Repository navigation
evidence upload host: Today card screenshots (do not merge) - #1
jingchaodev wants to merge 1 commit into
Conversation
The dashboard has no view of what was worked on today. The memory pipeline writes a dated daily-history file, but its only surface, the Daily history card on Overview > Memory, is hidden whenever vector memory is active (the gate hides the three editable text documents once the store migrates), so a migrated install shows nothing. The Usage card counts sessions started today and the sidebar lists sessions, but neither is a recap of the day. Add a fourth summary card, Today, to the Overview 2x2 grid with a URL-backed `?view=today` drill-in, the same pattern as Usage, WakaTime and Memory. Zero model calls: the card is a view over data the dashboard already holds. It counts sessions active in the browser's local day (live slots from redux plus archived rows from /api/sessions, the live slot winning on a shared key; incognito and temporary sessions never listed), states the running count, the entries already saved to memory and the last activity time, and shows one chip per folder. The drill-in lists the sessions grouped by folder, newest folder first and unfiled last, with a click that opens the session (switchSlot for a live slot, resumeFromHistory for an archived one), beside the day's memory entries split on the `#### HH:MM TZ` blocks append_history writes. The memory half lags by design, so its empty state names the consolidation window from memory settings and its Open Memory action switches to the Memory view, where Summarize now already lives. The MemoryTab hide is untouched. Backend: GET /api/memory/history accepts `?date=YYYY-MM-DD` (canonical form only, else 400 invalid_history_date) and returns that one day's file through the same guarded reader in both memory versions; a missing file is an empty day. The path is rebuilt from the parsed date, so query text never becomes a path component. The date names a server-local day, the clock that names the files. Copy: the Usage card's fragment now reads "sessions started" so its today count and the Today card's "sessions active today" read as two sets rather than a disagreement; every new string ships in the 12 catalogs with per-locale plural forms. Tests: dated read for global v1 and member v2, empty day, five malformed date shapes; vitest for the pure helpers (local date key, today filter with live-wins dedupe and incognito exclusion, folder ordering, entry parsing), the drill-in states and both open doors, and the Overview shell round trip including the interpolated running count.
|
Upload host only; the assets are referenced from the upstream PR. |
GPT 5.6 Review —
|
|
Evidence upload for kirodotdev#8985 head 15253299b (round 9 captures; label = AWS profile (optional)). |
|
Evidence upload for kirodotdev#11092 round 9 (card states the scope sentence only; folder chips carry a spoken label). |
|
Evidence upload for kirodotdev#8985 round 10 (pointerdown nudge, denial audit). |
|
Evidence upload for kirodotdev#8985 round 10 (pointerdown nudge, denial audit). |
|
Evidence upload for kirodotdev#11092 round 10 (error notices lead with a plain title; raw server text as detail). |
|
asset upload for kirodotdev#11092 round 10 (history-error notice with plain lead) |
|
Evidence uploads for kirodotdev#9314 (drain-capture placement, dark + light). |
|
Evidence uploads for kirodotdev#8985 (name input without the redundant placeholder; 8 captures from the built SPA). |
|
Evidence uploads for kirodotdev#8985 (kb_ids guidance as persistent helper text under its input; 8 captures from the built SPA at 51a1fd5). |
|
PR kirodotdev#8985 evidence at 987fa5e (round after 51a1fd5): 8 captures |
|
PR kirodotdev#8985 evidence, round after 987fa5e: 10 captures |
|
Bedrock KB kirodotdev#8985 captures, round on the retained-triple/narrow-stack head (2026-09-19 c77): |
|
Evidence carrier for upstream kirodotdev#11377 (compaction methods, round 15): the rotated-compaction notice cards, captured from a harness gateway of branch feat/compaction-methods. Not for merge; the PR body links these attachment URLs. |
…irodotdev#12915) * fix(chat): wire the paste-block sidecar into ChatPane and SideChat ChatInput collapses a large paste into a `[ Paste #N · M lines ]` token only when its host passes `onPasteBlocksChange`. ChatPage did; the split view / member DM pane (ChatPane) and the side panel (SideChat) rendered the composer without it, so a big paste stayed raw text there. Both hosts now keep the blocks with the same lifetime as their text: - ChatPane: pane-local `pasteBlocks` state, parked and restored with the text through chatPaneDrafts (a third sessionStorage store), expanded for the model in doSend and doSteer, kept as tokens on the bubble with `meta.pastes` + saveStoredPaste so the row renders a chip, handed back with the text on a refused send (carryPastes re-numbers a colliding block), and put back EXPANDED by a queue-card cancel because the stash carries no blocks. - SideChat: blocks live in the in-memory side-chat draft entry next to the text (a remount restores both; neither outlives the page, matching the store's documented intent), expanded in `send` on both the send and steer branch, pruned whenever the text is written. Tests: page-level ChatPane and SideChat scenes (paste, pill, send, expanded wire text, cleared composer), a rebind and a refused-send restore for the pane, a remount for the side panel, and unit coverage of the pane store's new paste field + carryPastes. Capture harness entry + script for the PR evidence. Closes kirodotdev#11337 * fix(chat): park pane drafts as one value; carry pastes with a synced ref Review rounds 1-2 on the sidecar wiring: - The pane store held text, files and paste blocks under three keys with independent byte caps, so an over-budget slot could lose its block while the text kept the token — a reload then sent the token literally. The parked draft is now ONE stored value (`mc-pane-drafts`: {text, files, pastes}) under one cap: a slot is evicted whole or not at all. Drafts a tab parked under the two-store layout (a text string per slot, paths under `mc-pane-file-drafts`) are folded into the unified shape on the first read after a reload, so an upgrade mid-session loses nothing. - Two recoveries landing in one React batch both read the render-time `pasteBlocksRef`, so the second `setPasteBlocks` dropped the first's blocks while both tokens stayed in the text. `carryIntoComposer` advances the ref synchronously per recovery before scheduling state, so the second carries on top of the first; the rebind path advances it the same way. - `carryPastes` moves to pasteTokens.ts as the one owner of the kept-ids -> remapCarriedBlocks -> concat rule; ChatPage's two inline spellings now call it. - `readSideChatPastes` had no product consumer; removed. The SideChat tests observe block state through the composer and the wire instead (a fresh paste numbers from #1 again once the blocks are gone). - SideChat's too-long message counts the EXPANDED question; when a collapsed paste is part of it the message now says so (`question_too_long_with_paste`, all catalogs), since the composer visibly holds one line and a pill. Tests: a two-refusals-in-one-batch scene on the pane (fails without the ref sync: two `#1` tokens), whole-slot eviction, legacy fold and malformed-value drop on the store, the paste-aware too-long message.
|
Evidence carrier for upstream kirodotdev#11377 (compaction methods, round 16), captured from a harness gateway of branch feat/compaction-methods. Not for merge; the PR body links these attachment URLs. |
|
Evidence carrier for upstream kirodotdev#11377 (compaction methods, round 17), captured from a harness gateway of branch feat/compaction-methods at its rebased head. Not for merge; the PR body links these attachment URLs. |
…ger off (kirodotdev#15851) ## Problem A Linux system install writes a gateway unit that bakes `XDG_RUNTIME_DIR=/run/user/<uid>` and `DBUS_SESSION_BUS_ADDRESS` into its environment, so every agent runtime scope is created under that account's `user@<uid>.service`. With `Linger=no` logind stops the account's manager when its last login session ends, and systemd takes every runtime down with it: in-flight turns and subagents get SIGTERM and new spawns fail with "Failed to connect to bus" until the account logs back in. The gateway, running under the system manager, stays up and reports nothing, so the install gives no sign that agent work needs linger to survive a disconnect. ## Why it matters The system-scope install is the path a user picks precisely so the gateway outlives their SSH sessions. Without linger that promise is silently broken for the agent runtimes, and `kirocrew service install` prints nothing about it even though the user-scope remedy already documents `loginctl enable-linger`. ## Not a goal The architectural fix for defect #1/kirodotdev#3/kirodotdev#4 in the issue -- running the runtime scopes under the system manager vs. making linger a hard requirement, plus the disconnect fallback and reconnect recovery -- is a placement decision the maintainer routed to needs-investigation. This change implements only the install-time warning the maintainer named as a small, separable piece. It never changes the install outcome. ## What changed - `service/linux.py`: add `linger_warning_for_service_account()` (and a `_linger_enabled` probe reading `loginctl show-user <user> -p Linger`) that returns an operator warning when linger is off for the resolved service account, naming the `sudo loginctl enable-linger <user>` fix. Returns None when linger is on or cannot be determined, so the install reports nothing it is not sure of. - `service/controller.py`: print that warning after a successful system install, next to the AppArmor and headless-auth notices. Non-fatal. ## Tests - `test_linger_warning_fires_when_linger_is_off` / `_silent_when_linger_is_on` / `_silent_when_undeterminable` cover the probe's three outcomes. - `test_install_systemd_prints_linger_warning` / `_silent_when_linger_ok` cover the controller wiring. - Full `test_service.py` (413 passed, 3 skipped), black/isort/flake8/mypy and the comment-history gate all clean. Fixes kirodotdev#15678 Co-authored-by: kirocrew-agent <agent@kirocrew.local>
|
Evidence for kirodotdev#16123, second capture (Title Case card title, "context sent" input row). |
|
Evidence for kirodotdev#16123, third capture (a gap before the Last Month column). |
|
Evidence for kirodotdev#16123, fourth capture (the note and the warning say which way each figure errs). |
|
Evidence for kirodotdev#16123, fifth capture (the note says output is roughly the reply length divided by 4 and leaves out tool calls and reasoning). |
|
Evidence for the +/- stepper revision of PR kirodotdev#17183 (title-refresh row now matches Message Font Size). |
|
Evidence for the UX-review revision of PR kirodotdev#17183: the row reads "Default" at 0 with − disabled, the label has no placeholder letter, and a failed save shows its banner. |
… stable id (kirodotdev#17382) Collapsed paste tokens `[ Paste #N · M lines ]` were paired to their backing block by the display number N, resolved by `seq` alone. `nextSeq` returns `max(existing) + 1`, so it restarts at 1 once a send clears the composer's blocks. A sent message keeps the literal token text, and up-arrow recall reads it back into the composer — so paste -> send -> paste -> recall made the recalled `#1` token resolve against the new draft's `#1` block, splicing the wrong content into the outgoing message and discarding the draft. The token now carries the block's stable `id` (makePasteId) right after `#N` as an ENTIRELY zero-width run: each ASCII id char as 7 bits over U+200B (0) / U+200C (1), fenced by two U+2063 INVISIBLE SEPARATORs (encodeIdInvisible). Every carried code point is zero-width, so the token renders byte-for-byte as `[ Paste #N · M lines ]` — none of the id is visible in the textarea composer or its backdrop mirror — while findTokenRanges recovers a durable key by decoding the run. A new-form token whose id names no live block stays literal text (never mis-paired); a legacy `#N`-only token falls back to seq, but only when exactly one live block owns it (ambiguous legacy left unresolved, not guessed). findTokenRanges branches on the PRESENCE of the id fence, not on decode success: a token that carries a fence but whose run fails to decode (corrupted outside the app) stays unresolved rather than falling back to seq, so it can never splice a live same-seq block's content in (GPT 6.1 review). Backend: optimizer.py PASTE_TOKEN_REGEX matches the optional invisible id run too, so paste-content scoping (by seq) and the placeholder-preservation multiset guard keep working for the new token form. Previously the backend saw no tokens, forwarded no content, and the empty-vs-empty guard accepted a token-dropping rewrite — silently losing the paste on the paste -> optimize -> send path. No DOM/attribute change (the `data-paste-seq` chip linkage is within one live composer where seqs are already unique, so it is left as is). Tested: - pasteTokens.test.ts: the rendered token strips to exactly `[ Paste #N · M lines ]` with no visible id; encode/decode round-trip; acceptance (recalled token resolves to its own block or to nothing, never to a reused-seq block); migration (legacy `#N` resolves by seq; ambiguous legacy left unresolved). - test_optimizer.py: seq extraction and full-token multiset counting over tokens carrying the zero-width id run. - Updated the composer/bubble/pinned-prompt/draft-store/sidecar/terminal-command suites. 49 impacted frontend test files pass (1430 tests); tsc, eslint, the i18n gate, flake8 and mypy are clean. Fixes kirodotdev#13851 Co-authored-by: kirocrew-agent <agent@kirocrew.local>
|
Evidence for kirodotdev#17183 (title-refresh row, readout keeps its widest width). |
|
Evidence upload for PR 11377, round 19: the mid-turn compaction notice ( |







































































































































































































r7 evidence host for kirodotdev#11092 (uploaded 2026-09-16T19:03Z)