Point the skills at auto mode before manual escalation - #21
Open
StefanoGuerrini wants to merge 4 commits into
Open
Point the skills at auto mode before manual escalation#21StefanoGuerrini wants to merge 4 commits into
StefanoGuerrini wants to merge 4 commits into
Conversation
Measured: with the skills installed, the agent recommended premium_proxy in 6 of 8 answers to a plain 'fetch this Cloudflare page' question, where nobody asked about configuration. That pair with js_render is 25 credits per request against 1 for a basic call. The old text said to start with auto mode, then listed the manual line as a peer example with no cost attached, so it read as the stronger option for a hard target. Manual mode now sits in its own block, carries its price, and is framed as the path after auto mode has already failed.
The extract skill offered --manual --js-render as an ordinary fix for missing fields, with no mention of auto mode and no cost. Auto mode already renders JS when the target needs it, and manual rendering is 5 credits per request against 1.
The failure map sent the agent straight from FETCH_FAILED to --manual --js-render and then --premium-proxy, with no mention of auto mode and no cost. Auto mode escalates on its own and bills only for the configuration that succeeds, so it is the first retry, not the last resort.
The multipliers live in cost-control. Repeating the numbers in three more skills means they go stale in three more places the next time pricing moves, and a skill that quotes a wrong price is worse than one that quotes none. The wording keeps what made the change work: the cost warning stays at the point of temptation, next to the flag, and it still says which direction is expensive. Only the figure moves to the one file that owns it.
StefanoGuerrini
changed the title
StefanoGuerrini
marked this pull request as ready for review
September 9, 2026 15:31
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Problem
Given a plain "fetch this bot-protected page" question, with no mention of configuration or cost, an agent reading our skills recommended
premium_proxyin 7 of 8 runs.That pair with
js_renderis the most expensive configuration the API offers.mode=autoexists so the agent never has to make that call itself, and it bills only for the configuration that succeeds.Cause
Three skills sold escalation as ordinary practice.
protected-fetchsaid to start with auto mode, then listed this as a peer example with no cost attached:"Full manual control" reads like the serious option for a hard target, and the cost model lived in a different file. Advice at the point of temptation beats advice somewhere else.
extractoffered--manual --js-renderas an ordinary fix for missing fields, never mentioning that auto mode already renders JS when the target needs it.trace-debugsent the agent fromFETCH_FAILEDstraight to--manual --js-renderand then--premium-proxy, with no mention of auto mode at all.Change
Manual mode moves out of the auto examples into its own block and is framed as what you reach for after auto mode has already failed and a trace shows why.
extractandtrace-debugnow point at auto mode first. Theprotected-fetchfrontmatter description carries it too, since that part is always loaded.No prices are repeated. Each warning names the direction and points at
cost-control, which owns the multiplier table. A figure copied into four files goes stale in four files.Result
Measured with the harness in #20, 8 runs, one variable changed:
Every run now chooses auto mode, and most say so explicitly: "no JS rendering or premium proxies on the first attempt, since auto mode handles the escalation".
Scope
Three markdown files under
skills/. No code, and no behaviour change in the CLI itself. It reaches anyone already receiving these skills throughzenrows plugin installor theagent-pluginpackage, independent of any other change.#20 stacks on this branch and adds a deterministic test that fails the build if a skill regresses this way again.