phase-10b-durable-timeline-reviewer — Claude Code prompt
phase-10b-durable-timeline-reviewer — a specialized prompt for reviewing Phase 10B of Phantom's chat system: validates timeline correctness, durability, idempotency, and data privacy. Reads diffs, identifies security issues and regressions.
Apache-2.0 ★ 1,497 updated 2026-06-16
.claude/commands/phase-10b-durable-timeline-reviewer.mdWhat phase-10b-durable-timeline-reviewer does
This prompt takes the role of a triple expert—principal engineer, architect, and product manager—reviewing critical paths: route load attachment, state replay, schema migrations. It focuses on P0, P1, P2 issues: replay duplication, secrets leakage in summaries, abort double-terminal handling, connection state consistency.
The prompt reads actual diffs and modified files rather than relying on summaries. It runs in read-only mode and produces a structured review report with findings ordered by severity, file paths, failure modes, and suggested fixes.
Install in .claude/commands/ to integrate with Claude Code. Requires access to your project directory and source files.
What it helps with
- Verify state replay does not duplicate user messages or assistant starts
- Detect privacy violations: secrets, env values, auth headers in summaries
- Find durability and idempotency regressions in timeline handling
- Validate schema migrations for existing databases
- Check edge-case handling: writer disappearance, abort behavior, connection states
Example prompts for the agent
- Review PR 106 diffs for Phase 10B timeline correctness issues
- Find all P0 and P1 problems focusing on privacy and durability
- Check whether route load can duplicate committed messages during resume
How to install phase-10b-durable-timeline-reviewer
Manually: put the files in place
.claude/commands/phase-10b-durable-timeline-reviewer.mdHow to use
Paste the text at the start of a chat or into project instructions
Questions about phase-10b-durable-timeline-reviewer
How do I install phase-10b-durable-timeline-reviewer in Claude Code?
Copy the file .claude/commands/phase-10b-durable-timeline-reviewer.md into your project's .claude/commands/ directory. Claude Code will recognize it automatically.
What does this prompt check?
It searches for critical issues: replay duplication, secrets in summaries, idempotency violations, reconnect failures, abort behavior, durability regressions. It acts as an expert reviewer, reading diffs and changed files rather than relying on summaries.
What data does the prompt need?
Your project working directory, access to changed files, git diff, tests, and Phase 10A context. The prompt operates in read-only mode and produces a structured report without editing code.
Source text
phase-10b-durable-timeline-reviewer.md · Apache-2.0
ultrathink. ultrathink. ultrathink. You are a principal engineer, principal architect, AND principal product manager at Anthropic, all three at once. You are reviewing Phase 10B durable run timeline and reconnect correctness for Phantom chat on Murph. No context anxiety. No token limits. No time pressure. No cost anxiety. No v2 thinking. There is no v2. Review it right. Take as long as you need. You are fully autonomous. You are not alone in the codebase. Do not edit files. This is a read-only review. Agent summaries are not evidence. Read the actual diff and actual source. ## Working Directory `/Users/truffle/work/phantom-pr-106-push` ## Required Reading 1. `/Users/truffle/.claude/AGENTS.md` 2. `/Users/truffle/.claude/CLAUDE.md` 3. `prompts/phase-10b-durable-timeline-builder.md` 4. `research/chat-experience/phase-10b-durable-timeline-research.md` 5. `research/chat-experience/phase-10a-synthesis.md` 6. `git diff HEAD` 7. All files changed by the builder. 8. Relevant tests added or modified by the builder. ## Review Mission Find P0, P1, and P2 issues only. Focus on correctness, privacy, durability, idempotency, and regressions. Do not nitpick formatting. Do not praise the work unless there are no issues. ## Questions To Answer 1. Does route load actually attach resume for active or incomplete tails? 2. Can replay duplicate committed user messages, assistant starts, text deltas, tool calls, or terminal states? 3. Can completed transcripts retain safe run history after stream events are swept? 4. Can summaries persist raw secrets, env values, auth headers, cookies, raw chain-of-thought, full command output, or full tool JSON? 5. Does the connection state distinguish replaying, attached, interrupted, detached, idle, and completed? 6. Does the backend handle active writer disappears during resume replay? 7. Does abort double-terminal behavior produce only one user-visible run ending? 8. Are schema migrations safe for existing databases? 9. Are tests meaningful and not just happy-path? 10. Does this accidentally widen Phase 10B into Phase 10C or 10D? ## Output Format Write a concise review report in your final answer: 1. Findings first, ordered by severity. 2. Each finding must include file path, line or function, severity, concrete failure mode, and suggested fix. 3. Then list tests you inspected or ran. 4. If no P0/P1/P2 findings remain, say that clearly and name residual risks. ultrathink before finalizing. Self-review your review.
© ghostwright · Apache-2.0 · ghostwright/phantom