review: per-reviewer prompt-token cap #16

New issue

Closed

opened 2026-07-18 22:05:17 +00:00 by crenshawdev · 1 comment

crenshawdev commented

2026-07-18 22:05:17 +00:00

(Migrated from github.com)

Tier: Improve Cadence (ranked #3) · small safety

Cadence sends artifacts to a paid review provider with no size ceiling. Adopt a per-reviewer prompt-token cap.

Why: GSD-core's max_prompt_tokens_per_reviewer is a one-key guard against an unbounded paid call. Cheap insurance.

Cost: one key under review.* plus the check that honors it.

Source: internal GSD comparison note, 2026-07-18.

**Tier:** Improve Cadence (ranked #3) · small safety Cadence sends artifacts to a paid review provider with no size ceiling. Adopt a per-reviewer prompt-token cap. **Why:** GSD-core's `max_prompt_tokens_per_reviewer` is a one-key guard against an unbounded paid call. Cheap insurance. **Cost:** one key under `review.*` plus the check that honors it. _Source: internal GSD comparison note, 2026-07-18._

john modified the milestone from v1.2.0-rc.1 to (deleted)

2026-07-30 05:13:39 +00:00

john modified the milestone from (deleted) to v2.1.0

2026-07-30 05:15:36 +00:00

john commented

2026-08-05 17:08:46 +00:00

Owner

Closed by v2.1.0 (REV-03). review.max_prompt_tokens (default 120000) bounds the payload before any paid call is issued: an over-cap payload is refused with ok:false, reason: over-cap, and a non-string payload field is refused as bad-payload first, since an unmeasurable field is an unbounded one. The cap governs cross-model provider calls; the free claude-subagent reviewer never runs that script.

v2.3.0 (RES-03) moved the check to measure the payload FILE's contents rather than a composed string, so the cap still binds now that reviewers receive references instead of inlined artifacts.

Verified against the tree at the v2.3.0 close: config.schema.json:65.

Closed by v2.1.0 (REV-03). `review.max_prompt_tokens` (default 120000) bounds the payload before any paid call is issued: an over-cap payload is refused with `ok:false, reason: over-cap`, and a non-string payload field is refused as `bad-payload` first, since an unmeasurable field is an unbounded one. The cap governs cross-model provider calls; the free `claude-subagent` reviewer never runs that script. v2.3.0 (RES-03) moved the check to measure the payload FILE's contents rather than a composed string, so the cap still binds now that reviewers receive references instead of inlined artifacts. Verified against the tree at the v2.3.0 close: `config.schema.json:65`.

john closed this issue

2026-08-05 17:08:59 +00:00