phase-10h-provider-thinking-research — Claude Code prompt
phase-10h-provider-thinking-research — a Claude slash command that digs through Murph, Pi and Phantom sources to find what providers really expose about thinking and reasoning events.
Apache-2.0 ★ 1,497 updated 2026-06-16
.claude/commands/phase-10h-provider-thinking-research.mdWhat phase-10h-provider-thinking-research does
The file is a research brief for Claude. It asks the model to work through the source of Murph, Pi and Phantom and pin down what each one actually carries for reasoning events, redacted thinking, progress updates, token usage and rate limits. Several ultrathink lines at the top put the session into extended thinking before any reading starts.
A fixed reading order follows: AGENTS.md, QUALITY-BAR.md, PROGRESS.md, the Murph core directories, the Phantom chat files such as sdk-to-wire.ts and thinking-block.tsx, and the provider types in pi-mono. Outside material is limited to official docs. The paths are absolute and point at the author's home directory, so on your machine the pass stops at step one until you rewrite them.
Everything lands in one report at research/chat-experience/phase-10h-provider-thinking-research.md: the event taxonomy, a capability matrix for OpenAI, Anthropic and ZAI, and the line between verbatim output and private chain-of-thought. It also has to say where thinking-token counts come from and which tests to run first. No code edits, no key values printed, no costly live model runs.
What it helps with
- Map the thinking events that Murph and Phantom already emit
- Compare reasoning support across OpenAI, Anthropic and ZAI in a single matrix
- Separate what can be shown verbatim from private chain-of-thought
- Name the missing event behind the absent thinking-token counter
- Collect the tests needed before shipping the thinking panel
Example prompts for the agent
- Run
/phase-10h-provider-thinking-researchover the Murph and Phantom sources and tell me what we can honestly label as thinking /phase-10h-provider-thinking-researchfor the three providers — build the Anthropic, Anthropic and ZAI matrix- Call
/phase-10h-provider-thinking-researchand list the event gaps plus the tests before release
How to install phase-10h-provider-thinking-research
Manually: put the files in place
.claude/commands/phase-10h-provider-thinking-research.mdHow to use
Paste the text at the start of a chat or into project instructions
Questions about phase-10h-provider-thinking-research
How do I install phase-10h-provider-thinking-research in Claude Code?
Place the file at .claude/commands/phase-10h-provider-thinking-research.md inside your project, then run /phase-10h-provider-thinking-research.
What does phase-10h-provider-thinking-research do?
It walks the Murph, Pi and Phantom sources and writes a report on which thinking and reasoning events can be surfaced without misleading the user.
Does the command touch my code?
No. It stays a research pass: no edits, no API key values printed, no expensive live model runs.
Source text
phase-10h-provider-thinking-research.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 provider thinking and reasoning event support across Murph, Pi, Phantom, OpenAI, Anthropic, and ZAI. Mission: Determine what Phantom can truthfully show as "thinking" or "reasoning" today without lying to the user or exposing unsafe private chain-of-thought. We need best-in-class progress, but it must respect provider semantics. Required reading, in order: 1. `/Users/truffle/.claude/AGENTS.md` 2. `/Users/truffle/work/murph/QUALITY-BAR.md` 3. `/Users/truffle/work/murph/PROGRESS.md` 4. Murph source: - `/Users/truffle/work/murph/packages/core/src/events/` - `/Users/truffle/work/murph/packages/core/src/substrate/` - `/Users/truffle/work/murph/packages/core/src/providers/` - `/Users/truffle/work/murph/packages/anthropic-sdk-shim/src/` 5. Phantom source: - `src/chat/sdk-to-wire.ts` - `src/chat/sdk-to-wire-handlers.ts` - `src/chat/types.ts` - `src/chat/run-timeline.ts` - `chat-ui/src/components/thinking-block.tsx` - `chat-ui/src/components/message.tsx` - `chat-ui/src/lib/chat-types.ts` 6. Pi source: - `/Users/truffle/work/pi-mono/packages/ai/src/types.ts` - `/Users/truffle/work/pi-mono/packages/ai/src/providers/openai-responses.ts` - `/Users/truffle/work/pi-mono/packages/ai/src/providers/anthropic.ts` - `/Users/truffle/work/pi-mono/packages/ai/src/providers/openai-completions.ts` - `/Users/truffle/work/pi-mono/packages/ai/src/providers/transform-messages.ts` External docs: - Use official provider docs only if needed. Prefer OpenAI and Anthropic official docs, ZAI official docs or source-backed docs. - Do not use SEO sites or tutorials. Deliverable: Write `/Users/truffle/work/phantom-murph-hardening/research/chat-experience/phase-10h-provider-thinking-research.md`. The report must include: 1. Event taxonomy: what events we currently have for thinking, redacted thinking, progress summaries, token usage, tool calls, and rate limits. 2. Provider capability matrix for OpenAI, Anthropic, and ZAI as routed through Pi/Murph. 3. What can be shown verbatim, what should be summarized, and what must be hidden or labeled as private/redacted. 4. Whether Phantom can show thinking tokens/counts today, and if so where the data comes from. If not, name the missing Murph/Pi event or field. 5. Recommended UI model for "thinking" that is honest across providers. 6. Tests needed before shipping thinking/progress UI changes. Non-goals: - Do not edit code. - Do not print or inspect API key values. - Do not run costly live model loops unless absolutely needed. This is a source and event-contract research pass. Self-review: Before finishing, check that each recommendation maps to an existing event or a clearly named event-contract gap.
© ghostwright · Apache-2.0 · ghostwright/phantom