Spotify's shunt hook blocks non-targeted Claude Code file reads past a configurable line threshold

When check-file-size fires on a Read call, the read does not happen. The hook blocks it and returns a message telling Claude to use the /bulk-reader skill instead ^[11]. Targeted reads pass through untouched, on the reasoning that Claude already knows which section it wants ^[11]. That block is the whole difference from the first attempt, which put the same policy in CLAUDE.md, where by the author's account the rules were advisory, Claude could ignore them, and every project needed its own copy ^[8]. Instructions ask; a PreToolUse hook returns a verdict, and hooks fire before every tool call ^[10].

The routing criterion is file length: 350 lines by default, overridable with SHUNT_MIN_LINES in a shell profile or .claude/settings.json ^[11][13]. Line count is a cheap proxy for how much reasoning a file needs, and a rough one. A dense 400-line parser goes to Flash; a 300-line one stays with the frontier model. The bash-side rule has a shape worth knowing too. check-bash-read intercepts cat, head, tail, less and more on large files, but lets piped commands through on the grounds that they are targeted reads ^[12]. By that definition, appending a pipe converts a blocked read into an allowed one.

Delegation moves tokens rather than deleting them. The worker model reads the file body at its own rate, then Claude pays frontier input rates for the bullets that come back. Which is why "output only the code" is the load-bearing line in the code-writer instructions: the post notes that without it the model wraps output in markdown fences and explanatory prose that Claude then has to parse ^[14][15]. Worker verbosity is billed twice, once as output and again as input.

The cost frame in the post is a survey figure, not the author's bill. A quarter of engineering leaders already spend $200 to $500 per developer per month on tokens, and some are well past $2,000 ^[3]. Take that top cohort literally and it is roughly $24,000 per developer per year; hold the 90 percent and it lands near $2,400 ^[1]. The 90 percent itself is one engineer reporting on his own workload, with no before-and-after token counts published in the post ^[1]. For the ratio to transfer, your spend has to be dominated by whole-file reads and pattern-copy generation rather than debugging, your files have to be long enough to trip the threshold, and the worker output has to be accurate enough that Claude does not quietly redo it.

The post says the fix needed no platform team and no new subscription ^[4]. It does need a Portal instance with the AiKA plugin enabled, because delegation runs through the Portal CLI actions registry ^[9], and a mode whose model field points at something already configured in that instance ^[6]. Portal supplies the ephemeral runtime, described in the post as AWS Lambda but for agents, and keeps the API keys out of the client ^[5]. The excerpt does not state whether Portal is available outside Spotify, or on what terms ^[17]. What copies without it is the enforcement layer: two PreToolUse hooks, a pair of shell wrappers, and any cheaper model reachable from a CLI ^[10][16].