Repository navigation
feat(extension): dogfood the decision features and add Budget (#241 PR 2) - #247
Merged
Merged
Conversation
AgentLoopConfig::new(provider, model) sets every other field to the defaults the struct-literal examples used. The struct is non_exhaustive, so new fields stop being breaking changes for agent_loop callers; fields stay public. Every in-tree literal (Agent, SubAgentTool, 30 test sites) and the docs now use new(). transform_context is kept (#241 needs it until on_context exists). Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01T7iq5hpndSiHQcnAsywKuG
…non_exhaustive] in listings Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01T7iq5hpndSiHQcnAsywKuG
Extension (a factory: name, mode, filters_tool_output, rechecks_modified_calls, start_run, on_event) and RunHooks (fresh per run: tools, on_input, before_model, before_tool, after_tool, on_stop, finish). Advisory or required failure handling (a required failure ends the run with an Error message prefixed EXTENSION_FAILED_PREFIX), partial tool output withheld while an extension filters it, a rechecking policy judges rewritten arguments again, on_stop continues capped by max_stop_continues. Agent / SubAgentTool::with_extension, and with_tree_extension to apply an extension to every delegated run at any depth (host policy reaches sub-agents; the tree's tools stay with the parent). The existing hooks are unchanged and the hook-order pin test passes unmodified. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01T7iq5hpndSiHQcnAsywKuG
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01T7iq5hpndSiHQcnAsywKuG # Conflicts: # CHANGELOG.md # src/agent.rs # src/sub_agent.rs
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01T7iq5hpndSiHQcnAsywKuG
- Required failures are never lost: settle() (flush the event observer,
then take the failure) at the top of each turn before the limit and
on_before_turn checks, before on_stop, and after run_loop on any path.
- on_event is deterministic: the observer acknowledges a flush only after
every event sent before it was observed; finish flushes first.
- A required verifier outvoted by an advisory Continue still fails at the
cap; an extension that filters tool output and cannot start fails the
run; started extensions still get finish when a later start fails.
- after_tool gets ToolOutput { result, is_error }; ExtensionMode is
non_exhaustive; run_label passes to delegated runs (ToolContext::run_label);
a failure outside a turn gets its own TurnStart/TurnEnd; required
failures log at error!; ModelHalt replaces the unreachable arm.
- 11 more tests (34): lost-failure, on_event determinism, outvoted
verifier, broken redactor, finish per ending, transcript after a stop,
child denial honoured and tree tools not offered, child required failure,
custom delegation tool, real concurrency (barrier), after_tool is_error.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01T7iq5hpndSiHQcnAsywKuG
…eard - fail_run latches the run as failed: a failure recorded while the first is reported (an on_event that fails on those very events) is only logged, so there is one Error message and one on_error. - on_stop Continue sends one line per extension that asked, so a required verifier's message reaches the model even when an advisory one asked first. - Tests: a sink that panics on every MessageEnd fails the run exactly once (mutation-checked); both continue messages are sent. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01T7iq5hpndSiHQcnAsywKuG
…R 2) - ToolGate, InputGuard and the decision advisor implement Extension; with_tool_gate / with_input_guard / with_decision_model install them as extensions after the agent's own (the gate runs last and judges arguments an extension rewrote). Their old trait impls stay for installing by hand. - extension::Budget: a dollar limit, priced from each assistant MessageEnd's usage, checked in before_model; per run or across_runs (a session, or a tree with with_tree_extension); for_model is None for an unpriced model. - Tests: the gate sees an extension's rewrite; budget per run, across runs, across a delegation tree, unpriced. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01T7iq5hpndSiHQcnAsywKuG
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01T7iq5hpndSiHQcnAsywKuG # Conflicts: # CHANGELOG.md # CLAUDE.md # docs/concepts/extensions.md # src/agent.rs # src/extension.rs # src/sub_agent.rs
- before_tool / after_tool take &self and run under a read lock, so the calls of one response are judged concurrently again (a slow gate no longer serializes parallel calls); the other hooks keep &mut self. - ToolGate rechecks modified calls, so a gate installed as a tree extension also judges an extension's rewrite. - Budget::spent_usd() is the across-runs total (None per run); the CHANGELOG and test name say when Budget really stops; docs note that a failed mid-stream attempt reports no usage. - Tests: parallel calls judged concurrently (a barrier; serializing them times out); the tree budget test now fails unless the spend is shared (child request count asserted). Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01T7iq5hpndSiHQcnAsywKuG
…efore tools run - The guide's first example used &mut self for before_tool. - Per-call hooks need interior mutability for state: said in the guide. - A gate installed as a tree extension can cost a second decision request when a later extension rewrites the call: said. - Extensions observe a turn's response before its tools run, so a sub-agent sees the tree's spend so far. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01T7iq5hpndSiHQcnAsywKuG
This was referenced Oct 6, 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
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.
PR 2 of #241: the test of the design. If yoagent's own features can't be written as extensions, the API is wrong. They can.
Dogfooding
ToolGate,InputGuardand the decision advisor implementExtension:ToolGatebefore_tool(samedecide)with_tool_gateInputGuardon_input(samescreen)with_input_guardbefore_modelnote (sameadvise, memo shared across the agent's runs)with_decision_modeldecision::wirenow returns extensions, appended after the agent's own, inAgent::build_configandSubAgentTool.ToolMiddleware,AsyncInputFilter) for installing by hand.Changed):gate_judges_the_arguments_an_extension_rewrote.extension::Budgetbefore_model, before the request that would exceed it, with[Agent stopped: budget of $… spent …].MessageEnd's usage, throughon_event, priced with oneCostConfig. feat: Extension — one plug-in contract for the agent lifecycle (#241 PR 1) #246's flush barrier makes this deterministic..across_runs()(one total for a session, or for a whole delegation tree withwith_tree_extension).Budget::for_model(max, &model)returnsNonefor an unpriced model, because a budget without a price is no limit.tests/extension_budget_test.rscovers per run, across runs, a tree including a sub-agent's spend, and an unpriced model.Not in this PR
yoagent-rutiscollapsing into oneimpl Extension. It touches the same bridge files as #240 (rutis 0.6), which is on hold. It follows once #240 lands.Checks
cargo test --all-features(58 binaries);--all-featuresand--no-default-features;decision;cargo doc -Dwarnings, fmt.Docs: extensions.md (Budget, "Built on extensions"), decision-models.md, CHANGELOG, CLAUDE.md.
🤖 Generated with Claude Code
https://claude.ai/code/session_01T7iq5hpndSiHQcnAsywKuG