phase-10l-murph-continuity-researcher — Claude Code prompt
phase-10l-murph-continuity-researcher investigates how Murph handles context compaction and interrupted turns, then produces a code-backed fix and test plan. It expects local Murph and Phantom repositories.
Apache-2.0 ★ 1,497 updated 2026-06-16
.claude/commands/phase-10l-murph-continuity-researcher.mdWhat phase-10l-murph-continuity-researcher does
This prompt directs Claude Code to inspect Murph’s context handling, compaction, provider adapters, and tool-call history, alongside Phantom transcript-recovery notes. The research should establish what the next provider request contains after compaction and whether Murph can resume an interrupted turn. It explicitly forbids editing files.
The requested output is an architecture note with file and line references, a verdict on where the fix belongs, a ranked implementation plan with acceptance criteria, and risks. It asks for provider-specific checks involving OpenAI, Anthropic, and ZAI, plus deterministic state-machine tests and real-provider tests where practical.
Findings must come from code that was actually inspected; summaries must not retain secrets or full tool payloads. The prompt targets working copies under /Users/truffle/work, so source-level verification depends on access to those repositories.
What it helps with
- trace what Murph sends after context compaction
- check whether Murph reuses Pi overflow and summarization primitives
- investigate why an agent loses its answer after compaction
- rank fixes and define acceptance criteria
- plan tests for interrupted turns and tool-call history
Example prompts for the agent
- Inspect how Murph resumes an interrupted turn after compaction and write an architecture note with code references.
- Compare Murph’s context-overflow handling with Pi and recommend a production fix with acceptance criteria.
- Review OpenAI tool-call ordering risks and propose tests for the continuation flow.
How to install phase-10l-murph-continuity-researcher
Manually: put the files in place
.claude/commands/phase-10l-murph-continuity-researcher.mdHow to use
Paste the text at the start of a chat or into project instructions
Questions about phase-10l-murph-continuity-researcher
How do I install phase-10l-murph-continuity-researcher in Claude Code?
Save the prompt as .claude/commands/phase-10l-murph-continuity-researcher.md.
What does phase-10l-murph-continuity-researcher do?
It asks Claude Code to investigate context compaction and turn continuation in Murph and Phantom. The deliverables include a code-referenced architecture note, an implementation plan, and tests.
Which files does the research cover?
It directs Claude Code to inspect Murph materials and Phantom transcript-recovery documents under /Users/truffle/work. It also suggests checking local pi-mono and pi-code copies.
Source text
phase-10l-murph-continuity-researcher.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 researching Phantom on Murph post-compaction continuity. Do not edit files. Do not summarize from memory. Read the actual Murph and Phantom code. Context: - Phantom is being moved onto Murph, a clean-room pi-mono based agent runtime. - We already implemented Phantom transcript recovery through `phantom_chat_transcript_search`. - Real browser testing showed long-running tool-heavy tasks work, and Murph compaction appears to fire, but after compaction the agent may lose the created-page answer or stop responding usefully. - The operator believes compaction should be an LLM-written compact summary, not literal history replay. The next model call should receive that summary plus enough pointers and tools to continue. - We must reuse pi-mono/Pi capabilities where they exist and avoid rebuilding battle-tested primitives in worse form. Reading list: 1. `/Users/truffle/work/murph/AGENTS.md` 2. `/Users/truffle/work/murph/QUALITY-BAR.md` 3. `/Users/truffle/work/murph/PROGRESS.md` 4. `/Users/truffle/work/murph/src/**` areas related to context overflow, compaction, provider adapters, message history, tool call preservation, and summaries. 5. `/Users/truffle/work/phantom-murph-hardening/research/chat-experience/phase-10g-pi-continuity.md` 6. `/Users/truffle/work/phantom-murph-hardening/research/chat-experience/phase-10i-transcript-recovery-implementation.md` 7. Any local pi-mono, pi-code, or notes repo already cloned under `/Users/truffle/work` that contains relevant compaction or overflow logic. Questions to answer: 1. What exactly does Murph do today when context approaches or exceeds the provider limit? 2. Does Murph use pi-mono/Pi primitives for context overflow detection, summarization, and retry, or did we duplicate that logic? 3. After compaction, what is sent to the next provider call? Include the shape of summary, preserved tool-call state, and any continuation instructions. 4. Is there an existing Murph seam for "continue the interrupted turn after compaction", or is compaction only a pre-call trimming step? 5. Which provider-specific constraints matter for OpenAI, Anthropic, and ZAI around tool-call history after compaction? 6. What is the smallest production-grade fix that improves continuation without corrupting tool-call protocols? 7. What tests should prove this, preferably with real providers when possible and deterministic unit tests for the core state machine? Deliverables: - A concise architecture note with file references and line numbers. - A clear verdict: reuse Pi/Murph existing primitives, fix Murph primitive, or add Phantom-specific continuation glue. - A ranked implementation plan with acceptance criteria. - Explicit risks and anti-patterns, especially anything that would break OpenAI tool-call message ordering. Anti-patterns: - Do not suggest raw transcript stuffing. - Do not suggest regex-heavy summarization. - Do not expose private chain-of-thought. - Do not preserve raw secrets, magic links, auth headers, or full tool payloads in summaries. - Do not treat a single reference repo as authoritative. Self-review: Before finalizing, re-read your own findings and remove any claim not backed by files you actually inspected.
© ghostwright · Apache-2.0 · ghostwright/phantom