feat(console): unread nudge on the read path (pull/list) + final-check convention#253
Open
ivanmkc wants to merge 6 commits into
Open
feat(console): unread nudge on the read path (pull/list) + final-check convention#253ivanmkc wants to merge 6 commits into
ivanmkc wants to merge 6 commits into
Conversation
…der, list summary) The nudge fired only on writes (push/patch/scoped status) — a human message arriving after the agent's LAST write was never surfaced. Close the gap on both halves of the read loop: - server: scoped GET /state (pull) now sets the x-termchart-inbox-unread header, same once-per-new-message semantics as the write path - cli pull: prints the content nudge off that header - cli list: prints a workspace-wide unread summary via GET /inbox?all=1 — the one read an agent still does when it's done writing, and the only place cross-scope messages can surface; re-announces until read (an explicit catalog read should show what's unread); stderr-only so --json stdout stays clean - skill: final-check convention — run `termchart list` as the last termchart action; plugin 0.16.4, marketplace entry synced (was stale 0.15.0) - eval: TC-INBOX-MIDTASK — seed_inbox.py gains @await scopes posted by a background watcher only after the agent's board appears (a human typing mid-task); run.sh wires the watcher up and kills it after the runner Verified live: pull nudges after a post-push message; list re-announces until read and shows cross-scope unread; quiet after a real read; --json stdout unpolluted. Touched tests 33/33.
…ecklist tick - TC-INBOX-DECIDE: agent posts two chips, blocks on inbox --wait; console_human.py (the simulated human, launched by run.sh) clicks 'green' via GET /suggest -> POST /inbox action once the chips are up. The choice reaches the agent only through the console. Deterministic gate: button color prop is the clicked id. - TC-INBOX-TICK: human ticks an editable Checklist out-of-band (POST /interact — produces NO inbox event by design); the agent must observe it by polling pull and reply naming the item. Gates: tick landed in the spec + reply names it. - pull-grep objectives now grep raw `pull` output (content in --json is an escaped string, so quote-anchored patterns silently miss — caught in dry-run). Dry-run verified: watcher clicks + ticks; inbox --wait returns the action; all objective commands pass against simulated-agent state.
…rop name Run 1 near-miss: the flow worked end-to-end (chips, --wait click, green button per judge, reply) but the agent expressed the color without a literal "color" prop. Gate on content containing the clicked id and NOT the un-clicked one (blue is also Mantine's default primary, so a click-blind agent still fails).
The inbox cases share one viewer/workspace; an agent following the list unread-summary in one case can consume another case's seed (observed: DECIDE read UNPROMPTED's seed before that case ran -> false FAIL). Singleton case-sets + a per-set clear/seed/watch/run cycle make that state-sharing unrepresentable.
…naming requirement Run-1 miss was real signal: the agent parsed pull --json (content is a double-encoded string), extracted an empty item id, and replied 'You checked: '. Keep the strict objective; stop steering the agent into the harder surface.
The scenario said the item 'shows "checked": true' — spaced — but interact
re-stringifies the spec compact ("checked":true), so agents that grepped the
prompt's literal form polled for 2 minutes against a pattern that cannot match.
Say compact explicitly; the parse robustness burden stays on the agent.
Owner
Author
K=3 × 2 agents × 5 cases — repeat-run grid (isolated mode)
Core console flows are stable: 24/24 case-runs green across PROMPTED, UNPROMPTED, MIDTASK, and DECIDE — every repeat, both agents. In particular:
TICK (checklist observation) exposed two real papercuts, both diagnosed from transcripts and fixed in-flight (each run re-reads the cases, and run 3 passed both agents on the final config):
Harness hardening shipped along the way: CI green. Ready for review. |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What
Closes the biggest gap left after #252: the nudge only fired on writes. If the human typed after the agent's last push, the message was never surfaced — the eval passed only because its seed existed before the push.
GET /state(pull) now setsx-termchart-inbox-unread(same once-per-new-message semantics as the write path).pull: prints the content nudge from that header.list: workspace-wide unread summary viaGET /inbox?all=1— the one read agents still do when done writing, and the only surface where cross-scope messages can appear. Re-announces until read (an explicit catalog read should show what's unread); stderr-only,--jsonstdout stays clean.termchart listas your last termchart action. Plugin 0.16.4; stale marketplace entry (0.15.0) synced.TC-INBOX-MIDTASKcase —seed_inbox.pygains@awaitscopes posted by a background watcher only once the agent's board appears, modelling a human typing mid-task. The write-path nudge can't fire; thelistsummary must catch it.Verification
Live (local viewer): push → human message after →
pullnudges with content;listre-announces until read + shows cross-scope unread; quiet after a realtermchart inboxread;--jsonstdout still valid JSON. Touched unit tests 33/33.Agent eval: K=3 × {Claude, Gemini} × 3 cases (PROMPTED / UNPROMPTED / MIDTASK) running now — results will be posted as a comment before undrafting.
Known race (documented in the case): if the agent's
listlands before the watcher posts the seed (~0.5s poll), MIDTASK can miss — acceptable for an eval of a genuinely asynchronous channel.