Dev to main, 1st Oct, 2026 - #136
Merged
Merged
Conversation
…d let a prompt injection use its own words
…ifficulty, varied attacks and complete briefs, and move voice-only guidance to the voice kind
…into the skills, and keep writers busy to the end
…rol cells, and writers never predict the agent or invent its rules
…mes are the writer's job
…our, no undeliverable voice conditions, plain-word proportions, a short review and distinct real names
…ch difficulty once per task, drop rather than overwrite, wider caller spread, and a fame test for names
…on protective refusals, and check the name still fits the content
…dings, people spread and attack share, with writer defaults when a brief is silent
…the ordinary-path branch; no tasks the agent does not handle; no sounds in the caller's words
…, attacks take the channel's form, and writers never reuse the agent's own names or places
…eir words, steps and rules stay inside tasks, and plain directory-style names
…ing them, deal accent spread per brief, and keep option picks and agent questions inside tasks
…g not the difficulty, a changed place alone is a walk-through, and use cases need instructions
…arger than a smoke test
…he skills on details, task levels and the attack share
…ositions are behaviour, and every slice carries a couple of attacks
…luded, appears in each suite, and rare interface levels stay rare
…ever to a scripted reply or a tidy close
… every attack angle several times
…ons, and cooperative callers are the rare case
…und, and some callers are away from home
…y address is a real place there
…ty survives, labels show in the text, and values read real
…nd pressures that compound, with attacks spread across every task
…ven share of the slices
…mes fit the offered accent, and spoken calls name a noise place
…ls, and build-environment keeps the voice limits
…rative level, attackers decide for themselves, and neutral accents fit the name
…as giving an identifier stays one line
… steps are never tasks, and only the caller speaks on a call
…un was generated with
…through and where it cannot
…ction-identifier checks
…ore than a third of a suite
fix(harness): make scenario quality a skill rule, not a wording check
fix(harness): keep guest PIN policy out of scenario distribution
…ise bed cannot play
KarthikAvinashFI
marked this pull request as ready for review
October 1, 2026 12:04
fix(harness): scenarios name the place, never a sound the noise bed cannot play
chore(release): bump agent-learning-kit to 0.2.2
khushalsonawat
approved these changes
Oct 1, 2026
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
4 participants
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.
Summary
Promotes
devtomain. Includes:Your details:block with everything the task needs, in-character callers, real places in the caller's location, name/accent/language from one background), every stated use case covered with negative cases, red-teaming in every suite, background noise on most calls, the share check kept for task/counterparty/disposition/interaction, wording-based submit checks removed, and the caller brings an unanswered question back on the next turn.agent-learning-kitto0.2.2sopublish-pypiruns.Verification
uv run ruff checkon the changed filestests/harnessandtests/test_harness.pyfail only wheredevalready failsLinear
To be linked.