Repository navigation
AbstractObserver 0.8.0
Requires AbstractGateway 0.15.0 (GET /runs coalesced per plane + query; /host/runs titles and waits) over AbstractRuntime 0.10.1 (the active-run index), and @abstractframework/ui-kit 0.8.7.
Added
- Board redesigned around the blocker, not the status word (backlog 0003 "observe agents, not cards", wave 1). Five lanes: Needs you (inline Approve/Reject/Answer, and Resume for a paused run), Working, Waiting on…, Scheduled, Done. Every card reads title first (the automation's title, else the run's task excerpt from
GET /runs/{id}/input_data, else the workflow id marked untitled), then the subject line (automation · occurrence #n · workflow, or session · workflow), then a state sentence with its object and actor: "needs your approval: send_email · 23m · you", "running subrun react loop · llm_call open 38m", "waiting for subrun X: asks a question", "sleeps until 02:00 UTC (in 1h 12m) · clock", "waiting for event inbox.message · external", "waiting, reason unknown (key) · inspect". A parent parked on its own running child reads running — the gateway's own reading of an occurrence — and inherits the child's open step and staleness; a running run always shows its last ledger touch and gets a no progress mark past the threshold you pick in the board header (5m / 15m / 1h, remembered per viewer). Automation controllers are the automation's own Scheduled card: next run, last run, number of turns, cadence and context mode. The (i) popover carries the full task, workflow, cadence, context mode, tools, model/provider, iterations, tool-approval policy, notify channels and workspace, Name it (an inline rename throughPATCH /automations/{id}with the revision the board read) and Open automation. Run ids never print; ⧉ copies them. - Observe → Agent, the view a board card opens on: the header (title, subject, state sentence, the one action it calls for, copy id), Can do (model/provider, tools, tool policy, workspace access — the automation's policy tagged as such), Inputs → Outputs as AbstractFlow's pins with their live values (start pins from the flow with
input_data, end pins with the run'soutput; "not provided" and flow defaults said; a pin-less end node renders as one untyped output), Context (the history window replayed vs dropped, the token gauge against the run's limit, flow warnings), Run tree (controller → occurrence → children, each with its own sentence), Spend (tree-inclusive, labelled so) and Ask. The flow comes fromGET /bundles/{b@v}/flows/{f}; the session's rows are re-listed when the subject moves and once a minute. - Agent profile (Observe → Agent, operator 2026-10-14 11:13): "Can do" became a profile in five groups — Who (workflow, subject, role, model · provider, system prompt, the task), Access (tools it may call vs tools it did call, with counts and "called outside the declared list"; tool approval; untrusted-input tools; email recipients; workspace mode, root, allowed and ignored paths), Does (the state sentence, LLM/tool/subrun counts, delegated subruns, the outcome), How long (started, finished, duration, iterations used of max, token limit, cadence · next run · count, retry policy) and After (notify channels, context mode and growth, next occurrence or why none). Every row names its source — used (the ledger), run (the inputs), flow default, automation — and absent facts say why ("not on this wire", "not provided, no default").
- No "…" anywhere on the Agent view (operator 2026-10-14 11:11): every value — provided, flow default, output, "also sent" — renders through one value component: short scalars inline, long strings collapsed by size with show all (N chars) / show less, arrays and objects as a complete key/item list with the full JSON behind a disclosure. Every field has a copy button on hover or focus (the full raw value; JSON pretty-printed), as cards and messages do.
- Flow tab (operator 2026-10-14 11:15): a run-state banner (ongoing with the state sentence and "at ", or finished with the sentence and "N nodes visited in D"), AbstractFlow's run-status vocabulary on every node (running ring and pulse, waiting dashed ring, failed ring, done check; one here chip on the current node), a facts panel on hover or keyboard focus (visits, first/last, time in steps, the LLM line, the tools table, waits, subruns, the last result and error) with Copy facts (JSON), Copy records (JSONL) and Copy context (the exact messages of the last LLM call), and a toolbar Copy run state that hands an AI one JSON document: where the run is and why.
- No duplicate blockers (operator 2026-10-15): a parent run that waits on a subrun which itself needs you is blocked by nothing else, so it no longer gets its own card in Waiting on…; the Needs-you card says blocks (a link) and, when the child's own name is a generated id, takes the nearest titled ancestor's title with "subrun · " as its secondary line. Waiting on… now holds only runs blocked by something other than you.
- Cancel all in a lane (operator 2026-10-16): the heads of Needs you, Working and Waiting on… carry a cancel icon right after the title when the lane has cards and the host wired cancel; it asks in a band under the lane head (Cancel all 53 runs in Needs you and the 36 tasks they belong to? Everything working under those tasks stops too; automation schedules are not touched. Cancel all / Keep, never squeezed into the head's row), then cancels through the host's batch handler with a live count (cancelling 3 / 36 · 1 refused) and a receipt (cancel asked for 35 tasks · 1 refused). Scheduled (pause is its verb) and Done have none.
- Task cards (operator 2026-10-16 01:31: "we should be handling sessions (runs)… see their subsessions… see the parent… be able to cancel either"; design adversary GO,
untracked/redesign/TREE-CARD-adversary.md). A board card is now one TASK — a parent-less run, an automation occurrence, or a child whose parent the listing lacks — in the lane of its deepest live blocker (an asking or stuck subrun → Needs you, a working subrun → Working, else the root's own wait). Inside: the task text as title, then the CHAIN root → leaf, one row per live run with its own words (waits for ↓ above a live child, else its state sentence), a name that opens that run and, on every subrun row, a ✕ that stops that subrun only and says what follows (Stop this subrun only: its parent continues and may delegate again); then one ask block per asking subrun (builder asks since …, the request, Approve / Reject / Review & answer), or, when a row is stuck, the attention block: the stuck run's name (it opens that run), stuck since (), the situation in a sentence built from the facts (Its subrun builder was cancelled 1h 43m ago and this run was not resumed.), then Cancel task and Open the run →; the card's status word says stuck. The footer's ✕ cancels the task. Children never get a card of their own, so "blocks " is gone; the Needs-you count counts tasks and its tooltip says N tasks · M asks; cancel-all counts tasks. Scheduled and Done cards are unchanged. - Readable names on the task card (operator 2026-10-16 02:36: "multiagent-coding 0.0.13 waits for ↓ └ multiagent-coding 0.0.13 … absolutely not readable; at least use the label of the step"). A run now reads by its FLOW ("multiagent-coder", "multiagent-coding"), a generated subworkflow by the NODE it was spawned from ("builder", "scout_code"), the whole id on hover; a parked parent says which step waits (waits at build ↓), and a stuck run names its subrun by that step (stuck: subrun builder was cancelled 1h 43m ago, this run was not resumed); a lone stuck run (no chain) gets the same attention block in place of the ask actions. Hovering a card's title opens the task's full text in the elevated panel (the shared markdown renderer), never a native tooltip.
- A Cancel cancels the task, not one run (operator 2026-10-16 01:04, "there are still 10 items?"): the gateway runner cascades a cancel DOWN (root + every descendant at apply time) but nothing cascades UP — a child cancelled on its own hands its parent
{success: false, cancelled: true}and the parent proceeds; a multi-agent parent delegates again and asks again (5 of the 10 cards that came back were children born within minutes of the wave, the other 5 were folded parents that resumed and asked again). Every Cancel on the board (a card's ✕, the (i) panel's, a lane's cancel-all) now sends ONE cancel per task root — the topmost parent below an automation controller — and lets the runner take the subtree down; the confirm says so: Cancel this task? Its parents and everything working under them stop too. - Board actions on one line (operator 2026-10-15): Open in Code and Cancel are kit icon buttons on the footer row of EVERY ongoing card (Needs you, Working, Waiting on…), next to copy and (i); a Needs-you card's act zone keeps its one primary move.
- Board polish (operator 2026-10-14 11:33–11:41): lanes are always open (the fold chevron is gone) and their counts are readable pills; copy affordances appear only on hover or focus and are legible when they do; a Needs-you card reads asks since Mar 29, 19:47 (198d 17h) (or needs approval for web_search since …, paused since …) with the request quoted under it, names the person when the wire names one and nobody otherwise, and offers Review & answer, Open in Code (deep link
#session=…&run=…into the AbstractCode web app when the gateway's app registry says it is installed and running; greyed out with an install hint otherwise) and Cancel with an inline two-step confirm; plain-run cards no longer print "session"; automation cards carry a trigger chip (every 24h / on email / manual) instead of "clock"; the (i) panel is a sectioned grid (Run · Spend · Where, or Automation · Schedule · Policy · After · Where) with ids in monospace, sources as tags, long text collapsed by size, copy on every row and the actions at the bottom; Done cards show the outcome glimpse and 24m · 99,921 tk · 11 tools · 3 artifacts (answer and artifact count fetched once per terminal run). - Titles come back: the cached brief of a run (its title and excerpt) is keyed anew, so a title that was once cached empty is fetched again instead of reading untitled for good.
src/ui/state_line.ts: one vocabulary for what a run is doing (sentence, lane, subject, title order, pill word, automation brief and pulse folds), shared by the board and the Agent view and tested per wait kind × lane.
Changed
-
Board: hovering (or clicking) a card's answer or request glimpse opens the FULL text beside the card in an elevated panel — a header (Answer / Request, whose card, words · chars, copy, close), a reading-measure body rendered with the shared markdown renderer (panel-chat
Markdown: headings, lists, quotes, code, tables at the board's type scale), and an "Open the run" footer — never a native tooltip of raw markdown (operator 2026-10-15 21:44, 23:54). The glimpse itself is unchanged: three clamped lines. The (i) brief shares the same elevated shell (depth, hairline highlight, backdrop blur, a 140 ms entrance that respects reduced motion). The "data Ns old · Refresh" line is gone (21:48); the board polls by itself, and only a one-sentence notice shows when the gateway is unreachable or before the first poll. -
Board layout rework (operator 2026-10-15 20:41): the stats strip is gone (its counts live on the lanes); data age and Refresh sit on one slim line above the lanes and the "no progress after" select in the Working lane's head; every card is one vertical flow whose footer — the always-visible kit icon buttons (automation: open · run now · pause/resume · info; run: open in Code · cancel · info) — is the last child inside the card, so the (i) can no longer fall outside it; only the copy icon appears on hover; a Scheduled card's lines wrap instead of clipping; Done cards carry the outcome glimpse with their metrics on the footer. Cards are memoized: an unchanged poll re-renders 0 cards (was 114 for 57 cards), commit time ~2 ms (was ~8.5 ms).
-
Loading feedback everywhere (operator 2026-10-15): one
LoadingState(spinner + "Loading …",role=status, still under reduced motion) shows wherever a page has asked the gateway and has nothing yet — the Automations list and detail, the Observe run navigator, the System Artifacts, Logs and Activity panes — instead of an empty list, a "0" or a "nothing here" that is not yet true. -
Automations page: a card's controls (Active, Run now, Edit, Discuss, Archive) sit on ONE line; the page frame's dark margins shrink (10 × 12 px instead of 16 × 18 twice) to give the panels the space back.
-
Launch is a modal, not a page (operator 2026-10-14): + New automation on the Automations page opens it in Automate mode, the header's Launch button in Run once mode, and
#launch/#launch/automatedeep links open it over whatever page is showing. The Launch entry leaves the sidebar; the form, its modes and its fields are unchanged and scroll inside the modal. Creating an automation closes the modal on the Automations page; starting a run lands on the run. -
The board's poll carries no metrics and no ledger folds:
status=running+status=waitingevery 5 s; terminal rows (with metrics) come from a 60 s background sweep that never delays the tick. Automation records, task excerpts, last-record pulses (only for runs older than a minute) and wait details are fetched lazily through one pool (3 in flight) and cached for the session across page switches. The Done lane says "loading…" until the first sweep lands. -
The wait modal no longer covers the Agent view when a waiting run opens from the board; the header's Answer leads to the Ledger view's controls.
-
Listing rows keep
is_draftandpause_reason.
Fixed
- "Fetch is aborted" banners inside a run (operator 2026-10-16 01:59: "apparently there are some errors and I have no easy way to check it"; adversary
untracked/redesign/ABORT-AND-SPEED-adversary.md). Those three banners were the app's own 30 s listing deadlines, worded by WebKit exactly like our own aborts and rendered among the run's steps. Every gateway request now goes through one helper that tells the two apart: a deadline is a GatewayTimeout ("GET /api/gateway/runs?status=running: no answer within 30 s — the gateway is busy"), our own abort (a newer request replaced this one, the view went away) is Superseded and is never reported. Real failures are coalesced into ONE connection notice ("Refresh runs failed: … · 02:01 ×2") beside the connection label and on the board's status line, cleared by the next success; the run's Ledger shows the run's steps only. - Run header counts failed steps of every kind ("6 LLM calls · 199,088 tokens · 19 tool calls · 2 failed steps"); before, only tool failures were summed, so a failed LLM or subworkflow step was invisible there.
- The board paints the last known rows at once (operator 2026-10-16 02:01: "the observer is extremely slow to load"; measured: the two active listings cost 4–5 s with one run and an idle GPU, minutes under load, and a reload started from nothing). The runs poll keeps a persisted snapshot per gateway and account; a reload paints it immediately as board as of 02:00:41 — refreshing… and the first real poll replaces it. One listing that lands is applied even when the other fails (a notice, not a blank board), and the archived-roots scan is skipped when the gateway says there are no archived conversations.
- Board: a parent parked on a subrun that the listing no longer has read waiting for a subrun to finish and nothing more (operator 2026-10-16 01:31: "waiting on what? from what?"). The board now reads the wait detail of every such parent (it names the subrun) and looks the subrun up once: waiting for subrun coding-agent 0.2.4 to finish while it lives beyond the listing's cap, subrun coding-agent 0.2.4 completed 12s ago · resuming (Working) within the minute the runner's repair pass gets, and stuck: subrun coding-agent 0.2.4 ended cancelled 3h ago and the gateway could not resume this run · 77d 17h past it (or stuck: subrun 028f89e5 no longer exists… when the gateway answers 404) — a stuck run is a Needs you card with a stuck pill, sorted after today's asks, its parents folded in as "blocks", and Cancel task as its one move. The ✕ on every card now says what it reaches: Cancel this task: this run, its parents below the automation, and everything working under them.
- Board: a run whose flow the gateway can no longer name (
pin_ids: [], the bundle version is gone) showed untitled with its prompt sitting in the response; the prompt titles the card again. - Board: cancelling many runs was extremely slow (operator 2026-10-16 01:04: 53 cancels took 6 minutes while the gateway accepted each command in 41 ms) — every cancel fired three forced listing refreshes (each listing stats ~10,000 run files in the runner's own process), the commands queued behind ~160 listings, and the runner's apply latency rose from 0.03 s to 10 s. A batch cancel path (
on_cancel_runs) now sends one command per task root, four at a time, with ONE refresh for the batch and one more 2.4 s later. - Run view: an LLM call made on the runtime's defaults showed no provider or model on its card — they are named only in the request as sent (
result.metadata._provider_request);llm_detailnow reads them there. - Run view: a long outcome expanded with More grew past the viewport with no way to scroll back to Less — the run panel clips and its Ledger page was not a scrolling child. The Ledger page (and the new Agent view) now scroll inside the run panel; the toolbar and tabs stay put.
- Board: the Entities strip landed after the lanes had painted and pushed them down (operator 2026-10-15). It is now in the page at first paint, one reserved height in every state (
.mc_entities_reserved,flex: noneso the column flex never squeezes it): shimmering tile placeholders under its head until the roster and cards land (as many as the roster says, once known), the tiles after, and a one-line Entities — no entities head when the gateway has none. Measured in Chromium: the strip's height and the lanes' top are the same before and after the data. - Run view: a tool call that needed approval showed as waiting, forever, beside a second, completed copy of the same call. The runtime records an approved call's execution under a new step with the same idempotency key; the steps are now merged, so each call is one step with its real status (root backlog 1016 tracks reading step state from the runtime instead).