feat: middleware hooks, context injection, mid-run messaging, tool profiles, todo lists by Enreign · Pull Request #87 · emberloom/sparks

Enreign · 2026-03-20T11:27:28Z

Summary

Five new capabilities added to Sparks:

Middleware Safety Net — SparkMiddleware trait with before_model_call / after_spark_complete hooks; ActivityLogFlushMiddleware registered by default. Guarantees deterministic post-run cleanup regardless of LLM behavior.
Rich Context Injection — TicketProvider::fetch_full_context() fetches issue comments (and future PR diffs) for GitHub and Linear at intake time. Opt-in via inject_full_context = true in [ticket_intake].
Mid-Run Message Injection — Messages arriving on Slack/Teams/Telegram while a spark is running are queued and injected as user-role messages before the next LLM call instead of being dropped. Configurable cap via max_queued_messages.
Tool Curation Profiles — Named [tool_profiles] config section; ghost configs reference via profile = "name" eliminating tool list copy-paste. Validated at startup by the doctor command.
Per-Spark Todo Lists — todo_write and todo_check tools for sparks to track tasks during execution. Observable via /session command on Slack and Telegram. Persisted to spark_todos SQLite table on session close.

Test Plan

462 tests passing (up from 433 baseline), 0 failures
cargo clippy — zero new warnings
Spec compliance reviewed per-task (two-stage review: spec then quality)
Quality issues fixed: pub(crate) scoping on executor fields; floor_char_boundary for non-ASCII safety in TicketContext::format
Manual smoke test: enable inject_full_context = true and verify a GitHub ticket spawns with comments in context
Manual smoke test: send a message to a running spark via Slack and verify it appears in the next step

🤖 Generated with Claude Code

5 Open SWE patterns evaluated for Sparks: middleware safety net, rich context injection, mid-run message injection, tool curation profiles, and per-spark todo lists — ordered by impact × effort ROI with module touchpoints and effort estimates per item. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

- Correct middleware hook sites to strategy/react.rs + strategy/code.rs - Fix mid-run inject queue owner to Executor (keyed by session ID) - Add CoreHandle::inject() to item 3 design decisions - Rename tool_allowlist -> tools throughout item 4 - Add src/profiles.rs to item 4 touchpoints - Fix src/strategy.rs -> src/strategy/mod.rs in item 2 Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

1. middleware-safety-net — SparkMiddleware trait + lifecycle hooks 2. rich-context-injection — full ticket context at spark invocation 3. mid-run-message-injection — queue messages during running sparks 4. tool-curation-profiles — named tool profiles in [tool_profiles] 5. per-spark-todo-lists — todo_write/todo_check tools with SQLite persistence Each plan is TDD-ordered, bite-sized steps, exact file paths. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

…eware Introduces the SparkMiddleware trait with before_model_call and after_spark_complete lifecycle hooks, plus an ActivityLogFlushMiddleware no-op placeholder, ensuring the safety net infrastructure is in place for deterministic post-run guarantees. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

…complete Adds SharedMiddlewares field to Executor, wires invoke_after_spark_complete into Executor::run() after the strategy loop completes (on both success and failure paths), and updates manager.rs to pass vec![] at the call site. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Wire the before_model_call middleware hook into every LLM invocation site across ReactStrategy and CodeStrategy (native + text-fallback paths, EXPLORE and VERIFY phases), using docker.session_id() as the session identifier. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

…ls in code.rs Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

…ew; add MiddlewareConfig Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

…ethod Introduces TicketContext struct with comments and diff_summary fields, a format() method with progressive trimming to a char budget, and a default fetch_full_context() on TicketProvider so existing impls need no changes. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

…g loop Add inject_full_context and rich_context_char_cap fields to TicketIntakeConfig with serde defaults (false / 4000). Thread both values through spawn_ticket_intake and apply fetch_full_context + TicketContext::format in the per-ticket dispatch loop, warning on errors rather than failing. Includes a config round-trip test. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

…field Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

…n in doctor Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

…ister in ToolRegistry Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Add session_todos() to ExecutorInjectHandle, CoreHandle, and Manager. Wire todo block into /session responses for both Slack (mrkdwn code block) and Telegram (HTML <pre> block). Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Before removing a session's TodoList on close_session(), serialize it to JSON and emit a structured tracing::info! event (spark_todos persisted) so the todo state is captured in the observer log for post-session analysis. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

… in TicketContext Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

src/todo.rs

+    }
+    async fn execute(&self, _session: &crate::docker::DockerSession, params: &Value) -> Result<crate::tools::ToolResult> {
+        let session_key = crate::executor::Executor::current_activity_context()
+            .map(|c| c.session_key)


src/todo.rs

+    }
+    async fn execute(&self, _session: &crate::docker::DockerSession, params: &Value) -> Result<crate::tools::ToolResult> {
+        let session_key = crate::executor::Executor::current_activity_context()
+            .map(|c| c.session_key)


Enreign and others added 29 commits March 20, 2026 10:30

feat(middleware): hook before_model_call on step-limit escape LLM cal…

373f7d0

…ls in code.rs Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

feat(middleware): register ActivityLogFlushMiddleware from Manager::n…

116cf41

…ew; add MiddlewareConfig Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

feat(ticket-intake): implement fetch_full_context for GitHubProvider

7b7c380

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

feat(ticket-intake): implement fetch_full_context for LinearProvider

dcc4f72

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

docs: add rich context injection config fields to config.example.toml

8a9be97

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

feat(inject): add inject queue and active session tracking to Executor

efaaa5f

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

feat(inject): add CoreHandle::inject() and is_session_active()

697a63a

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

feat(inject): drain inject queue before each LLM call in strategy loops

947b2c8

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

feat(inject): route mid-run messages to inject queue in frontends

ac690fd

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

feat(inject): add max_queued_messages config; wire into Executor

2d30b49

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

feat(profiles): add ToolProfiles config type and GhostConfig.profile …

c932e85

…field Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

feat(profiles): resolve tool profile references during ghost loading

a35b2fd

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

feat(profiles): validate tool profile references and tool registratio…

a2b1b6d

…n in doctor Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

docs: add tool_profiles example to config.example.toml

ff5f99c

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

feat(todo): add TodoList state type

3d57ecd

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

feat(todo): add spark_todos migration and save helper to db.rs

6845a5b

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

feat(todo): add per-session TodoList to Executor with accessors

1a32074

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

feat(todo): implement Tool trait for TodoWriteTool/TodoCheckTool; reg…

e018aa4

…ister in ToolRegistry Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

fix: downscope executor fields to pub(crate); use floor_char_boundary…

f68746a

… in TicketContext Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Enreign changed the title ~~feat: Open SWE adoption — middleware, context injection, mid-run messaging, tool profiles, todo lists~~ feat: middleware hooks, context injection, mid-run messaging, tool profiles, todo lists Mar 20, 2026

github-advanced-security bot found potential problems Mar 20, 2026

View reviewed changes

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

feat: middleware hooks, context injection, mid-run messaging, tool profiles, todo lists#87

feat: middleware hooks, context injection, mid-run messaging, tool profiles, todo lists#87
Enreign wants to merge 29 commits intomainfrom
feature/open-swe-adoption

Enreign commented Mar 20, 2026 •

edited

Loading

Uh oh!

Check failure

Check failure

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Conversation

Enreign commented Mar 20, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Test Plan

Uh oh!

Check failure

Uh oh!

Check failure

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Enreign commented Mar 20, 2026 •

edited

Loading