Dump Context

📄 Extension source

Why. Captures the exact system prompt text sent to the LLM each turn and writes it to ignore/ on demand. Console shows stats: bytes, estimated vs actual token count, tools, skills, context files, and active injections (caveman, ponytail, lessons).

How it works. Hooks before_agent_start and stores event.systemPrompt (the fully assembled string after all extensions have patched it). The zzz- prefix ensures alphabetical load order places this last, so the capture includes modifications from caveman and ponytail. On /dump-context command, writes the stored prompt and JSON-structured options to ignore/, then prints summary stats to terminal.

Location: .pi/extensions/zzz-dump-context/

Why use it

The system prompt is assembled from many sources — pi’s default prompt, the global APPEND_SYSTEM.md append (cheasee-pi operating instructions), AGENTS.md, skills XML, plus injections from caveman and ponytail. There’s no built-in way to see the final string. This extension makes the full text inspectable and debuggable: verify what the model actually receives, audit token budget allocation, debug prompt injection conflicts.

Commands

Command	Effect
`/dump-context`	Writes `ignore/dump-context.txt` + `ignore/dump-context-options.json` and prints stats

Stats

The terminal output shows:

System prompt: 22,762 B  ·  context: 8,103 tok  ·  0.8%
~5,691 tok (est)  ·  API: 8,103 tok (+2,412)
11 tools (7 ext)  ·  14 skills  ·  1 ctx files  ·  27 guidelines  ·  13 sections
Injections: caveman, ponytail, lessons
→ ignore/dump-context.txt
→ ignore/dump-context-options.json

Stat	Source
`B`	`event.systemPrompt.length` (UTF-8 bytes)
`tok (est)`	`chars / 4` — rough heuristic (undercounts code/XML content)
`context` / `API`	`ctx.getContextUsage().tokens` — actual token count from provider
`tools (ext)`	`selectedTools` length, filtered by built-in set
`skills`	Counted from XML `<skill>` tags in prompt
`ctx files`	`contextFiles.length` — AGENTS.md files found
`guidelines`	`promptGuidelines.length`
`sections`	Markdown heading count (`##` / `###`)
Injections	Detected by substring match in prompt text

Section attribution

dump-context.txt labels each part of the assembled prompt with its source:

Section	Source
`[1] Base system prompt`	pi built-in (no custom `system.md`)
`[2] System prompt append (APPEND_SYSTEM.md)`	`~/.pi/agent/APPEND_SYSTEM.md` — the global cheasee-pi operating instructions, appended by pi in every repo (split seam: the file’s H1 anchor)
`[3] Project context`	`contextFiles` — AGENTS.md/CLAUDE.md files
`[4] Skills`	model-invocation-enabled skills
`[5] Working directory`	cwd line
`[6] Extension injections`	appended via `before_agent_start` (caveman, ponytail, lessons)

Details

Architecture

before_agent_start (fires every turn)
  → capture event.systemPrompt (final, after all extensions)
  → capture event.systemPromptOptions (structured metadata)
  → return undefined (don't modify prompt)
         ↓
/dump-context command (on demand)
  → write ignore/dump-context.txt
  → write ignore/dump-context-options.json
  → print stats to terminal
  → notify file paths

Why zzz- prefix

Extensions modify the system prompt in a chain — each before_agent_start handler receives the prompt as modified by all previous handlers:

caveman → ponytail → zzz-dump-context (captures final)
                                       ^
                                  runs last alphabetically

Without the zzz- prefix, the capture would miss later injections. The prefix ensures this handler sees the final string.

Files written

File	Content
`ignore/dump-context.txt`	The exact system prompt text sent to the LLM
`ignore/dump-context-options.json`	`BuildSystemPromptOptions` — structured data: tools, skills, guidelines, context files

Both files are overwritten each call. Saved to ignore/ per project conventions (temp files go in ignore/).

No session persistence

The extension does not use pi.appendEntry() — there’s no need to persist dump state across sessions. The prompt is captured fresh each turn.