Repository navigation
feat: add rules that limit image reads, bound delegated work, and finish background work - #87
Conversation
A review session read 59 key frames 4,010 times in two hours and cost $1,032 at list prices. Claude Code kept about 35 images in each request and put "[media removed: request limit]" in place of the older ones. The session took the marker for a failed render and read the frames again, so every request wrote its whole context to the cache again. The rule asks for one read of each image, notes in place of a second read, contact sheets, and image limits in each brief that asks for images.
A brief asked a child agent to view every key frame in a directory and set no limit. The child ran for two hours, 191 requests, and 4,024 tool calls while the parent waited. The same brief one round earlier took 11 requests and about 75 tool calls. A limit of 150 tool calls would have stopped it after 19 requests and $11. The rule adds a limit, a stop on repeated calls, and an overrun check to each delegation.
Three of five delegated implementers ended their turn while a per-commit typecheck still ran in the background. Each sent a second report about a minute later, and the parent checked the typecheck by hand each time. The rule asks for a wait on each background command before the final report.
|
review/correctness stood down on |
abominable-intelligence
added
the
review:runeseer
|
review/correctness round 1 started on |
Runeseer review — cleanLooks good. The head is unchanged since the last review, the three rules stay prose-only inside the rule length limits, and the
No open findings · Reviewed |
github-actions
Bot
removed
the
review:runeseer
![runeseer[bot]](https://avatars.githubusercontent.com/u/304134085?s=60&v=4){kind=small}
Codex Review Summary: ✅ CompletedThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
|
review/correctness round 2 started on |
|
Claude finished @N4M3Z's task in 1m 19s —— View job Runeseer review — cleanLooks good. The head is unchanged since the last review, the three rules stay prose-only inside the rule length limits, and the
No valid concerns from the other reviewers. The only open comment is a Codex status notice with no finding; the Codex code and security reviews were still running on this head when this round closed. The ledger carries no open threads and no finding from round 1. No open findings · Reviewed |
![macroscopeapp[bot]](https://avatars.githubusercontent.com/in/900172?s=60&v=4){kind=small}
ApprovabilityVerdict: Approved at Macroscope's review found this PR approvable — Adds agent guidance to bound image reads and delegated work, and to finish or disclose background commands before reporting. These are limited workflow safeguards with explicit reporting of incomplete work and no direct infrastructure or customer-data changes. Notes:
You can add or adjust custom eligibility rules. Learn more. |
abominable-intelligence
added
the
review
|
| Case | Rule | With rules | Baseline | What changed |
|---|---|---|---|---|
Write a subagent brief to check docs/ for broken links |
BoundDelegatedWork | 1.00 | 0.33 | All 3 baseline briefs had no numeric limit and no stop conditions. All 3 briefs with the rule had both, plus a report shape. |
| Find the first red frame in a 40-frame PNG recording | LimitImageReads | 1.00 | 1.00 | With the rule: 0 image reads in 3 of 3 runs. In a kept trace, the model decoded the PNGs with a script. Baseline: a binary search with 6 to 8 image reads. |
An image shows only [media removed: request limit], and the prompt gives the earlier note |
LimitImageReads | 1.00 | 1.00 | No difference. Neither side read the image again. |
Run ./build.sh in the background and say whether it passed |
FinishBackgroundWork | 1.00 | 1.00 | Not a fair test: when the job ends, claude -p starts the model again, so both sides finish with the result. In one kept trace for each side, the first report with the rule said "still running", gave the log path, and said the result was not verified. The baseline said "I'll tell you once it finishes". Neither side waited before its first report. |
Mean change: +0.17.
What this shows: BoundDelegatedWork has a clear effect. LimitImageReads cuts image reads. The media-removed case does not reproduce the real loop, because that needs about 35 images in context and the harness removing them. FinishBackgroundWork improves what the first report says, but the model still ended its turn and did not wait. The Runeseer note about its first sentence may be the reason.
Limits: Bash was available in every case, so both sides could use a script in the frames case. 3 runs for each side, one model (the original loop was Fable 5.1), and Haiku graded the LLM checks. The suite is outside the repo.
|
Your first merged contribution here — welcome, @abominable-intelligence, and thank you. Every rune in this deck was reviewed the same way yours just was; may it be the first of many. |
Three core rules from the t3code/cd run review. One judge subagent read 59 screenshots 4,010 times. It took Claude Code's
[media removed: request limit]marker for a failed render. After that, every request missed the cache, and that judge cost $1,032 of the run's $1,145.Release Notes
🤖 Generated with Claude Code
Note
Add rules to limit image reads, bound delegated work, and finish background work
Macroscope summarized ff6fbac.