ctx-kit: file-based task handoffs for Codex and Claude Code, with a two-chat example

[bestiaya]

I maintain ctx-kit, an MIT-licensed set of workflows for Codex and Claude Code. It keeps a task's current decisions, open questions, unfinished work and next step in project Markdown files so another chat can pick up that task.

The workflow is small to describe: start a task, save a handoff, then give a new chat the task-file path. A project board helps find the right task; a single current owner maintains its record. This is an agent-followed procedure, not a storage engine that enforces updates.

There is a fictional example you can run in a temporary project. Chat 1 checks an existing CSV command, records an exact-integer constraint, and leaves a filter option unfinished. Chat 2 receives only the task-file path, implements the option, and runs a separate acceptance script. That makes missed decisions visible without sharing a private project.

Repo and install packages: https://github.com/bestiaya/ctx-kit
Example: https://github.com/bestiaya/ctx-kit/blob/main/docs/TRY-IT.md

The example passed in two independent agent contexts. Native desktop skill discovery and the full Claude client flow have not been verified for this release; one standalone Codex CLI attempt stopped at a TLS connection error. The guide is English, while the shared workflow instructions are currently Chinese. It was developed with AI assistance and is an independent community project.

I haven't established that this is better than a good handoff note or native recovery. If you work across several chats, I'd value one concrete comparison: what does your current handoff already preserve, and what still makes you repeat work? A small, redacted example or a report that the extra process isn't worth it would both help.

[bomtobe-hgkim]

One concrete comparison: run the same two-chat CSV task with (a) your task-file handoff and (b) a concise conventional handoff note. Keep the integer constraint and unfinished filter identical; give Chat 2 only the handoff plus the same filter request. Check that it preserves exact integers, implements the filter, leaves the open option unresolved, and avoids reopening accepted decisions. That tests continuity, not just demo completion. I work on Orke at BOMTOBE, a Claude Code/Codex workspace; this is a proposed eval, not a result from Orke.

[bestiaya]

I tried the CSV task in two fresh Sol contexts, one with a compact handoff note and one with ctx-kit, carrying the same facts. Both passed the unchanged tests, left JSON output unapproved and kept large-file performance unverified, so this small synthetic example doesn't show an advantage over a good handoff note.

[bestiaya]

I checked our earlier validation records. We have a cross-session recovery test, and a real compaction trace where previously saved state was read and used to finish pending work.

The motivation behind ctx-kit is to capture decisions, their rationale and open questions during the discussion, so continuation doesn’t depend only on the compacted context. My CSV comparison started with those facts already recorded, so it didn’t test that capture-and-update process.

Your plain-note baseline remains useful for evaluating that part. The existing evidence shows the save-and-resume path working in specific cases; whether ctx-kit does it more reliably than a well-maintained note remains open.

[TheodorNEngoy]

For Turnfeed, the handoff detail I most need is the boundary of each verification: source revision, what was checked locally, and what remains untested on the published Site. Otherwise, ‘tests passed’ is too easy to inherit as a broader claim.

A small example: ‘Photo ownership tests passed locally at revision X; hosted identity/storage behavior is untested.’ In a new chat, I’d check whether that unresolved condition stays unresolved until new evidence arrives. Recording the evidence also helps the next chat decide which checks need repeating after a change.

I haven’t tried ctx-kit, so this is a comparison case from our handoff process. It would be useful alongside your CSV example.

[bestiaya]

I included an explicit "large-file performance untested" item in both handoffs. Both receivers kept it open after the local CSV checks passed; native client loading and deployment behavior remain outside what this run established.

[TheodorNEngoy]

Thanks for running that check. Keeping the untested item open after local tests passed is the distinction I meant. The comparison showing no advantage over a good handoff note is useful to know, too.

[bestiaya]

You could keep your full test reports where they are. With ctx-kit, a small task overview would link to them and track what was checked, against which revision, and what still needs checking. The workflow asks the agent to update that overview as decisions and results change.

Would you be open to trying it on one small task across two sessions? The useful test would be whether the next session picks up those boundaries correctly, and whether the extra record is worth maintaining compared with your usual handoff.

[TheodorNEngoy]

Thanks for the invitation. I can't commit to a two-session trial right now; I'm focused on Turnfeed's own user feedback. If I do try it, I'll compare it with our usual handoff and include the upkeep cost.