use-cept

name	use-cept
description	Use when the user asks to use Cept or cept, asks for outside-in steering, wants an agent session reviewed, is stuck or looping, needs debugging triage, or wants architecture/research guidance from the recent agent trajectory. This skill tells Codex to call the local Cept CLI or MCP tool first and treat Cept output as evidence to reconcile.

Use Cept

Cept is proprioception for coding agents: a way for the AI to inspect its own recent work, uncertainty, and next move before continuing. Use it when the agent needs a deliberate outside-in review of its recent trajectory.

Cept is adapter-based at the transcript boundary. Claude Code is the default adapter, but any agent can use cept by writing agent-neutral JSONL and passing --transcript.

Cept's grounded default is perplexity/sonar-pro through OpenRouter. Use OPENROUTER_API_KEY; this plugin does not require or read PERPLEXITY_API_KEY. For keyfile setup and troubleshooting, read docs/CEPTKEY.md in the cept repo.

The source-owned local plugin path follows the Conduit pattern: /Users/dshanklinbv/plugins/cept -> /Users/dshanklinbv/repos-eidos-agi/cept.

Primary Rule

Run the smallest relevant cept command first, then reconcile the result with repo evidence before acting.

Useful entrypoints:

cept-cli --help
cept-keyfile show
cept-keyfile where
/Users/dshanklinbv/plugins/cept/scripts/cept-registry doctor
/Users/dshanklinbv/plugins/cept/scripts/cept-registry registry --json
cept-cli --goal "<goal>" --headline "<3-4 word headline>" --dry-run
cept-cli --provider openrouter --goal "<goal>" --headline "<3-4 word headline>"
cept-cli --goal "<goal>" --headline "<3-4 word headline>" --mode debug
cept-cli --goal "<goal>" --headline "<3-4 word headline>" --mode research
cept-cli --goal "<goal>" --headline "<3-4 word headline>" --mode architecture
cept-cli --goal "<goal>" --headline "<3-4 word headline>" --transcript /tmp/agent.jsonl --dry-run
cept-cli --self-assess --transcript /tmp/agent.jsonl --dry-run
cept-hud-install

If the MCP tool is available, call the cept tool with a fresh session-verification nonce.

When To Use Cept

Use Cept when:

• The agent is stuck, looping, or retrying the same fix.
• A bug needs ranked likely causes from the recent transcript.
• A design choice needs outside architecture review.
• Current work depends on external facts or version gotchas.
• The user explicitly says "use Cept" or "use cept."

Modes

• steer: broad blind spots and next-step guidance.
• debug: rank likely causes from errors and evidence.
• research: check external facts, docs, and version-specific gotchas.
• architecture: compare design alternatives and tradeoffs.

Secret Boundary

Cept redacts before it sends, but it still packages recent transcript and optional file content. Use --dry-run first when the packet may contain sensitive material. Do not include secrets, private keys, raw tokens, protected personal data, or unrelated files.

Source-Of-Truth Rules

• Treat Cept output as evidence, not a verdict.
• Verify any concrete claim against local repo state or primary docs before shipping.
• Keep the headline short. If you cannot summarize the ask in 3-4 words, clarify the goal first.
• Prefer --file only for files that are directly relevant to the requested critique.
• For non-Claude agents, normalize transcript rows into user/assistant messages, tool calls, and tool results before invoking cept.

Plugin Boundary

This plugin is a thin signpost to the cept CLI/MCP server. The CLI owns transcript discovery, redaction, OpenRouter calls, progress events, and HUD integration. The source-owned registry under registry/ documents Cept storage locations and stores local proof records.