From 779b2d8753f1b9ab4574ca9093f7fee023f79613 Mon Sep 17 00:00:00 2001 From: "Claude Opus 5.5 (claude-opus-5-5)" Date: Mon, 9 Nov 2026 18:32:27 +0100 Subject: [PATCH 1/3] feat: add a rule that limits image reads 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. --- runes/core/rules/LimitImageReads.md | 7 +++++++ 1 file changed, 7 insertions(+) create mode 100644 runes/core/rules/LimitImageReads.md diff --git a/runes/core/rules/LimitImageReads.md b/runes/core/rules/LimitImageReads.md new file mode 100644 index 0000000..d117bb5 --- /dev/null +++ b/runes/core/rules/LimitImageReads.md @@ -0,0 +1,7 @@ +Read each image one time. A screenshot costs more than a thousand input tokens, and every later request in the session sends it again. + +Before a task that needs images, list the files and pick only the images that answer the question. Prefer one contact sheet or montage over many single frames, and prefer a key frame over the full frame sequence of a recording. When you read an image, write one line that says what it shows, and use these notes after that. When you ask another agent to view images, name the images or give one contact sheet. Put these limits in its brief. + +Do not read an image again to check it. If an image seems to be gone from your context, or you think that it did not render, use your notes and continue. A harness can remove older images from a long session and show a marker such as `[media removed: request limit]` in their place. That marker is not an error, and a new read only removes another image. Do not request the image again. + +After 20 image reads in one task, stop and decide if you need more. Never read more than 100 images in one task. If the task needs more, tell the user and ask how to continue. From cdb52444a6b549f51fdb497a8b25390cb1ca3802 Mon Sep 17 00:00:00 2001 From: "Claude Opus 5.5 (claude-opus-5-5)" Date: Mon, 9 Nov 2026 18:32:27 +0100 Subject: [PATCH 2/3] feat: add a rule that bounds delegated work 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. --- runes/core/rules/BoundDelegatedWork.md | 7 +++++++ 1 file changed, 7 insertions(+) create mode 100644 runes/core/rules/BoundDelegatedWork.md diff --git a/runes/core/rules/BoundDelegatedWork.md b/runes/core/rules/BoundDelegatedWork.md new file mode 100644 index 0000000..0c42770 --- /dev/null +++ b/runes/core/rules/BoundDelegatedWork.md @@ -0,0 +1,7 @@ +Give every task that you delegate a limit and a stop condition. A child agent cannot see what the work is worth, and it can repeat one step for hours while you wait. + +Write the limit into the brief: the inputs to read, the most tool calls or minutes the task may use, and the shape of the report. Base the limit on a similar task that finished, and add a margin. Do not ask for every file in a directory when a few files answer the question. LimitImageReads sets the limits for images. + +Tell the child to stop and report when it reaches the limit. Tell it also to stop when it calls the same tool with the same input a third time. A partial report that names the cause is a good result. + +When the child returns, compare its tool calls and time with the limit. Report an overrun to the user with the numbers. Change the brief before you delegate the task again. From ff6fbac51a3fd2479c5be6d539e6cbbd29da3c84 Mon Sep 17 00:00:00 2001 From: "Claude Opus 5.5 (claude-opus-5-5)" Date: Mon, 9 Nov 2026 18:32:27 +0100 Subject: [PATCH 3/3] feat: add a rule to finish background work before a report 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. --- runes/core/rules/FinishBackgroundWork.md | 3 +++ 1 file changed, 3 insertions(+) create mode 100644 runes/core/rules/FinishBackgroundWork.md diff --git a/runes/core/rules/FinishBackgroundWork.md b/runes/core/rules/FinishBackgroundWork.md new file mode 100644 index 0000000..4c87180 --- /dev/null +++ b/runes/core/rules/FinishBackgroundWork.md @@ -0,0 +1,3 @@ +Do not end your turn while a command that you started in the background still runs. The user or the parent agent reads your report as final, and a later report makes them check the work again. + +Before the final report, wait for each background command that the report depends on. Read its result and put it in the report. Use the harness's wait tool, or run the command in the foreground when it ends within the tool timeout. If you must end the turn first, stop the command, or say that it still runs, where its log is, and what is not verified.