refactor: tlon-aligned agent architecture (UI core, abed peek/watch, linear migrations) by arthyn · Pull Request #4 · arthyn/notes

arthyn · 2026-05-01T02:57:31Z

Summary

Five focused structural refactors to align app/notes.hoon with tloncorp idioms. No semantic changes; backend tests stayed 53/53 green at every step.

03ccf8b Extract HTTP handling into ++serve-http; move PWA static assets (manifest, sw, favicon-svg, icon-svg) into a |% core in lib/notes-ui.hoon. New scripts/build-notes-ui.sh regenerates the ++index arm from app/notes-ui/index.html without touching the static asset arms.
c1a6080 Mass-rename type imports: /- notes → /- n=notes; all :notes refs in agent / lib / tests / mark files become :n.
8bc56de Per-notebook peek/watch via abed: ++no-peek, ++no-watch in no-core; ++se-watch in se-core. Top-level +peek collapses to a generic [%x %v0 kind=@ ship=@ name=@ rest=*] delegate. ++no-abed relaxed (works on any net); ?=(%sub -.net) guard pushed into write/sign arms.
90ceaac Wrap +poke in a |^ kelt — six poke-only handlers (join-remote, leave-remote, handle-{send,notify,accept,decline}-invite) nest inside +poke. Move 9 utility arms (slugify, strip-query, get-book, can-view-flag, etc.) to a contiguous block just before ++se-core.
82de44f Restructure +load to follow the homestead/groups pattern: |^ kelt with +$ any-state head-tagged union and one ++state-N-to-N+1 arm per version step (1-to-2, 2-to-3, ..., 9-to-10). Replaces the prior nested ?: cascade where states 1–7 jumped directly to state-8.

Test plan

mcp__sidwyn__run-tests desk=notes path=/tests/app → 53/53 OK at every commit
Playwright suite — 12/12 in isolation; 2 environmental flakes under sequential load (pre-existing, identified in prior session — the host ship slows down running 12 tests serially with 30s timeouts; same tests pass at 11.6s in isolation)
Manual UI sanity (refresh published note, invite flow, sub→host edit propagation) — recommend doing once the ships are quiet

🤖 Generated with Claude Code

… core Two related changes: 1. lib/notes-ui.hoon is now a |% core. The +index arm holds the HTML (regenerated from app/notes-ui/index.html via the new scripts/build-notes-ui.sh). +manifest, +service-worker, +favicon-svg, +icon-svg moved out of app/notes.hoon into this core. Agent imports as `/= ui /lib/notes-ui` and references arms as `index:ui`, `manifest:ui`, etc. 2. The %handle-http-request branch of +poke is lifted into a top-level ++serve-http arm (named to avoid shadowing the http namespace used by `response-header:http`). +poke now just delegates with a one-liner. Net effect on app/notes.hoon: -45 lines. scripts/build-notes-ui.sh splices index.html into the ++index arm body without touching the static asset arms, so manifest/sw/svg can be hand-edited independently. Triple-quote safety check kept. AGENTS.md (CLAUDE.md symlinks to it) updated for the new workflow and to switch the default dev ship reference from bospur to sidwyn. Tests: 53/53 backend. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

@ud

Mass-rename type references from :notes to :n by changing the import declaration in each consumer: /- notes → /- n=notes flag:notes → flag:n state-9:notes → state-9:n ... Affects desk/app/notes.hoon, desk/lib/notes-json.hoon, desk/tests/app/notes.hoon, and the four mark files under desk/mar/notes/. Two collision fixes in the test suite where local face names shadowed the import: - ++nb-flag took `n=@ud` for the count, masking the n alias for the ^- flag:n cast. Renamed the parameter to `nid`. - ++ex-history-len took both `nid=@ud n=@ud` for the same reason. Renamed the count parameter to `expected`. Tests: 53/53 backend. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Move per-notebook peek and watch logic into arms inside the per-notebook cores, so the top-level dispatchers shrink to flag-parse + delegate. The notebook-state, flag, members, etc. are loaded once at abed time and visible to all the inner arms. - ++no-peek (in no-core) — handles all 7 per-notebook scry paths (%notebook, %folders, %notes, %note/<id>, %note-history/<id>, %folder/<id>, %members). Top-level +peek catches them with a single [%x %v0 kind=@ ship=@ name=@ rest=*] arm. - ++no-watch (in no-core) — handles the local UI %stream subscription. - ++se-watch (in se-core) — handles remote-subscriber %updates watch (delegates to existing ++se-watch-sub). - ++no-abed relaxed: no longer asserts ?=(%sub -.net), so it works for any net type. The %sub assertion moved into ++no-action, ++no-start-watch, ++no-leave, ++no-agent — the write/sign arms that actually require subscriber semantics. ++no-response left without an outer guard because adding one breaks type narrowing on the %update branch. Cross-cutting peeks (%notebooks, %published, %invites, /x/ui, /x/debug/dummy) and watches (%inbox %stream) stay at the top level. Tests: 53/53 backend. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Two structural moves in the helper core, no logic changes: 1. Wrap +poke's body in a |^ kelt so the six poke-only handler arms (join-remote, leave-remote, handle-{send,notify,accept,decline}- invite) live inside +poke instead of polluting the outer subject. The kelt arms still reach the outer subject (state, cor, give-inbox-*, se-core, no-core) since |^ extends rather than replaces it. 2. Move the nine utility arms to a contiguous block just before ++se-core, in this order: slugify, a-notebook-to-c-notebook, get-book, strip-query, can-view-flag, find-flag-by-nid, notebooks-changed-card, give-inbox-received, give-inbox-removed. Net effect on +helper-core arm order: dispatchers (init, load, poke, serve-http, watch, peek, agent, arvo) up top, utilities below them, per-notebook cores at the bottom — so a reader sees the entry points before the helpers. Tests: 53/53 backend. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Restructure +load to follow the tloncorp/homestead groups.hoon shape: a |^ kelt with a +$ any-state head-tagged union and one ++state-N-to-N+1 arm per version step. Replaces the prior "jump to state-8" cascade where states 1–7 each migrated directly to state-8 in a deeply nested ?: chain. ++ load |^ |= =vase ^+ cor =+ !<(old=any-state vase) =? old ?=(%1 -.old) (state-1-to-2 old) =? old ?=(%2 -.old) (state-2-to-3 old) ... =? old ?=(%9 -.old) (state-9-to-10 old) ?> ?=(%10 -.old) =. state old cor +$ any-state $%(state-10:n state-9:n ... state-1:n) ++ state-1-to-2 |= s=state-1:n ^- state-2:n [...] ... -- Each step is small and focused (5–30 lines) with a `~> %spin.[...]` trace label. Cumulative 1 → 10 transformation matches the prior cascade. The legacy v0/v8 entity types and intermediate state types remain in sur/notes.hoon (used by the test suite to seed each starting version). Net: -39 lines in app/notes.hoon (-193 of the old cascade, +154 new). Tests: 53/53 backend. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Scry endpoints used to all return %json+!>(...). Now each returns a typed mark with both noun and json grow arms, so callers can scry the same path and get either the raw noun (e.g. via %notes-notebooks) or JSON (.json extension on Eyre, or %json mark coercion). 10 new mark files in mar/notes/: - notebook — wraps notebook-detail (flag + notebook record) - notebooks — list of notebook-summary (flag + notebook + visibility) - folder — wraps folder - folders — list of folder - note — wraps note - notes — list of note - note-history — list of note-revision - members — list of member-record (ship + role) - invites — list of invite-record (flag + invite-info) - published — list of published-record (flag + note-id + html) 5 listing wrapper types added to sur/notes.hoon (notebook-summary, notebook-detail, member-record, invite-record, published-record). JSON encoders for the new types added under enjs:notes-json (notebook-detail, notebook-summary, notebook-summaries, member-record, member-records, invite-record, invite-records, published-record, published-records). ++peek and ++no-peek now build the noun-typed value first, then return it via the typed mark — Eyre's grow:json handles JSON on the wire. "Not found" cases switched from json+!>(~) to outer-null `~`, the idiomatic absent-scry response. Test peek helpers updated: ++ex-mark asserts the new mark; ++peek-history binds (list note-revision:n) instead of (list json); ++ex-gone simplified to outer-null check. Tests: 53/53 backend. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

@ud

Audit pass: every =/ that just renames a wing (e.g. =/ nid=@ud id.nb-act) is reconsidered: - 12 multi-use + verbose aliases switched to =* (the proper aliasing rune): n-act, who, fid in se-rename-folder/se-move-folder/ se-delete-folder/se-create-note, nid in se-rename-note / se-move-note / se-delete-note / se-update-note / se-restore-note, nid-nb in se-batch-import-tree. - 15 aliases inlined: short or single-use (nb-act inlined since a-notebook.act is only 14 chars; u-nb in no-apply-update — 11 uses, but u-notebook.upd is 14 chars so still better inlined; title in handle-send-invite at 1 use; etc.). - 1 dead-code =/ deleted (fid in se-dispatch-folder, never referenced; the dispatch arm passes cmd wholesale). Computed values (function calls, literals) keep =/ — they're not pure aliases. The =/ =face:type source style (e.g. =/ =flag:n flag.act) also stays since the =face syntax is a face-mold cast, not aliasing. Net: -16 lines in app/notes.hoon. Tests: 53/53 backend. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Per the rule: use the natural type name as the face when the type is short or used in only 1-2 places; only deviate when a real shadowing collision forces it. Renames in app/notes.hoon (no logic changes): - new-nb / placeholder-nb / nb (×2) → notebook - nt (×12 sites in se-create-note, se-rename-note, se-move-note, se-update-note, se-delete-note, se-restore-note, batch-import arms, loop bodies, scry binding) → note - nf (folder, ×2) → folder - nf (flag, ×3 in state-9-to-10 migration) → flag - f=flag:n in se-abed / no-abed gate samples → flag Wing resolution (right-to-left dot lookup) means notebook.notebook-state still finds the .notebook field of notebook-state correctly even with a local face named notebook. The slight stutter in lines like se-core(flag flag, ...) is acceptable per the rule — readability of the type name wins over avoiding stuttery wing-replace syntax. Tests: 53/53 backend. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Hoon's =type syntax auto-creates a face matching the type name, so `=/ flag=flag:n ...` is just `=/ =flag:n ...`. Same for =notebook:n, =note:n, =folder:n in the spots where the previous pass renamed abbreviations to their natural type names. 19 sites updated. Tests: 53/53 backend. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Two style fixes across the helper-core arms: 1. ++se-rename-notebook had the read-mutate-write triplet: =/ =notebook:n notebook.notebook-state =. notebook notebook(title ..., updated-at ..., updated-by ...) =. notebook.notebook-state notebook Three lines for one in-place field update. Collapsed to a single =. notebook.notebook-state ...notebook.notebook-state(...) write using %_ for the multi-field block. No alias, no copy. 2. 21 arms had ?> ?=(...) before ^+ se-core, opposite of the convention (cast right after the gate sample, then assertions). Swapped order in se-create-notebook, se-rename-notebook, se-delete-notebook, se-set-visibility, se-invite, se-member-{join,leave}, se-dispatch-folder, se-dispatch-note, se-create-folder, se-{rename,move,delete}-folder dispatch arms, se-create-note, se-{rename,move,delete,update,restore}-note dispatch arms, plus a few more. Tests: 53/53 backend. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

@ud

…cells Two stylistic rules applied across the helper core: Rule 1 — when typing a =/ as a tuple, use * for parts not accessed at that site: - two %notes-action dispatch arms (poke %note branch + main): only -.net.entry is checked, so entry=[=net:n *] - handle-send-invite: only title.notebook.notebook-state.entry is read, so entry=[* =notebook-state:n] - the /v0/notebooks peek gate sample: only notebook-state half used - the find-flag-by-nid gate sample: only notebook-state half used - the state-9-to-10 xlat builder: only notebook-state half used Rule 2 — same-subject cells collapse to [a b c]:subject when the cell is the LAST element of the outer expression (right-associative noun shape splices through): - (slugify title.notebook.notebook-state id.notebook.notebook-state) → (slugify [title id]:notebook.notebook-state) - /v0/notebooks listing item: [flag notebook.notebook-state visibility. notebook-state] → [flag [notebook visibility]:notebook-state] - se-create-notebook constructions: [...our.bowl now.bowl now.bowl our.bowl] → [...[our now now our]:bowl] (notebook + root folder) - se-create-folder construction: bowl tail collapsed - se-batch-import-tree folder: bowl tail collapsed Note constructions (se-create-note, batch-import, batch-import-tree notes) tried-and-reverted: revision=@ud follows the bowl block, so the bowl-cell is NOT the last sub-expression and the right-associative splice doesn't apply. Left as-is. Tests: 53/53 backend. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

The %notes-action branch of +poke was a chain of ?: ?=(%X -.act) tests ending in a ?> %notebook fall-through, with another nested chain inside the %notebook case for %invite/%note/default-route, and a third nested chain inside %note for %publish/%unpublish/default-route. Restructured to: ?. ?=(%notebook -.act) :: switchable cases — %notebook is the meaty fall-through ?- -.act %create-notebook ... %join ... %leave ... %accept-invite ... %decline-invite ... == :: meaty %notebook body =/ =flag:n flag.act ?+ -.a-notebook.act :: default: route to host (se/no) based on net ... :: %invite (handle-send-invite ...) :: %note ?+ -.n-act :: default: route ... :: %publish :: local :: %unpublish :: local == == Default-routing logic appears in two places (outer ?+ default and inner %note ?+ default) — kept the duplication; small enough. Tests: 53/53 backend. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Follow the +go-a-group pattern from tloncorp/homestead/groups.hoon: the action handler converts to a c-notes command and pokes the host unconditionally. If host == our.bowl, Gall loops the poke back through +poke %notes-command and dispatches to se-core. Removes the host/sub discriminator from the action surface. Agent changes: - +poke %notes-action default cases (both outer ?+ and inner ?+ for %note) drop the `?: ?=(%pub -.net.entry)` branch. Each just calls no-abet:(no-action:(no-abed:no-core flag) act). - ++no-action drops the `?> ?=(%sub -.net)` assertion. Other no-* arms keep their %sub guards (those are genuinely sub-only). Test harness changes (lib/test-agent.hoon): - ++drain-self-pokes + ++drain-loop helpers detect any [%pass wire %agent [our.bowl <our-dap>] %poke =cage] cards in the emitted list and re-deliver them through on-poke. Bowl state is threaded via play-cards so wex/sup remain coherent. Recursion has a depth-16 cap. - ++do-poke and ++do-agent run the drain after the initial agent call, so test assertions on state see the fully-applied result, not just the queued self-poke card. Tests: 53/53 backend (no test changes needed — the drain is transparent). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

@tas

AGENTS.md still described state-4 (state-0..4 in sur, current state-4) from before the migration cascade was rebuilt. Updated the State and API Surface sections to match what's actually deployed: - Current state is state-10, with shape books/next-id/published/invites - flag is [=ship name=@tas] with slugified names - notebook-state contains notebook/members/visibility/folders/notes/history - Migrations follow the tlon |^ kelt + linear =? chain pattern - Scry endpoints now return typed marks (notes-notebooks, notes-note, etc.) with both noun and json grow arms - Action surface routes through no-action which always pokes the host (Gall loops self-pokes back through %notes-command); only %publish / %unpublish stay local - HTTP routes lifted into ++serve-http - Watch arms split: ++no-watch for local UI, ++se-watch for remote subs For /v0/published, the published-record type carried html=@t but the encoder never emitted it — the listing has only ever returned host / flagName / noteId pairs in JSON (which is what the FE consumes). Dropping html from the type so the noun-form mark also reflects the documented shape; saves the wasted bandwidth on the noun-encoded peek. Tests: 53/53 backend. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

The %publish / %unpublish action handlers used to mutate published.state directly with no checks — any local poke could publish HTML keyed under any flag, including subscribed remote notebooks. Tightened to: - new ++publish-note / ++unpublish-note arms in cor (since published is in outer state, not notebook-state). Each arm: ?> =(src.bowl our.bowl) — local-action only ?> =(ship.flag our.bowl) — host-only (only the host's ship maintains its public URL) abed se-core to get notebook context ?> (se-can-edit:se src.bowl) — edit permission publish also asserts the note exists - action handler delegates: (publish-note flag nid html) / (unpublish-note flag nid). New tests: - test-publish-note-rejects-non-host (~bus tries to publish under ~zod's flag → crash) - test-publish-note-rejects-missing-note (publish a non-existent note id → crash) Tests: 55/55 backend. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Inbox events were the last surface emitting raw json+!>(...) cards; every other client surface now uses typed marks. Consistency fix. - New +$ u-inbox tagged union in sur/notes.hoon — three variants: %invite-received, %invite-removed, %notebooks-changed. - ++u-inbox encoder under enjs:notes-json reproduces the existing {type:'update', update:<payload>} wire shape so the FE's applyNotebookUpdate / applyInboxEvent dispatch keeps working. - New mar/notes/inbox-update.hoon with grad %noun and grow noun+json. - ++give-inbox-received / ++give-inbox-removed / ++notebooks-changed-card emit notes-inbox-update+!>(<u-inbox value>) instead of building the JSON inline. Tests: 55/55 backend. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Master added several functional improvements between when this branch forked and now. Reconciled them onto the refactored layout: - v0.10.1 / v0.11.0 UI work (drafts in localStorage, faster autosave, preview shows title, list/task editing UX, checkable preview checkboxes, suppress join modal on already-joined notebooks): merged via auto-merge of index.html; lib/notes-ui.hoon regenerated via scripts/build-notes-ui.sh. - leave-remote: emit %member-leave to host before unsubscribing (so the host's members.notebook-state reflects reality). Ported into the |^ kelt arm. - /v0/notebook scry: now includes visibility. Updated notebook-detail type in sur and the encoder; the no-peek %notebook arm builds [flag notebook visibility]. - +agent: handle [%notes %leave ship name ~] poke-ack as a no-op (best-effort). - +arvo: handle [%behn %wake] for [%notes %rewatch ship name ~] — re-arms a subscription after a transient failure. - ++no-agent watch-ack failure: instead of silently dropping the sub, schedule a 30s rewatch via behn timer. - Syncer (Tauri menubar app) v0.2.0 alignment with state-10 agent. Tests: 55/55 backend. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

@ud

…ioning The previous publish-note / unpublish-note arms in cor were checking host ownership and edit permission via se-core — too much. The action poke layer already enforces ?> =(our.bowl src.bowl), and no-abed already validates that the notebook exists. Beyond that, publish is just a per-ship cache write (the local /notes/pub/<flag>/<id> URL serves whatever this ship caches), not an authority assertion that needs the host's edit-role gate. Replaced with two arms in no-core: ++ no-publish |= [nid=@ud html=@t] :: write published.state ++ no-unpublish |= nid=@ud :: del Action handler dispatches via no-abet:(no-publish:(no-abed:no-core flag) ...). no-abed crashes if the flag isn't in books; the action poke layer crashes on non-self src. That's the full check set. Tests: - test-publish-note-rejects-non-host renamed to *-rejects-non-self (the rejection is the action-poke self-check, not a ship-flag check). - test-publish-note-rejects-missing-note replaced with test-publish-note-rejects-unknown-notebook (no-abed rejects unknown flags; we don't validate that the note-id exists within the notebook anymore — published is a key-value cache). Tests: 55/55 backend. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

The agent-door header has =| current-state followed by =* state -, so the agent state value sits at the head of the subject. Wing resolution walks into it for free, which means books, next-id, published, invites each resolve directly without the .state suffix. 52 sites updated across the helper core: books.state → books, etc. The local-face .s in migration arms (state-N-to-N+1) is unrelated and stays — that's the migration input, not the agent state. Tests: 55/55 backend. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

The host's self-poke from the action handler comes back as %poke-ack on the [%notes %sub ship name ~] wire — same wire used by genuine subscriber agent signs. The outer ?> ?=(%sub -.net) at the top of no-agent crashed the host because its net is %pub. Move the sub-only check into the per-sign-type branches that actually need it: - %fact: route to no-response unconditionally (no-response itself already short-circuits when the host emits to itself, since we don't watch our own /updates wire). - %kick: only re-watch if we're a sub; host doesn't watch self. - %watch-ack: only re-arm the rewatch timer if sub. - %poke-ack (default): always no-op. Surfaced by an action poke crashing on sidwyn with `take %poke-ack failed at no-agent` — host's self-poke ack was tripping the assertion. Tests: 55/55 backend. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Two FE bugs surfaced while exercising the e2e flows: 1. ++triggerAutoCreate captured the user's pending title/body once early (right after the create-note poke) and restored them after several awaits (loadNotes, selectNote). Any keystrokes typed during that ~150-400ms window were silently dropped on restore — symptom: typing a title quickly then switching to body sometimes "snapshotted" an empty body. Fix: capture pending values just before selectNote. 2. channelId was initialized to "" and only set inside ++openChannel, which runs from ++connect. Fast user clicks (or e2e tests) firing a poke before connect() finished produced PUT /~/channel/ (empty id) → 400. Fix: initialize channelId at module load with a Date.now() id; the channel comes into being on the first PUT, so the name is just a label we pick eagerly. Bumped ++dummy to v0.11.0-refactor. Tests: 55/55 backend. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Same race shape as the channelId one — SHIP is fetched via /~/name during ++connect, and pokes that fire before that resolves carry ship:"" → Eyre 400. - Expose window.__notesGetShip() returning the current SHIP value once /~/name has resolved. - Test fixture (notes + openSubscriberContext) waits up to 15s for the probe to return a non-empty ship before yielding the page to the test body. Tests: 55/55 backend. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

When a cross-ship-invite test fails before the sub accepts/declines, the entry stays in the sub's invites.state and accumulates across runs. Same for cross-ship-decline-invite if it fails before the decline. - wipe-leftovers: scry /v0/invites for each ship; if any e2e- titled invite is present, navigate the page to /notes/ and call declineInvite() via the FE. - tryDelete (per-test cleanup): sweep matching invites for the title before attempting the notebook deletion. Best-effort, errors swallowed. Tests: 55/55 backend. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Ergonomics for diagnosing flaky cross-ship behavior — when a sub edit doesn't appear on the host, you usually want to know what SSE events the FE actually received. Adds: - handleEvent gates a `console.log("[sse]", JSON.stringify(data))` on localStorage.e2e-log-sse === "1". Off by default (no-op for users). - Test fixture sets the localStorage flag on every page (host + sub), and pipes browser console / pageerror through to the test runner — but only when E2E_TRACE=1 is set, so green runs stay quiet. Usage: `E2E_TRACE=1 npm run test:e2e -- --grep 'cross-ship edit'` Tests: 55/55 backend. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Playwright UI's Console tab auto-captures browser console.log output per action — but only if the FE actually logs. Gating the localStorage flag on E2E_TRACE=1 meant UI mode showed nothing. Always set the flag now; E2E_TRACE=1 still gates the runner-side pipe (attachConsole) for CLI runs. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

handleEvent only sees fact/snapshot/inbox messages — poke-ack short-circuits before it. Move the trace log up to eventSource.onmessage so every raw SSE payload (including poke-acks) is captured. Both layers now log when localStorage.e2e-log-sse === "1". Bumped ++dummy to force a reload. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

forceSave's "Saved" indicator only means the FE's local poke-ack came back — for cross-ship subscribers, the host's se-update and the fact's trip back through the stream subscription happen AFTER that. Without a settle, cross-ship-edit asserts on the host's editor before the update event has been delivered. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Host was already in the notebook from the create+invite steps and is subscribed to its /stream — the sub's create-note should propagate via the existing subscription. The page.goto + selectNotebook re-navigation masked any real propagation bug in the original test. If this fails now, the failure is the actual stream-propagation bug we want to surface, not a workaround. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Two more rules from arthyn/notes#4 follow-up commits (68e0e1a, b6b8d48). architecture.md - New "Always poke the host" subsection in ACUR: the action handler emits a poke to ship.flag unconditionally; Gall loops self-pokes back through +poke %notes-command. Removes the host/sub branch from the action layer entirely. c-command processing becomes the single dispatch point for all state-changing logic. - New "State fields resolve without .state" callout under standard boilerplate: =| / =* state - puts state at subject head, so wing search finds books, invites, etc. directly. Prefer the unprefixed form except where local shadowing forces it. patterns.md - Updated existing examples (Tuple Types with *, Per-Entity Engines) to use books / cor(books ...) instead of books.state / cor(books.state ...) for consistency with the new rule. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Marks the refactor cycle on PR #4: tlon-aligned dispatch (action → no-action → command → host), typed marks for every scry endpoint, linear migrations, ?- / ?+ dispatch over chained ?:, =* aliasing, =type face shortcut, no-publish/no-unpublish through the engine, inbox-update mark, and a handful of e2e harness fixes (drain self pokes, settle after forceSave, eager channelId, wait for SHIP). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

arthyn and others added 12 commits April 30, 2026 20:13

arthyn mentioned this pull request May 5, 2026

Apply lessons from arthyn/notes PRs #3 and #4 tloncorp/hoon-reference#2

Open

3 tasks

arthyn and others added 12 commits May 5, 2026 16:13

arthyn force-pushed the refactor-peek-watch-via-abed branch from cdba5ff to 4a76627 Compare May 6, 2026 21:49

arthyn and others added 4 commits May 6, 2026 16:53

arthyn merged commit 47e6476 into master May 7, 2026
1 check passed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

refactor: tlon-aligned agent architecture (UI core, abed peek/watch, linear migrations)#4

refactor: tlon-aligned agent architecture (UI core, abed peek/watch, linear migrations)#4
arthyn merged 29 commits into
masterfrom
refactor-peek-watch-via-abed

arthyn commented May 1, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Conversation

arthyn commented May 1, 2026

Summary

Test plan

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant