Block publish on missing Notion instead of silently degrading#45
Block publish on missing Notion instead of silently degrading#45parsakhaz wants to merge 1 commit into
Conversation
An epic publish (bloomapi/bloom-mono#338) ran with no Notion connection and a stale local copy of these files: the Notion page was never created, the GitHub issue was a ~10-line summary that referenced a 'linked work item' that didn't exist, and the refs/ artifacts (data pulls, code research) existed only on one machine. The degraded path executed silently when the right move was to stop and connect. - notion skill setup: with no Notion tools/CLI, interactive sessions now block with the connection path and retry after connecting; only an explicit user skip returns NOTION SKIPPED BY USER. Headless runs keep the non-blocking NOTION UNAVAILABLE so cron/cloud never hangs on a prompt. - publish procedure: degraded mode is gated on those two explicit signals, sets the issue body to the full item.md (not the summary), records the skip in frontmatter, and the done-when now names the failure this PR is preventing. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 63c17bcd34
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| `NOTION UNAVAILABLE` from a headless run. | ||
| 6. Degraded mode: the issue must carry everything a remote `/do` needs — | ||
| set the issue body to the **full `item.md` content** (not the summary), | ||
| record `notion: SKIPPED — re-run notion publish when connected` in the |
There was a problem hiding this comment.
notion: for skipped publishes
When a user explicitly skips Notion, this writes a non-URL sentinel into the notion: frontmatter. Later /do wrap-up treats any notion: reference as a real Notion target (tyler/.claude/skills/do/SKILL.md lines 170-173) and invokes the notion upload operation with it, so degraded items can fail at wrap-up or try to upload plan/wrapup to SKIPPED — ... instead of simply skipping or re-publishing. Use a separate flag/status or update /do to recognize this sentinel before adding it to the canonical notion: field.
Useful? React with 👍 / 👎.
1 participant
What happened
An epic publish today (bloomapi/bloom-mono#338) ran in a session with no Notion connection. The notion skill returned
NOTION UNAVAILABLEand the publish continued per the letter of the procedure: the Notion page was never created, the GitHub issue went out as a ~10-line summary whose "full spec lives in the linked work item" pointed at nothing, and therefs/artifacts (PostHog funnel pulls, code research with file:line findings) existed only on one machine. The refs layer is what justifies keeping specs lean — a publish that drops it is a failed publish, but the current wording treats it as an acceptable degraded outcome and lets it happen silently.(Aggravating factor: the session's synced
~/.referencescopy predated #44, so it didn't even run the artifact-comment fallback. That's a sync problem, but it shows silent degradation compounds.)Change
tyler/.claude/skills/notion/SKILL.md— setup step 1 no longer silently returnsNOTION UNAVAILABLEwhen a human is present. Interactive sessions block, surface the connection path (claude mcp add --transport http notion https://mcp.notion.com/mcp+/mcp), and retry once connected; an explicit user skip returnsNOTION SKIPPED BY USER. Headless runs (cron/cloud) keep the non-blockingNOTION UNAVAILABLEso nothing hangs on a prompt nobody can answer.tyler/references/publish-work-item.md— degraded mode (the #44 artifact-comment fallback) is now gated on those two explicit signals instead of firing automatically; it must set the issue body to the full item.md (not the caller's summary) and record the skip in frontmatter so a laternotion publishre-run is an obvious TODO. The done-when now names the failure mode this prevents: artifacts stranded on one machine, or an issue referencing a Notion page that doesn't exist.Why blocking is right here
Publish is the one moment the artifacts leave the local machine. Every caller (
/create-feature,/create-epic,/create-issue) is human-invoked, so there is always someone present to either connect Notion or consciously accept degraded mode — the cost of pausing is seconds, the cost of silent degradation is a work item that can't be picked up remotely and dangling references in the issue. Headless is carved out explicitly so/do-adjacent automation never deadlocks.🤖 Generated with Claude Code