fix: repair the #1331 merge — one generation of #1295, one engine, one bus

[macanderson]

What & why

main has not compiled since #1331. That branch merged main in and resolved three conflicts by keeping both sides, which left two crates unparseable, docker-serve red, and a shipped knob unreachable. Every fix below takes the side main already had.

#1295 had two generations, and both landed. #1328 shipped the refined design to main — judge_evidence_demand, gated on a tracked command so the ask is only raised where it can be answered (#1211 §1). The #1331 branch still carried its own earlier draft, require_evidence_for_lone_judge_pass. The merge kept both config fields, both CLI knobs, and spliced both bodies into verify_candidate, leaving a revise_candidate( call unterminated. The older draft is removed — its const, field, CLI plumbing, and its two tests, whose four cases tests/judge_evidence_demand.rs already covers, plus the no-tracked-command case they never had.

That draft's TOML plumbing had displaced the shipped knob's, so pipeline_judge_evidence_demand parsed as an unknown key, never reached PipelineConfig, and was dropped by AgentEngineConfig::overlay — the same silently-lower-scope-wins bug the headless_scope_bypass comment describes. Renaming that plumbing restores it. This one is a silent behavior regression, not a compile error, so it would have survived a green build.

ServeConfig lost a field initializer. extensions (#1298) was spliced into the body of with_sub_agents (#1297), which then returned nothing.

run_session built two engines. One over &tools carrying the hook bus and calibration map (#1298); one over the delegating tool_view that makes sub-agents reachable (#1297), carrying neither. They are now one engine over tool_view with both attachments. The duplicated drive_turn after the goal-run match is gone, and the bus now reaches every round of a goal run and every sub-agent's tool executor — an operator policy that can refuse a tool call must not be escapable by delegating it.

Also: the SessionSpec literals in both serve suites carry the other side's fields; api_sources() counts the two files it gained; judge_evidence_demand is switched off in the OOM scenario the way its siblings already do; and routes.rs — 1521 lines once both sides' routes landed — moves /v1/sessions into a submodule, with the two now-obsolete size-baseline entries retired.

Refs #1331

The witness

• No witness needed (merge repair / CI) — because the witness already exists on both sides and was not running: stella-pipeline and stella-serve did not compile, so cargo test, clippy, doc, fmt and the container build all failed before reaching an assertion. The restored suites are the proof, including tests/judge_evidence_demand.rs, which pins the surviving generation's four behaviours and never ran on main.

Verified that no shipped feature was dropped: every file added by #1330 is present and byte-identical to main, and the key symbols of #1290, #1291, #1293, #1294, #1295, #1297, #1298 and #1301 all resolve.

The gate

• cargo fmt --check
• cargo clippy --workspace --all-targets -- -D warnings
• cargo test --workspace — 108 binaries, 0 failures
• Docs updated where behavior/flags changed (doc comments on the removed knob's references now name judge_evidence_demand)
• CLA signed
• Also run locally: check-file-size.sh, check-wire-schema.sh, RUSTDOCFLAGS="-D warnings" cargo doc, make cache-correctness, stella context validate

Refs, not Closes: this repairs a merge rather than finishing an issue.

Ground-rule check

• No I/O added to stella-core; no new deps
• No new outbound network calls

Anything reviewers should know?

The one judgement call is which generation of #1295 survives. Both compile in isolation, so this could not be settled by the build. The evidence: git log -S shows evidence_demand_is_worth_a_turn entered main with #1328 (9a509cc), and require_evidence_for_lone_judge_pass has no history on main at all — it existed only on the #1331 branch. The surviving generation is also the stricter one (bounded to one demand per candidate, and only raised when a tracked command could answer it) and is the one whose default is on. If that reading is wrong, the revert is contained to stella-pipeline/src/pipeline.rs and the CLI settings plumbing.

Second call: routes.rs needed to lose ~21 lines to clear the 1500-line ratchet. Splitting out its #[cfg(test)] mod tests would have cleared it by 33 lines; extracting the four /v1/sessions handlers instead leaves ~330 lines of headroom, which matters because both merge parents were actively adding routes to this file. The handlers are re-exported, so server.rs's dispatch is unchanged — the move is mechanical, with no logic edits.

The size baseline's obsolete entries were retired with make file-size-update, which also retightened several stale ceilings. I confirmed against both merge parents that those files genuinely are that size and no content was lost in the merge.

---

Generated by Claude Code

Summary by Sourcery

Repair post-#1331 merge regressions by removing the obsolete pipeline evidence knob generation, fixing serve/session engine wiring and configuration initialization, and splitting /v1/sessions routes into a dedicated module while keeping APIs and tests in sync.

Bug Fixes:

• Restore ServeConfig extensions initialization and keep with_sub_agents returning a complete config
• Ensure run_session builds a single engine over the delegating tool view and propagates the hook bus and calibration into goal runs and sub-agent executors
• Fix CLI/TOML plumbing so the pipeline judge evidence demand knob is parsed correctly and reaches PipelineConfig
• Disable judge evidence demand in the OOM verification-hardening scenario to match other OOM behaviours

Enhancements:

• Extract /v1/sessions handlers and wire types from routes.rs into a sessions submodule, re-exporting handlers to keep call sites stable
• Propagate the operator hook bus through goal runs so every round is fully observable
• Align calibration reporting and inspection text with the judge evidence demand knob name

Build:

• Update file-size baseline and add routes/sessions.rs to keep routes.rs under the configured line-count limit
• Expand parity test API sources to include new serve routes and suites

Documentation:

• Update inline documentation and comments to reflect the surviving judge evidence demand generation and new naming

Tests:

• Adjust serve SessionSpec literals in goal and hooks tests to carry extensions and calibration consistently
• Update pipeline judge evidence demand tests to explicitly disable coverage tooling

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

1 Skipped Deployment

Project	Deployment	Actions	Updated (UTC)
stella-cli-docs	Ignored		Aug 4, 2026 4:17am

[sourcery-ai]

Reviewer's Guide

Merge repair that restores compilation and intended behaviour after #1331 by removing the earlier #1295 evidence knob implementation, wiring the shipped judge_evidence_demand knob correctly through CLI/TOML/pipeline, fixing ServeConfig initialization and session engine construction, ensuring the hook bus/calibration reach goal runs and sub-agents, and extracting the /v1/sessions handlers from routes.rs into a new submodule to satisfy file-size constraints.

Sequence diagram for goal runs and sub-agents sharing one engine and hook bus

sequenceDiagram
    actor Host
    participant routes_handle_session_turn
    participant Session
    participant run_session
    participant Engine
    participant HookBus
    participant drive_goal
    participant drive_turn
    participant RemoteToolExecutor

    Host->>routes_handle_session_turn: POST /v1/sessions/{id}/turns
    routes_handle_session_turn->>Session: Session::start(SessionSpec{extensions,sub_agents,calibration})
    Session->>run_session: run_session(spec,bus)

    run_session->>RemoteToolExecutor: RemoteToolExecutor::new(advertised,frame_tx,pending,spec.reverse_request_timeout,bus.clone)
    run_session->>Engine: Engine::with_sleeper(provider,tool_view,spec.config,sleeper)
    run_session->>Engine: Engine::with_gate(gate)
    run_session->>Engine: Engine::with_steering(steering)
    run_session->>Engine: Engine::with_bus(bus)
    run_session->>Engine: Engine::with_calibration(spec.calibration)

    alt spec.goal is Some
        run_session->>drive_goal: drive_goal(engine,spec.messages,run,spec.budget,event_tx,cancel,bus.as_ref)
        loop each_round
            drive_goal->>Engine: Engine::with_turn_instance(base_turn_offset)
            drive_goal->>drive_turn: drive_turn(round_engine,messages,budget,events,round_cancel,bus)
        end
    else no goal
        run_session->>drive_turn: drive_turn(engine,spec.messages,spec.budget,event_tx,cancel,bus.as_ref)
    end

Loading

Flow diagram for wiring the judge_evidence_demand knob from CLI/TOML into PipelineConfig

flowchart LR
    CLI[AgentEngineConfig.pipeline_judge_evidence_demand]
    TOML[AgentsSection.pipeline_judge_evidence_demand]
    Overlay[AgentEngineConfig::overlay]
    Pipeline[PipelineConfig.judge_evidence_demand]

    CLI --> Overlay
    TOML --> Overlay
    Overlay --> Pipeline

Loading

File-Level Changes

Change	Details	Files
Consolidate #1295 to the shipped judge_evidence_demand implementation and remove the older require_evidence_for_lone_judge_pass generation.	Delete the lone-judge evidence request constant and its pipeline config field/default. Remove the CLI and settings plumbing, TOML fields, and helper accessors for the old knob. Update calibration and ground-truth comments to reference judge_evidence_demand instead of the removed knob. Drop the dedicated tests for require_evidence_for_lone_judge_pass, relying on the existing judge_evidence_demand suite.	stella-pipeline/src/pipeline.rs stella-pipeline/src/pipeline/tests/verification_hardening.rs stella-cli/src/settings.rs stella-cli/src/settings/toml_config.rs stella-cli/src/agent/engine.rs stella-pipeline/src/replay.rs stella-pipeline/src/replay/ground_truth.rs stella-serve/src/goal.rs stella-cli/src/inspect.rs stella-cli/src/settings/unknown.rs stella-pipeline/src/pipeline/tests/judge_evidence_demand.rs
Restore correct wiring for the pipeline_judge_evidence_demand knob through CLI settings and TOML so it reaches the pipeline config.	Rename the TOML and unknown-field entries to pipeline_judge_evidence_demand. Ensure AgentsSection raise/lower functions and AgentEngineConfig overlay use the pipeline_judge_evidence_demand field. Update comments and unknown-field whitelists to match the new field name.	stella-cli/src/settings.rs stella-cli/src/settings/toml_config.rs stella-cli/src/settings/unknown.rs
Fix ServeConfig initialization and run_session engine construction so extensions are preserved and a single engine over the delegating tool view carries gate, steering, hook bus, and calibration.	Restore the extensions field initialization in ServeConfig::default and make with_sub_agents return self without clobbering extensions. Change run_session to build one Engine over tool_view instead of two different engines. Attach the hook bus and calibration map to that single engine and pass the bus into drive_turn and drive_goal so hooks see all turns and goal rounds. Update tests to set SessionSpec.extensions and calibration explicitly in relevant suites.	stella-serve/src/server.rs stella-serve/src/session.rs stella-serve/src/goal.rs stella-serve/tests/goal_and_subagents.rs stella-serve/tests/hooks.rs
Extract /v1/sessions route handlers and wire types from routes.rs into a new routes::sessions submodule to reduce file size while preserving API behaviour.	Create stella-serve/src/routes/sessions.rs containing the four session handlers and their request/response structs, reusing the parent module imports via use super::*. Remove the duplicated handler and type definitions from routes.rs and add a mod sessions; plus pub(crate) use re-exports for all four handlers. Update file-size baseline and parity tests to account for the new file and additional API source files.	stella-serve/src/routes.rs stella-serve/src/routes/sessions.rs scripts/file-size-baseline.txt stella-parity/src/lib.rs
Tighten test configuration and wiring around OOM and judgment evidence demand scenarios.	Explicitly disable judge_evidence_demand in the persistent memory kill OOM scenario PipelineConfig used in verification_hardening tests. Ensure judge_evidence_demand tests run with no coverage tooling to keep axes independent. Update SessionSpec literals in serve tests to carry extensions/goal/sub_agents/calibration fields consistently.	stella-pipeline/src/pipeline/tests/verification_hardening.rs stella-pipeline/src/pipeline/tests/judge_evidence_demand.rs stella-serve/tests/goal_and_subagents.rs stella-serve/tests/hooks.rs

---

Tips and commands

Interacting with Sourcery

• Trigger a new review: Comment @sourcery-ai review on the pull request.
• Continue discussions: Reply directly to Sourcery's review comments.
• Generate a GitHub issue from a review comment: Ask Sourcery to create an
  issue from a review comment by replying to it. You can also reply to a
  review comment with @sourcery-ai issue to create an issue from it.
• Generate a pull request title: Write @sourcery-ai anywhere in the pull
  request title to generate a title at any time. You can also comment
  @sourcery-ai title on the pull request to (re-)generate the title at any time.
• Generate a pull request summary: Write @sourcery-ai summary anywhere in
  the pull request body to generate a PR summary at any time exactly where you
  want it. You can also comment @sourcery-ai summary on the pull request to
  (re-)generate the summary at any time.
• Generate reviewer's guide: Comment @sourcery-ai guide on the pull
  request to (re-)generate the reviewer's guide at any time.
• Resolve all Sourcery comments: Comment @sourcery-ai resolve on the
  pull request to resolve all Sourcery comments. Useful if you've already
  addressed all the comments and don't want to see them anymore.
• Dismiss all Sourcery reviews: Comment @sourcery-ai dismiss on the pull
  request to dismiss all existing Sourcery reviews. Especially useful if you
  want to start fresh with a new review - don't forget to comment
  @sourcery-ai review to trigger a new review!

Customizing Your Experience

Access your dashboard to:

• Enable or disable review features such as the Sourcery-generated pull request
  summary, the reviewer's guide, and others.
• Change the review language.
• Add, remove or edit custom review instructions.
• Adjust other review settings.

Getting Help

• Contact our support team for questions or feedback.
• Visit our documentation for detailed guides and information.
• Keep in touch with the Sourcery team by following us on X/Twitter, LinkedIn or GitHub.

[sourcery-ai]

Sorry @macanderson, you have reached your weekly rate limit of 500000 diff characters.

Please try again later or upgrade to continue using Sourcery