CtxWise

The local-first observability and control layer for OpenAI Codex.
Audit context, compile intentional profiles, catch configuration drift, and attach honest usage receipts.

Formerly CtxRay. Public name, CLI, and package are now CtxWise, ctxwise, and @framy2/ctxwise. This is not the unrelated PyPI project named ctxray. The ctxray command remains a compatibility alias.

Install · 20-second demo · Benchmark evidence · Reproduce it

CtxWise is a local-first CLI and Codex plugin for context diagnostics, safe profile compilation, drift detection, reproducibility lockfiles, and honest post-turn usage receipts. It calls no model of its own, requires no API key, and has no telemetry.

Community project. Not affiliated with or endorsed by OpenAI.

See what Codex carries in one command

npx @framy2/ctxwise snapshot

snapshot runs locally and makes no model or API call. It shows the largest known startup contributors, labels configuration and MCP surfaces it cannot measure, and suggests the next safe action. It never reads session transcripts or prints configuration values.

CtxWise snapshot · attention · ~3,738 known startup tokens (estimated)
2 guidance files · 41 skill descriptions · 10 plugins · 1 MCP server

Largest known contributors:
- ~299 · guidance · codex-home/AGENTS.md
- ~170 · guidance · project/AGENTS.md
- ~145 · skill discovery · imagegen

7 config/profile sources and 1 MCP tool-schema surface are unmeasured;
unknown is not treated as zero.

See it in 20 seconds

The demo uses the public benchmark and the shipped CLI behavior. Click it for the compact MP4, or use the square social cut.

The papercuts it removes

You should not have to...	CtxWise gives you...
Guess which skills, instructions, plugins, or MCPs load	A private audit and bounded context map
Hand-edit model profiles and hope the change is reversible	Reviewable YAML, native Codex TOML, dry-runs, and automatic backups
Discover context changes only after a worse or costlier run	A redacted lockfile plus a CI-ready drift check
Confuse token estimates, subscription quota, and API bills	Receipts that label exact, estimated, comparison, and unknown data

Live evidence

The first public microbenchmark ran 10 paired tasks across requested GPT-5.6 Luna, Terra, and Sol profiles: 20/20 answers passed the same exact validator. Removing irrelevant discovered context reduced the estimated model-visible prompt by 73.3% and exact aggregate turn tokens by 29.1%. Every pair holds task, requested model, effort, sandbox, commit, and quality gate constant.

These are bounded microbenchmark results, not a universal productivity claim. The runtime stream records the requested profile but does not independently attest the served model, and prompt size remains a character-based estimate. Read the method and limitations, inspect the machine-readable summary, or watch the 20-second MP4.

A second maintainer run measured a conservative 28.5% exact aggregate reduction across nine comparable pairs and exposed a stale package-name validator. The raw repeat run and erratum are public; it is evidence of repeatability and fixture hardening, not an independent community reproduction.

The first external code review found that a nested invocation could omit project-root and intermediate AGENTS.md guidance. Version 0.2.2 fixes that gap, adds root-marker/fallback/byte-budget compatibility tests, and narrows lockfiles to the active root-to-working-directory guidance chain. This is an actionable independent finding, not an independent benchmark reproduction.

The first external code contribution added targeted benchmark task selection. Version 0.2.4 completes it with fail-closed ledger scoping, consistent report denominators, safe continuation commands, and a named two-turn smoke test. The contributor commit and maintainer RED/GREEN follow-ups remain visible in PR #7.

Run the smallest useful reproduction

Clone the repository and run one explicit baseline/optimized pair:

git clone https://github.com/FramY2/ctxwise.git
cd ctxwise
git checkout v0.4.0
npm ci
npm run benchmark:quick

The command states the quota-consuming turn limit before execution, creates a fresh community-* ledger, and writes share.md, summary.json, report.md, and SHA256SUMS.txt. It never uploads them. To choose another fixed task, run npm run benchmark:reproduce -- --task semantic-version. Unknown tasks and mixed-scope ledgers fail before Codex starts. Share successes or failures in Issue #1; independence is verified by the maintainer rather than asserted by the tool.

Why CtxWise exists

Codex already exposes excellent runtime primitives such as /status, /usage, /statusline, codex exec --json, profile files, and the app-server. The hard part is connecting them into one answer:

• Which instructions, skills, plugins, agents, and MCP declarations are active?
• Is a large context intentional or accidental?
• Which model/subagent profile should this task use?
• Can another developer reproduce the same capability surface safely?
• Was a dollar amount actually billed, or is it merely an API comparison?

CtxWise does that glue work without becoming another chat wrapper.

Features

Command	What it does	Network/model call
ctxwise snapshot	Ranks known context contributors, coverage gaps, and safe next actions	None
ctxwise audit	Inventories config plus active root-to-CWD guidance, skills, plugins, agents, and MCPs	None
ctxwise map	Renders a bounded Mermaid map of context sources and discovery overhead	None
ctxwise xray	Summarizes model-visible prompt JSON without echoing its text	None
ctxwise profile	Compiles YAML into native ~/.codex/<name>.config.toml, with dry-run and backups	None
ctxwise lock	Hashes a redacted capability surface for reproducibility	None
ctxwise drift	Compares a capability lock with a file or live setup; can fail CI on drift	None
ctxwise quota	Reads the current plan and quota window through local Codex app-server	Codex account read only
ctxwise receipt	Calculates a receipt from saved codex exec --json usage	None
ctxwise run	Runs Codex and appends exact usage plus an optional pre-turn prompt X-Ray	The requested Codex turn

Cost honesty by design

The dollar display is deliberately asymmetric:

Authentication	Default display	Dollar meaning
OpenAI API key	Exact runtime tokens + dated API estimate	Estimated billable API charge
ChatGPT Plus/Pro/Business	Tokens + credit equivalent + quota snapshot	No dollar amount
Subscription with --api-equivalent	Same data + API comparison	Comparison only; not charged

CtxWise never calls included subscription usage “money spent”. OpenAI states that ChatGPT credits have no cash value, so CtxWise does not invent a universal credit-to-dollar conversion. See Cost semantics.

Quick start

Requires Node.js 20 or newer and a working Codex CLI installation.

npm install --global @framy2/ctxwise
ctxwise doctor
ctxwise snapshot
ctxwise audit
ctxwise map --out ctxwise-context.mmd
ctxwise lock
ctxwise drift --fail-on-drift

Run these commands from the directory where Codex will work. CtxWise discovers the same project root markers and root-to-current-directory guidance chain, including configured fallback filenames and the aggregate project-doc limit.

If ctxwise doctor reports that Codex is unavailable, install the official CLI with npm install --global @openai/codex. On Windows, do not rely on directly executing the private binary inside the packaged desktop app. CtxWise detects the public npm launcher automatically.

For a custom Codex executable, set CTXWISE_CODEX_BIN. The former CTXRAY_CODEX_BIN name remains a compatibility fallback for existing setups.

GitHub renders the generated Mermaid file locally. Labels contain only the metadata already returned by audit, not prompt text or config values. The headline is a known startup estimate: AGENTS.md text and skill discovery metadata are counted; configuration files are marked as metadata, not falsely treated as prompt text.

Add a receipt after a Codex answer

ctxwise run --receipt --prompt-xray --model gpt-5.6-terra "Review the current diff"

Example output:

Fake answer...
CtxWise receipt · prompt ≈ 1,003 / 1,050,000 (0.1%) · 10,000 input (8,000 cached) + 500 output · credit equivalent ≈ 0.29 · quota 37% used · rates 2026-08-08

For a subscription-only API comparison, opt in explicitly:

ctxwise run --receipt --prompt-xray --api-equivalent --model gpt-5.6-terra "Review the current diff"

--prompt-xray asks Codex's experimental local debug command to render the model-visible input before the turn; CtxWise converts its character count into an explicitly estimated token value. The consumed input/output counters come separately from turn.completed and may aggregate several model calls. The footer itself is rendered locally after completion and consumes no model tokens.

Inspect model-visible prompt structure

Capture the experimental Codex diagnostic, then analyze the saved JSON:

codex debug prompt-input "Review this repository" > prompt-input.json
ctxwise xray prompt-input.json

CtxWise reports role counts, characters, and explicitly estimated tokens. It does not include prompt text in its report.

Compile native Codex profiles

ctxwise profile examples/ctxwise.yaml --dry-run
ctxwise profile examples/ctxwise.yaml

The second command stages files under .ctxwise/profiles. Installing into CODEX_HOME is a separate, explicit action:

ctxwise profile examples/ctxwise.yaml --install

Existing profiles are copied to ~/.codex/.ctxwise-backups/<timestamp>/ first.

Detect unreviewed context drift

Create a private, redacted baseline, then compare it with the live setup:

ctxwise lock
ctxwise drift
ctxwise drift --fail-on-drift

The last command exits with status 2 when a skill, instruction, profile, or plugin surface was added, removed, or changed, making it suitable for CI. For a fully offline comparison between saved files, use ctxwise drift baseline.json --current current.json.

Create a reproducibility lockfile

ctxwise lock --out ctxwise.lock.json

The lockfile contains hashes and relative paths, not prompt history. Secret-like config values and all MCP environment values are redacted before hashing.

Install the Codex plugin from a checkout

The repository includes a validated marketplace and plugin bundle:

codex plugin marketplace add .

Restart the ChatGPT desktop app, open the Plugins Directory, select the CtxWise marketplace, and install the plugin. After the repository is public, the same marketplace can be added using its GitHub owner/repository shorthand.

The bundled $ctxwise skill has implicit invocation disabled. Its instructions are loaded only when the user explicitly invokes it.

Architecture

flowchart LR
  A["Codex config, AGENTS.md, skills, plugins"] --> B["Audit + X-Ray"]
  P["CtxWise YAML policy"] --> C["Profile compiler"]
  J["codex exec --json"] --> D["Exact usage parser"]
  S["Codex app-server"] --> E["Plan + quota snapshot"]
  R["Dated OpenAI rate catalog"] --> F["Receipt policy"]
  D --> F
  E --> F
  B --> O["Local reports"]
  C --> O
  F --> O
  O --> L["Redacted capability lock"]
  L --> G["Drift guard"]

Loading

See Architecture and Privacy and security.

Measurement labels

• Exact: returned by the Codex runtime or account surface.
• Estimated: derived from a declared character proxy or dated rate card.
• Unknown: unavailable. CtxWise never replaces it with zero.

Claims about savings require comparable tasks that pass the same quality gate. CtxWise does not translate token estimates into a weekly allowance when Codex does not expose that conversion.

Current limitations

• A literal inline footer is available through ctxwise run. Codex does not currently document a plugin API that mutates a native desktop assistant message after generation, so the desktop plugin uses a separate result.
• codex debug prompt-input and app-server are version-sensitive surfaces. CtxWise fails closed to unknown when data is unavailable.
• turn.completed.input_tokens is aggregate consumption, not current context occupancy. Without --prompt-xray, CtxWise prints prompt context unknown instead of dividing that aggregate by the model window.
• Runtime MCP tool schemas and built-in tool schemas are not included in the static audit estimate; the audit reports that gap explicitly.
• Static discovery reads project-root markers from the active user config. A one-off Codex CLI override that is not present in that config cannot be inferred by a separate CtxWise process; pass --project explicitly.
• The bundled 2026-08-08 catalog covers GPT-5.6 Sol, Terra, and Luna. Supply a reviewed catalog with --pricing for other models or newer prices.
• Token-derived dollar estimates exclude unobserved tool-call fees and cache write classes.

Development

npm ci
npm run check
npm run build
npm run validate:plugin
npm pack --dry-run

The test suite includes unit, integration, and process-level CLI tests. Coverage thresholds are at least 80% for statements, branches, functions, and lines. See the v0.1 TDD evidence and the community-reproduction TDD record, plus the task-filter TDD record.

Read CONTRIBUTING.md, SECURITY.md, and the roadmap before opening a substantial change. Efficiency claims follow the public evaluation plan.

Official references

• Codex CLI commands and prompt-input
• Codex pricing, credits, and usage limits
• OpenAI API model prices
• Codex profile files
• Codex app-server protocol
• Codex plugins

License

Apache-2.0. See LICENSE and NOTICE.