fix(agent): auto-retry mid-stream + tighten system prompt

[code-crusher]

Summary

Two changes shipped on release/v0.3.4 and rolled up into one PR against main.

1. fix(agent): allow auto-retry mid-stream when partial output can be rolled back

Previously streamWithRetry only retried before the first chunk, because once any text or reasoning had streamed, re-issuing the request would duplicate on-screen output. The user-visible effect was that a dropped connection after partial progress surfaced as a failed step, even though the model was happy to continue.

This adds an optional onRestart callback to streamWithRetry. When the caller can cleanly undo the partial output (cleared buffers, reset accumulators) it returns true and the stream is re-issued.

• The main agent loop installs rollbackForRetry, which resets assistantText, reasoningOpen, reasoningStart, reasoningDetails, and pending tool calls, then emits a new stream-reset event. It declines the restart if a reasoning row was already committed to the transcript, since that cannot be undone.
• The compaction path installs a simpler reset that just clears its in-memory summary buffer (compaction only streams text and commits once at the end).
• The UI handler for stream-reset clears textBufferRef, streamingText, reasoningBufferRef, and streamingReasoning, and resets the busy label to Working so the spinner reflects the restarted attempt.

Files: src/core/events.ts, src/core/agent.ts, src/ui/App.tsx.

2. refactor(prompts): rewrite system prompt for speed and editing discipline

Replaces the "always gather exhaustive context" guidance with a "gather enough, then act" principle. The model is now told that a small, localized change typically needs about 3-6 tool calls and that further exploration after the edit point is identified is waste. The TODO list rule is scoped to multi-step tasks (3+ steps) instead of being mandatory for any work.

The file_edit / multi_file_edit section adds an explicit editing-discipline block: copy old_string verbatim from a same-turn read, treat earlier reads as stale after a successful edit, and never guess at a corrected old_string when a multi_file_edit batch fails.

The read_file and search_files sections collapse their repetitive parameter tables and examples into a short reference plus a "Reading Strategy" / "Search Hygiene" ruleset (read whole regions in one call, budget re-reads, verify the output matches the parameters sent, exclude test/spec/mock paths by default, scope path narrowly).

Two new cross-cutting sections are added: Verifying tool results and avoiding loops (check that outputs match the sent parameters; do not repeat an identical failing call) and Plan before editing (write the full change plan once, then execute edits in one batched pass with a single typecheck/build at the end).

Also fixes a "prefer to let the user to that" typo and a few list-formatting inconsistencies in the TODO list section.

Files: src/prompts/system.ts.

Test plan

• Start a long generation, drop the network mid-stream, confirm the agent auto-retries without duplicating output and finishes the step.
• Confirm a reasoning row that has already been committed is not silently lost on a mid-stream retry (it should surface as an error rather than a duplicated row).
• Run the compaction path against a large transcript and confirm a mid-stream drop still produces a single committed summary.
• Spot-check the system prompt on a 1-line edit (no todo list), a 5-step task (todo list), and a wide refactor (no exhaustive re-search after the edit point is identified).

[matterai-app]

Summary By MatterAI

🔄 What Changed

Implemented automatic retry logic for transient LLM streaming failures with exponential backoff and state rollback. Introduced a new /task slash command to reference and summarize previous conversations. Significantly refactored and tightened the system prompt to improve context gathering and editing discipline.

🔍 Impact of the Change

Improves agent resilience against network instability and API timeouts. Enhances user experience by allowing seamless context injection from past tasks. Reduces token usage and improves instruction following through a more concise system prompt.

📁 Total Files Changed

Click to Expand

File	ChangeLog
Version Bump package.json	Incremented version to 0.3.4.
Retry Logic src/core/agent.ts	Added streamWithRetry with backoff and state rollback safeguards.
Event Type src/core/events.ts	Added stream-reset event to handle UI buffer clearing during retries.
Prompt Refactor src/prompts/system.ts	Tightened instructions, emphasizing context gathering and verbatim editing.
Task Command src/ui/App.tsx	Implemented /task command and logic to summarize previous sessions.
UI Component src/ui/components/SessionPicker.tsx	Added customizable title prop to the session selection component.

🧪 Test Added/Recommended

Recommended

• Unit tests for isRetryableStreamError to verify status code filtering.
• Integration tests for streamWithRetry to ensure state rollbacks (text/reasoning) function correctly.
• Mock tests for the /task command to verify conversation truncation logic.

🔒 Security Vulnerabilities

No critical security vulnerabilities detected. Input validation for the /task command is handled via internal session listing.