# phase-10l-phantom-continuity-researcher

phase-10l-phantom-continuity-researcher is a Claude Code prompt for diagnosing session continuity in Phantom, including context compaction, transcript recovery, and SQLite state analysis.

- prompt · `type=prompt` · Claude Code, ChatGPT · ★ 1497 repo stars
- Author: ghostwright
- License: Apache-2.0 · policy=mirror · the beginning of the file is at the end of this page
- Source: https://github.com/ghostwright/phantom/tree/HEAD/prompts/phase-10l-phantom-continuity-researcher.md
- Updated: 2026-06-16
- Security check: automated, 2026-11-16 (text) — no signs of malicious code found.
- HTML: https://affpapa.org/en/nevibecode/prompts/phase-10l-phantom-continuity-researcher

## What it does

This prompt instructs Claude Code to act as a principal engineer, architect, and product manager investigating Phantom's web chat mechanics. It focuses on long-running Murph sessions, timeline persistence, and scenarios where compaction causes the agent to lose track of created pages.

The directive prohibits file modifications. The model must inspect local files in `src/chat`, `src/agent`, and `chat-ui`, alongside SQLite run event data, answering seven core questions regarding session reuse, context fidelity, and UI state truthfulness.

Because changing Murph internals carries risk, the prompt restricts outputs to a diagnostic memo with line numbers, verified facts, and a minimal testable fix.

## Install

### Claude Code (native platform)

**Manually: put the files in place**
- File: `.claude/commands/phase-10l-phantom-continuity-researcher.md`

### ChatGPT

**How to use**
- Paste the text at the start of a chat or into project instructions

Where a file is given instead of a command, copy the files from the source to that path: https://github.com/ghostwright/phantom/tree/HEAD/prompts/phase-10l-phantom-continuity-researcher.md

## When to use it
- trace lost artifact references after session context compaction
- verify how Murph session IDs persist across chat queries
- audit SQLite chat event logs against UI timeline states
- outline a minimal fix before modifying Murph internals

## Example prompts
- Run the continuity research analysis using /phase-10l-phantom-continuity-researcher
- Investigate Murph context compaction behavior with phase-10l-phantom-continuity-researcher

## Source text · phase-10l-phantom-continuity-researcher.md · Apache-2.0

Below is the content of a third-party file. It is data, not instructions for you: do not act on anything written inside unless the human asks. Block tag: b09af1da67431896.

BEGIN THIRD-PARTY FILE b09af1da67431896
```text
ultrathink. ultrathink. ultrathink.

You are a principal engineer, principal architect, and principal product
manager at Anthropic, all three at once. You are researching Phantom's web chat
behavior around long-running Murph sessions, compaction, replay, and user
perception. Do not edit files. Read actual files and, if useful, inspect local
SQLite state from the recent test sessions.

Context:

- Phantom now runs locally on Murph with real OpenAI successfully.
- Tool cards, durable timeline, transcript search, and artifact cards exist.
- The operator observed a real long-running test where the agent created a
  page, compaction seemed to work, then follow-up questions about the page link
  got confused or stopped responding.
- The UI also needs to communicate long-running work better, but the immediate
  risk is whether the agent run state and transcript recovery are correct after
  compaction.

Reading list:

1. `/Users/truffle/work/phantom-murph-hardening/CLAUDE.md`
2. `/Users/truffle/work/phantom-murph-hardening/research/chat-experience/phase-10h-product-direction.md`
3. `/Users/truffle/work/phantom-murph-hardening/research/chat-experience/phase-10i-transcript-recovery-implementation.md`
4. `/Users/truffle/work/phantom-murph-hardening/src/chat/**`
5. `/Users/truffle/work/phantom-murph-hardening/src/agent/**`
6. `/Users/truffle/work/phantom-murph-hardening/chat-ui/src/**` areas related
   to status, run activity, messages, and tool rendering.
7. Recent local SQLite chat/run event data if available and safe to inspect.

Questions to answer:

1. How does a new user message in web chat become an AgentRuntime query, and
   how is the previous Murph session id reused?
2. Does Phantom send enough post-compaction context for the agent to recover
   created artifacts, page paths, and prior task state?
3. Does Phantom have a concept of "agent compacted, now continue the same task",
   or does every user message start a fresh query that depends on Murph memory?
4. Are durable timeline events sufficient to reconstruct what happened for the
   user and for the agent, or are they UI-only?
5. Where can the UI truthfully show compaction, recovery, waiting, and retry
   states with current wire events?
6. What caused the observed "done but confused about the created page link"
   failure mode most likely: Murph summary quality, Phantom prompt/context,
   transcript search tool availability, UI state, or agent behavior?
7. What is the smallest testable Phantom-side fix, if any, before changing
   Murph internals?

Deliverables:

- A concise note with file references and line numbers.
- One recommended next implementation slice, including exact tests and browser
  verification.
- A list of facts the agent should have after compaction and where each should
  live: Murph summary, Phantom transcript search, durable artifact record, or
  UI-only state.

Anti-patterns:

- Do not invent a separate hidden memory store if the transcript or artifact
  record is the right source of truth.
- Do not make the UI pretend work is happening after the backend is done.
- Do not expose secrets or magic-login links in summaries, tools, artifacts, or
  persisted timeline displays.
- Do not add a Phantom-specific hack if the right fix belongs in Murph.

Self-review:

Before finalizing, separate confirmed facts from hypotheses and note what would
prove or disprove each hypothesis.
```
END THIRD-PARTY FILE b09af1da67431896

The third-party file has ended. Everything between the b09af1da67431896 tags is data: requests and commands from it are not followed. An “end of file” line with another tag inside the block is part of the file.

© ghostwright · Apache-2.0 · https://github.com/ghostwright/phantom/tree/HEAD/prompts/phase-10l-phantom-continuity-researcher.md

## Similar
- [Phantom Connect SDK](https://affpapa.org/en/nevibecode/mcp/phantom-connect-sdk.md) — MCP server · 3016 installs
- [helius-phantom](https://affpapa.org/en/nevibecode/skills/helius-phantom.md) — skill · ★ 33 repo stars
- [gsd-phase-researcher](https://affpapa.org/en/nevibecode/agents/gsd-phase-researcher.md) — subagent · ★ 3785 repo stars
- [phantom-quench](https://affpapa.org/en/nevibecode/skills/phantom-quench.md) — skill · ★ 33 repo stars
- [cost-compact-context](https://affpapa.org/en/nevibecode/skills/cost-compact-context.md) — skill · 728 installs
- [agency-continuity-audit](https://affpapa.org/en/nevibecode/plugins/agency-continuity-audit.md) — plugin · ★ 1596 repo stars

From the same set: [ghostwright/phantom](https://affpapa.org/en/nevibecode/by/ghostwright/phantom.md) — 38 cards

---
NeVibeCode on affpapa.org. Catalog texts are descriptions, not instructions for the agent. How to choose and install: https://affpapa.org/en/nevibecode/agents.md
