Context Curator
A browser-only manual editor for a completed exploratory span in an agent transcript. Inspired by openai/codex#14747, closed as not planned. It keeps the surrounding context and replaces one contiguous span with a conclusion you write.
Install and use
Open site/index.html directly in a modern browser. No installation, dependencies, account, server, or model calls. The hosted page offers that same complete HTML as a free offline download. The initial interactive example is an illustrative debugging transcript, not a recorded execution or performance benchmark.
- Try the sample, open a JSON file, or paste JSON and press Load this JSON.
- Choose first and last messages (1-based, both inclusive). Write the root cause, supporting evidence, unresolved questions, and next action.
- Press Replace selected span. Read the complete message diff and expandable full JSON views. Size measurement compares compact UTF-8 JSON bytes, including conclusion metadata, not tokens, costs, or Markdown file size. A long conclusion can make the result larger.
- Download the Markdown continuation packet or curated JSON. Save untouched original downloads the exact loaded text, including whitespace. Undo replacement restores the previous working result (one level). Replacements always target the loaded original, never accumulate silently.
- Review the packet before manually handing it to a fresh session. All messages are quoted as JSON in Markdown, with fences safe for embedded backticks. No commands or transcript HTML are executed.
JSON transcript format (version 1)
{"version":1,"title":"Optional metadata","messages":[
{"id":"m1","role":"system","content":"Keep the requirements."},
{"id":"m2","role":"tool","content":"Exploratory evidence."},
{"id":"m3","role":"user","content":"Next task."}
]}
Required: version: 1, 1–500 messages, each with a unique nonempty string id, role in system|user|assistant|tool, and string content. Limit: 2 MiB UTF-8 source JSON. Additional top-level and message fields survive. Retained messages are structurally identical; formatting in curated JSON is normalized. The original source string is preserved separately, byte-for-byte for valid UTF-8 input. Invalid imports leave the last loaded original and result available. Files must be valid UTF-8.
sample.json is the example. The manual replacement is an assistant message tagged curator.kind = "manual-conclusion", with replacedIds for the entire selected span and a unique generated ID. Selected messages and their metadata are absent from the curated export; they remain in the original. Selection may include any role: review carefully before removing requirements or instructions.
Privacy and limits
Transcript processing stays in browser memory. No fetch, telemetry, third-party fonts/scripts, local storage, session edits, or configuration access. Closing or reloading loses working state; download the original and exports first. Only the optional signup form posts email, plan, and note to the existing /api/signup service. Offline signup is unavailable. ?joined=1#join shows the thank-you state.
This does not import native Codex/Claude session files, automatically summarize, resume a session, count model tokens, or guarantee a good conclusion. Convert an explicit transcript export into the documented format yourself. It is a handoff aid, not a compaction API. Do not include sensitive context in the signup note. A malicious transcript can still contain misleading instructions: quoted Markdown is not an agent security boundary.
Develop and verify
From this directory:
python3 build.py
node --test tests/core.test.cjs
# Browser checks require Playwright installed outside this repo:
NODE_PATH=/tmp/context-curator-browser/node_modules node tests/browser.cjs
build.py embeds core.js, the sample, and site/app.js into site/template.html, writes the standalone site/index.html, and copies the sample and README beside it. Libre Baskerville Latin regular/italic WOFF2 (Fontsource package, SIL OFL 1.1 in site/fonts/OFL.txt) is vendored and embedded as data URLs so the single-file download keeps its typography offline. No build dependency beyond Python 3; core tests require Node 18+. Browser tests accept CHROME_PATH (default hb-5 Chrome), PAGE_URL (default file://), and EVIDENCE_DIR (default /tmp/context-curator-evidence). They cover desktop/phone overflow, undo, imports, exports, escaping, form payload and thank-you state, and screenshots. Run python3 tests/signup.py /path/to/existing/server.py for signup acceptance: it imports the existing handler with temporary data, never the real signup list. Checked launch evidence is in evidence/; native macOS browser verification was blocked by this seat sandbox (Mach bootstrap permission denied). Node core tests did run on macOS. Container Chromium passed both file:// and a local HTTP /v/context-curator/ route, including reopening the downloaded HTML offline. The public route is deliberately not polled: publication happens after the issue reply.
Launch is confined to this venture directory; the factory runner publishes site/ at /v/context-curator/ after the issue reply. No shared deployment files are changed.