SP Render Inspector (scripts/dogfood_sp_render.py)

A CLI for verifying system-prompt output during skill development — replacing the one-off python -c "..." scripts that were previously the only way to preview what the OS actually injects into a LLM call.

Why

Before this tool, checking a system prompt required writing an ad-hoc Python snippet each time: import the SP builder, construct a fake context, call render, print. The snippet varied by what you wanted to inspect (size? sections? legacy-literal audit?), and it was never saved. Debugging a routing problem could cost five variants of the same throwaway script in a single session.

dogfood_sp_render.py canonicalises that workflow into six named modes. You pass the flags that describe your agent's configuration, pick a mode, and get a consistent, repeatable output.

Setup

No installation beyond the project's standard dependencies. The script lives in scripts/ and uses only modules already in the virtualenv:

# From the project root
python scripts/dogfood_sp_render.py [flags] [mode]

Usage

Default mode — full SP preview

python scripts/dogfood_sp_render.py \
  --agent-name my_agent \
  --agent-role "Coding assistant" \
  --skill "skill_builder=Builds and improves skills" \
  --skill "eval=Runs evaluation scenarios"

Prints the rendered system prompt to stdout. Pipe to less for long prompts.

--stats — character and line counts

python scripts/dogfood_sp_render.py \
  --agent-name my_agent \
  --skill "skill_builder=Builds and improves skills" \
  --stats

Output:

2635 chars / 47 lines

Use this to monitor SP size during development. A sudden size increase after a commit indicates an unexpected injection.

--show-sections — SP structural overview

python scripts/dogfood_sp_render.py \
  --agent-name my_agent \
  --skill "skill_builder=Builds and improves skills" \
  --show-sections

Output:

## Capabilities (routing guide)
## Action categories
## Behaviour

Use this to confirm expected sections are present without reading the full SP.

--grep-legacy — legacy literal audit

python scripts/dogfood_sp_render.py \
  --agent-name my_agent \
  --skill "skill_builder=Builds and improves skills" \
  --grep-legacy

Exits with code 0 if no legacy literals are found. Exits with code 1 and prints the offending lines if any are found. Designed for pre-commit hooks and CI checks — see Integration with workflow.

--compare-legacy — size delta between wrapper and legacy SP

python scripts/dogfood_sp_render.py \
  --agent-name my_agent \
  --skill "skill_builder=Builds and improves skills" \
  --compare-legacy

Output:

Legacy: 9,029 chars / Wrapper: 2,635 chars / Reduction: 70.8%

Use during dogfood prep to confirm the wrapper SP is significantly smaller than the legacy SP. A regression (wrapper larger than legacy, or reduction below 50%) is a signal to investigate what is being injected unexpectedly.

--legacy-check — boolean pass/fail for legacy mode

python scripts/dogfood_sp_render.py \
  --agent-name my_agent \
  --skill "skill_builder=Builds and improves skills" \
  --legacy-check

Prints PASS or FAIL to stdout and exits with the corresponding code. A lighter alternative to --grep-legacy when only pass/fail status is needed.

Flag reference

Agent identity flags

Flag	Description
--agent-name NAME	Name of the agent (used in SP preamble)
--agent-role TEXT	Role description injected into the SP

Capability flags (repeatable)

Flag	Description	Format
--skill name=desc	Skill available to this agent	name=description
--agent-peer name=role	Peer agent this agent can delegate to	name=role
--mcp-servers name=desc	MCP server this agent has access to	name=description
--indexed-sources name	Indexed source name (for RAG-aware agents)	plain name

Scope and context flags

Flag	Description
--file-scope read=path write=path	File scope for permission-aware rendering
--output-language LANG	Output language code (e.g. ja, en)
--project-context TEXT	Project context injected into the SP

Rendering mode flags

Flag	Description
--hide-legacy-tools	Render with legacy tool definitions hidden
--universal-wrappers-enabled	Enable universal wrapper mode

Output mode flags (mutually exclusive)

Flag	Description
--stats	Print char/line count only
--show-sections	Print section headers only
--grep-legacy	Audit for legacy literals; exit 1 if found
--compare-legacy	Print wrapper vs legacy size comparison
--legacy-check	Print PASS/FAIL and exit accordingly

Integration with workflow

Before commit — leak check

Run --grep-legacy to confirm that wrapper-only mode does not contain any legacy literals that would break routing:

python scripts/dogfood_sp_render.py \
  --agent-name my_agent \
  --universal-wrappers-enabled \
  --skill "skill_builder=Builds and improves skills" \
  --grep-legacy

Exit code 1 means a literal leaked through; inspect the printed lines to identify the source.

During dogfood prep — size delta observation

Check the reduction ratio before each dogfood batch to establish a baseline. A ratio below 50% after a code change is worth investigating before running expensive LLM sessions:

python scripts/dogfood_sp_render.py \
  --agent-name my_agent \
  --skill "skill_builder=Builds and improves skills" \
  --compare-legacy

When debugging routing problems

Use --show-sections to confirm all expected sections are present, then switch to the default mode to read the full SP for the specific section where the LLM is misbehaving:

# 1. Confirm sections
python scripts/dogfood_sp_render.py --agent-name my_agent --skill "..." --show-sections

# 2. Read full SP
python scripts/dogfood_sp_render.py --agent-name my_agent --skill "..." | less

When NOT to use this tool

• Inspecting what a live session actually sent to the LLM. This tool renders the SP from flags you provide; it does not read a running session's state. For that, use REYN_LLM_TRACE_DUMP + dogfood_trace.py.
• Auditing LLM behavior. SP rendering is input-side only. For output-side behavior, use llm_replay.py.

See also

• LLM Payload Tracing — capturing and inspecting live LLM payloads (REYN_LLM_TRACE_DUMP, dogfood_trace.py, llm_replay.py)