commands. Each such read re-executes on every dispatch, including tight repair rounds; at ~40k+ real tokens per primed context this is pure priming tax on the highest-volume dispatches. ## Change - Two STANDING variants in the workflow scripts: verify-type stages read only the `## Skills plugin: project facts` build/test commands from `CLAUDE.md`; tree/extraction stages read nothing beyond their own instructions. `git log` drops out of both. Judgment roles (implementer, spec-reviewer, quality-reviewer, tester and all named agents) keep the full convention unchanged. - `docs/conventions.md` § Standing reading gains the tiering sentence; `docs/agent-template.md` § Standing reading list mirrors it. ## Precedent Narrowed standing reading already exists per role and is the established pattern, not a new convention: `tester` reads `git log -3`, `plan-recon` and `grounding-check` read `-5`, `fieldtester` `-8`, `docwriter` `-10 --oneline` scoped to code roots. The change extends the same principle to the script-defined stages, which have no agent file of their own.

claude self-assigned this 2026-07-17 15:20:06 +02:00

claude referenced this issue

2026-07-17 15:20:21 +02:00

Remove the postmortem skill #27

Brummel referenced this issue from a commit

2026-07-17 16:01:20 +02:00

feat(workflows): tier the standing reading for schema-bound stages

Brummel closed this issue

2026-07-17 16:01:20 +02:00

Sign in to join this conversation.