Repository navigation
fix(extension): holistic review fixes for #241 - #248
Merged
Merged
Conversation
API (unreleased, reshaped before 0.25):
- impl Extension for Arc<T>: share an extension, read Budget::spent_usd
after installing, pass ToolContext::tree_extensions() on.
- AgentLoopConfig::delegated_from(&ToolContext) / Agent::delegated_from:
a custom delegation tool's child gets the tree (depth, label, inherited
extensions whose tools are not offered); SubAgentTool uses it.
- on_event moves to RunHooks (sync, &self), so per-run state stays in the
hooks; an advisory on_event that panics is not called again.
- RunOutcome: private fields, end() -> RunEnd { Completed, Stopped,
Rejected, Cancelled, Failed { error, extension } }, stop_reason().
- Stateless -> ClonedHooks, with filters_tool_output() and
rechecks_modified_calls() builders.
- EXTENSION_MESSAGE_PREFIX = "[Extension message: ".
- Budget: rejects a negative or NaN limit, with_name, counts a sub-agent's
reported spend unless it runs in the child too; per-run spend lives in
the run's hooks (no leak on a dropped run).
Safety:
- A pending required failure stops tool calls not started yet (before
tools, between sequential calls and batches).
- Every hook races the run's cancel; finish is bounded (FINISH_TIMEOUT).
- A run cancelled before its first turn closes that turn.
- InputContext.text excludes the filters' warnings.
- name()/mode() panics are contained; logs carry run_id and tool_call_id;
a dead observer is an error; SubAgentTool fails when its event forwarder
fails instead of reporting an empty success.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01T7iq5hpndSiHQcnAsywKuG
Tests (extension_test 58, budget 11, decision +1, wasm32 +1): - a hand-written delegation tool runs its child under the tree (an Arc'd tree policy, depth 1, tools not offered); - a required failure on a response stops its tool calls, and after_tool failing stops the next sequential call; - a run cancelled before its first turn closes it; a hung hook does not hang a cancelled run (each mutation-checked); - loop-injected prefix, an advisory on_event switched off after a panic, on_input without filter warnings, ClonedHooks filtering output; - a tree redactor across a sub-agent, the output store holding only the filtered result, required panics in on_input/after_tool/finish, recheck scope, older hooks first, on_stop with follow-ups and limits, finish seeing every event, after_tool skipped for denied/unknown calls, a provider error in finish, on_event before the consumer, multi-thread variants; - Budget: read through an Arc, a per-run budget counting a sub-agent's spend, the exact limit, a named budget, negative/NaN refused; - a tree gate rechecking a rewritten call; - wasm32: an extension gating and observing a run on the host. Docs: extensions.md rewritten for the final API (examples compile: ModelConfig import, no shadowed Budget), the input guard's position, the on_stop/before_model rules, turn-hook note order, delegation via delegated_from, cancellation, wasm32 panics, Length; rustdoc for the gate, guard, advisor, is_loop_injected, ToolContext; struct listings; TurnEnd row; CLAUDE.md; CHANGELOG (behaviour changes from #243 flagged). Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01T7iq5hpndSiHQcnAsywKuG
- A cancel during the input stage is not a rejection: a cancelled on_input
lets the run start, and it ends Cancelled at its first check, before any
model request (was InputRejected).
- Budget decides per delegation which child spend it already counted: a
child run records parent_run_id/call_id (RunContext::{parent_run_id,
delegated_by}, new), instead of one sticky flag on the shared Budget
that stopped another agent counting its sub-agents' spend.
- run_outcome: a run whose model finished stays Completed even if the token
is cancelled afterwards.
- Docs: on_event can wait behind a running &mut hook; cancel semantics.
- Tests: cancel before the input check ends Cancelled (mutation-checked);
on_event strictly before the consumer (multi-thread); a shared budget
across a tree agent and a plain agent counts each child once.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01T7iq5hpndSiHQcnAsywKuG
- A SubAgentTool child cancelled before any assistant message fails the delegation with ToolError::Cancelled instead of reporting an empty success (the pinned test_sub_agent_cancellation expected the old behaviour and is updated; docs/sub-agents.md and CHANGELOG say so). - RunContext::inherited (in start_run): the extension came from the calling run's tree. Budget records a child delegation only then, so no key is left behind when the parent does not hold the budget. - The on_event ordering test sleeps before checking the consumer, so a forward-first observer fails it (mutation-checked). - Stale doc comment on ActiveExtensions::on_input. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01T7iq5hpndSiHQcnAsywKuG
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.
Fixes everything the holistic five-reviewer pass over the merged #241 work found (#242–#247). The extension API is unreleased, so the reshaping happens now, before 0.25.
API (commit 1)
impl Extension for Arc<T>: share an extension, readBudget::spent_usd()after installing, pass tree extensions on.AgentLoopConfig::delegated_from(&ToolContext)/Agent::delegated_from: a hand-written delegation tool's child now gets the tree properly: depth, label, inherited extensions whose tools are not offered. Before, tree policy failed open for these tools.SubAgentTooluses it too.on_eventmoves toRunHooks(sync,&self), so per-run state stays in the hooks.RunOutcome: private fields,end() -> RunEnd { Completed, Stopped, Rejected, Cancelled, Failed { error, extension } },stop_reason().StatelessbecomesClonedHooks, with.filters_tool_output()and.rechecks_modified_calls().EXTENSION_MESSAGE_PREFIX = "[Extension message: "("[Extension "also matched real user text).Budget:with_name;Safety
finishis bounded (FINISH_TIMEOUT).on_inputsees the input without the filters' warnings.name()/mode()panics are contained.run_id/tool_call_id; a dead event observer is anerror!; an advisoryon_eventthat panics is switched off for the run.SubAgentToolfails when its event forwarder fails, instead of reporting an empty success.Tests and docs (commit 2)
Tests:
extension_testgrows from 37 to 58 cases andextension_budget_testfrom 5 to 11, plus one decision test and one wasm32 smoke test (run under Node, 9/9).Docs:
extensions.mdrewritten for the final API. Examples now compile, the guard's position is right, theon_stop/before_modelrules are right, turn-hook note order is stated, plus delegation, cancellation, wasm32 panics andLength.is_loop_injectedandToolContext; struct listings; CLAUDE.md.Deliberately not done: deprecating
ToolGate'sToolMiddlewareimpl andInputGuard's filter impl. That's a separate decision for 0.25.Checks
cargo test --all-features(58 binaries; hook_order unchanged), clippy with--all-featuresand--no-default-features, wasm32 clippy (lib withdecision, and the test target), the wasm32 tests under Node,cargo doc -Dwarnings, fmt. The extension tests passed 10 repeated runs.🤖 Generated with Claude Code
https://claude.ai/code/session_01T7iq5hpndSiHQcnAsywKuG