Slim Buzz agent context - #6128
Closed
salman1993 wants to merge 3 commits into
Closed
Conversation
Signed-off-by: Salman Mohammed <smohammed@squareup.com>
Signed-off-by: Salman Mohammed <smohammed@squareup.com>
Revert engram_fetch.rs, pool.rs, and config.rs to main. The once-per-agent-process onboarding latch was a change to memory injection semantics, not to prompt content, and belongs in its own PR where the session-invalidation behavior can be reviewed on its own terms. This also drops a stale --no-memory doc comment that described the intermediate no-nudge behavior. Restore two rules to the always-on base prompt. Both guard states in which the agent cannot discover a reference: an unreported todo is a silent stall, and a context compaction destroys the reasoning that would tell the agent to go read recovery guidance. The todo rule is worded as "before you start" rather than the original "before sending the pickup acknowledgment" so it does not contradict the messaging rule that silence beats an acknowledgement-only message. Signed-off-by: Salman Mohammed <smohammed@squareup.com>
Contributor
Author
|
closing this PR since it didn't improve the benchmark, see results in #6161 |
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Why
Every agent turn loaded a ~16 KB prompt that was mostly procedure for tasks the turn wasn't doing. The goal isn't fewer tokens — it's a base prompt with only what an agent needs on every turn, and everything else loaded on demand.
What changed
Split into three layers. The base prompt keeps the session model, turn contract, messaging rules, and a list of
buzzcommand groups.buzz-cli/SKILL.mdis now just a router. Seven references (messaging,agents,repositories,memory,recovery,contracts,workspace) hold the details and are read one at a time.What stays in the base prompt is based on whether the agent could find the reference when it needs it. Two cases can't be deferred: silent failures (
--content 'a\nb'posts literal backslashes and succeeds, so nothing tells the agent to go read anything) and context loss (a compaction wipes out the reasoning that would say "read the recovery doc"). Rules that surface through a failed command are safe to defer.[Context]now gives a command instead of instructions. Was four sentences ofIMPORTANT:prose; nowReply: buzz messages send --channel <uuid> --reply-to <id> --content <message>. Event blocks drop what the batch already says once — channel identity,npubnext to hex, andh/p/thread-etags that are already parsed intoThread:andMentions:. Unknown tags are kept asMetadata:.Skill description says what it does, not what it isn't. Retrieval matches on that field. The "don't load this for ordinary replies" instruction lives in the base prompt instead.
Nest installs and version-refreshes the references next to
SKILL.md, with the existing0700permissions and per-provider symlinks. WorkspaceAGENTS.mdpoints at the references instead of repeating them.Not in this PR
Memory injection is unchanged —
engram_fetch.rs,pool.rs, andconfig.rsare identical tomain. An earlier revision made the no-core onboarding nudge fire once per agent process. That changes injection semantics, not prompt content, so it should be reviewed separately.Risk
Medium — this changes the prompt and per-turn event context for every managed agent. Reply, mention, and multiline rules stay inline. Prompt contract tests check the base prompt keeps its key rules and defers the rest; installer tests cover the reference tree, permissions, and version refresh. No relay protocol or event schema changes.
Main risk is under-loading: an agent not realizing it needs a reference. Worth watching in practice.