feat(agents): harden picker ranking, pruning, and status chips - #60
Merged
AnzoBenjamin merged 6 commits intoAug 18, 2026
Merged
Conversation
Rank file-lister/file-picker programmatically and emit set_output instead of an LLM mock path. Keep discovery-linked constraints ahead of file-picker facts, spawn context-pruner in threshold e2e via spawn_agent_inline, hide healthy index chips behind overflow-aware status-bar selection, and accept spawn_agent_inline on dynamic-agent templates.
|
The latest updates on your projects. Learn more about Vercel for GitHub.
|
The selectStatusBarChips shrink ladder could exit while the chip cluster still exceeded the width budget. With a warning or error index chip present, the ladder ran out of steps after dropping cost, model, git, and context, so the cluster measured 18 against a budget of 17 and the test-cli overflow test failed in CI. Add a last-resort step that drops the timer even during an active run, after the context-drop step. The index chip is never dropped or relabeled by that step, so an index alert still survives the narrowest rows. Follow-up hardening from review: export shared statusBarClusterWidth, statusBarChipBudget, and STOP_BUTTON_WIDTH so tests reuse the production width math instead of duplicating it; name the budget constants; truncate model names by stringWidth on grapheme cluster boundaries; drop an index chip that would clamp to an uninformative ellipsis-only label; and derive the context label from a single source shared with the overflow-shorten step. Expand status-bar-chips coverage: tone thresholds, sub-cent, zero, and negative cost, zero-diff git and deletions, xs timer, stop-hint, and non-error index paths, active versus idle timer priority, budget invariants, and wide-character and ZWJ truncation. 48 pass / 0 fail locally, typecheck green.
createBase2 no longer takes progressiveToolDisclosure; base2 always ships the full mode-resolved surface and the only narrowing control is the new optional unlockedTiers option. It pins progressiveToolDisclosure: false and publishes fullToolSurface from the default all-tier resolve, so stale persisted unlockedToolTiers cannot narrow a resumed surface. Tier constants and the UnlockedToolTier type now live solely in packages/agent-runtime/src/util/base2-tool-tiers.ts and the template derives its surface from them, removing FULL_SURFACE_ORDER and the duplicated list that had to stay in sync. filterByUnlockedTiers now requires an explicit ceiling (a predicate or the new ALLOW_ALL_TIER_TOOLS sentinel) so a caller with no ceiling can no longer fail open into allow-all, and getEffectiveAgentToolNames fails closed when programmaticConfig.fullToolSurface is absent. Also drops the inert OPENBUFF_PROGRESSIVE_PROMPT_DISCLOSURE and OPENBUFF_PROGRESSIVE_TOOL_DISCLOSURE env canaries and the docs advertising them.
Unify skipIfMissing description and deletion-only validation across str_replace and edit_transaction schemas so provider and input surfaces cannot drift and both reject non-empty newString cleanly. Add structured anchor_scope_mismatch and capability_scope failureKind handling, anchored candidate line offsets, bounded occurrence walks, and ordering fixes that report idempotent cleanup as no-op skips instead of scope mismatches or large-file blocks. Harden the str_replace handler with shared circuit-breaker guidance, centralized failure-budget accounting, structured errorCode/recovery payloads, and an all-skip zero-patch guard so mixed batches preserve real changes while fully-skipped cleanups succeed without file writes.
The schema commit b124f29 added the skipIfMissing doc comment for idempotent deletion replacements. Regenerate the four tool-definition sources via generate-tool-definitions so the CI 'Check generated tool definitions are current' step passes.
A failed str_replace under a context_compacted re-read marker left the sticky whole-file read authorization intact, so a later write_file could overwrite the file from a hash the model can no longer observe. The failure path now re-marks the path as requiring a fresh read while markEditRequiresFreshRead preserves the original context_compacted reason and sourceTool instead of downgrading it. This fixes the failing agent-runtime read_files edit-state recovery test. Also hardens the handler: empty-but-valid files are authorized by the auto-re-read gate instead of failing closed on an unhashable-content check, the structuralRecovery failure budget is released on the failure path and not only after a successful apply, stale content is no longer reused for auto-re-read when recovering from a failed edit, and untyped result casts are replaced with a typed StrReplaceResultWithMetadata intersection. Adds circuit-breaker tests for failed-recovery budget release and empty-file auto-re-read authorization, and makes the unreachable client stub throw so an accidental call fails the test.
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.
Rank file-lister/file-picker programmatically and emit set_output instead of an LLM mock path. Keep
discovery-linked constraints ahead of file-picker facts, spawn context-pruner in threshold e2e via
spawn_agent_inline, hide healthy index chips behind overflow-aware status-bar selection, and accept
spawn_agent_inline on dynamic-agent templates.
This change is