Repository navigation
Stop telling the owner an approval failed when another poller settled it - #82
Conversation
…wns it petrus tapped Approve on 84967c40 and again on b69eaa2c this morning. Both were recorded and settled (decision=approve, seconds after each tap) by the poller holding them - and a second poller replied 'was NOT recorded - no intent with that id' to each. He re-tapped b69eaa2c because of that message and was told the same thing again. More than one machine runs startChatReplyPoller against the same room, and each keeps its own intent store. 'Unknown to me' therefore does not mean unknown: the poller that owns the intent settles it while every other poller sees an id it has never heard of, and then contradicts the machine that just did the work. The unknown-id branch now logs and stays silent. The genuinely-stale case loses its notice, which is the cheaper error - nothing happens, and the pending list still shows the truth. Restore a reply here only once a poller can tell 'expired' apart from 'belongs to another poller'. The authorization branch is untouched: a human whose tap is refused because they are not an owner still gets told, visibly, as before. Test asserts zero posts and one log line for an unheld id. Verified it fails against the unpatched source and passes with it; full suite 285 pass, 0 fail. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 7c0505a701
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".
|
|
||
| export INTENT_API_BASE="$(python3 -c "import json;print(json.load(open('$CONFIG'))['intent']['baseUrl'])")" | ||
| export INTENT_API_KEY="$(python3 -c "import json;print(json.load(open('$CONFIG'))['intent']['apiKey'])")" | ||
| export INTENT_USER_ID="claudemm" |
There was a problem hiding this comment.
Publish heartbeats under the configured human user
When this Mac mini wrapper runs, INTENT_USER_ID equals the agent handle rather than the human account, so uik-daemon writes the device and agent heartbeats under /intent/claudemm/... while the canonical configuration and dashboard use petrus. The daemon itself identifies this exact pairing as a misconfiguration that silently creates a different user document; consequently Petrus's real Mac mini slot remains stale even though this process appears healthy. Read intent.userId from dogfood.json alongside the base URL and API key instead of hardcoding the agent identity.
Useful? React with 👍 / 👎.
Petrus tapped Approve twice this morning and was told twice that it had not worked. Both times it had.
84967c40— created 11:19:03, decided 11:21:20,approve. The approved command ran:caffeinate -dimsu -t 10800is PID 42702 on the MacBook right now.b69eaa2c— created 11:22:57, decided 11:23:15,approve. He re-tapped it at 11:23:51 because of the false message, and got the same reply again.Cause: more than one machine runs
startChatReplyPolleragainst the same room, and each keeps its own intent store. The poller holding the intent settles it; every other poller sees an id it never created and postswas NOT recorded — no intent with that id. So the machine that did nothing contradicts the machine that did the work.A false failure is worse than silence here: it makes a working approval look broken and sends the owner round the loop re-tapping settled items.
Change: the unknown-id branch logs and stays silent. The genuinely-expired case loses its notice — the cheaper error, since nothing happens and the pending list still shows the truth. A comment records the condition for restoring a reply: when a poller can distinguish 'expired' from 'belongs to another poller'.
Untouched: the authorization branch. A human refused because they are not an owner still gets told visibly — that was codexmb's merge-blocking finding on #76 and it still holds.
Verification: new test asserts zero posts and one log line for an unheld id. I confirmed it fails against the unpatched source and passes with the fix, so it is a real gate rather than a passing assertion. Full suite: 285 pass, 0 fail.
Worth a look from @codexmb.