feat(desktop): restore observer-feed regressions from #1381 and classify 4 new session/update types by wpfleger96 · Pull Request #1412 · block/buzz

wpfleger96 · 2026-06-30T23:28:40Z

What

Restores observer-feed display regressions introduced in #1381, classifies 4 new session/update types, and fixes system-prompt display ordering.

1. Restored observer-feed behaviors (6 items)

Regressions from #1381 that collapsed the observer transcript to a single undifferentiated stream:

System prompt — session/new systemPrompt sections rendered as collapsible "System prompt" block
Prompt context — session/prompt context sections rendered as collapsible "Prompt context" block
User message — session/prompt user text rendered as a message bubble with channel link
Steering — _goose/unstable/session/steer user text rendered as a steering message
Tool results — tool results correctly parsed and attached to their call items
Usage coalescing — current_usage_update accumulated across a turn into a single usage item

2. New lifecycle types (4 items)

Types present in the ACP stream since #1381 but not previously classified:

ACP type	Displayed as
`current_mode_update`	"Mode: `<mode>`" lifecycle row
`usage_update`	Usage lifecycle row (coalesces within turn)
`available_commands_update`	"Commands available" lifecycle row
`config_option_update`	"Config: `<key>` = `<value>`" lifecycle row

3. System-prompt display ordering

Bug: The observer feed rendered Prompt context before System prompt on the first turn.

Root cause: pool.rs emits turn_started before session/new, so turn_started creates the turn bucket before the system-prompt item is produced. A simple turnId: null approach appends the system-prompt single to displayOrder after the turn block, placing it after Prompt context.

Fix: System-prompt items are tagged acpSource: "session/new". buildTranscriptDisplayBlocks holds them as a pendingSystemPrompt buffer and injects them into the first following turn's classifyTurnItems call as externalSystemPrompt. The prompt segment renders: user message → System prompt → Prompt context. If no session/prompt follows (incomplete stream or isolated test), the system-prompt falls back to a standalone single.

The PromptContextInline renderer uses context.title (not hard-coded "Prompt context") so both blocks display under their correct headings.

Display order per turn:

<user message bubble>
<System prompt sections>   ← from session/new, heading: "System prompt"
<Prompt context sections>  ← from session/prompt, heading: "Prompt context"

Tests

11 E2E screenshot specs (shots 01–10: individual restored behaviors; shot 11: full first-turn ordering with the real pool.rs wire sequence)
2 ordering regression tests using the full realistic wire sequence (turn_started → session/new → session_resolved → session/prompt); both fail on revert (1473 pass / 2 fail), pass on fix (1475/1475)
All existing grouping unit tests pass

…sions Adds a Playwright smoke spec that seeds observer relay events directly into the production appendAgentEvent → processTranscriptEvent ingestion path via a new __BUZZ_E2E_SEED_OBSERVER_EVENTS__ e2e bridge hook, then screenshots four surfaces restored by PR #1412: - 01-prompt-context-inline: collapsed inline prompt-context block - 02-system-prompt-title: system prompt with title visible in feed - 03-permission-approved: permission row with Approved (allow_once) outcome - 04-permission-cancelled: permission row with Cancelled outcome Supporting changes: - injectObserverEventsForE2E exported from observerRelayStore.ts - __BUZZ_E2E_SEED_OBSERVER_EVENTS__ Window hook wired in e2eBridge.ts - spec added to smoke project in playwright.config.ts Co-authored-by: Will Pfleger <pfleger.will@gmail.com> Signed-off-by: Will Pfleger <pfleger.will@gmail.com>

wpfleger96 · 2026-07-01T20:16:24Z

Shots 01–10: restored behaviors and new lifecycle types

Screenshots 01–10 show the individual behaviors restored from the #1381 regression (prompt context, system prompt, permission outcomes, prompt sections expanded, lifecycle status lines). These are unchanged from the previous post — included for completeness.

Shot 11: first-turn ordering fix (new)

Full pool.rs wire sequence (turn_started → session/new → session_resolved → session/prompt). Shows the corrected per-turn display order: user message bubble → System prompt (Base/System sections) → Prompt context (Buzz event/Thread context sections).

…sions Adds a Playwright smoke spec that seeds observer relay events directly into the production appendAgentEvent → processTranscriptEvent ingestion path via a new __BUZZ_E2E_SEED_OBSERVER_EVENTS__ e2e bridge hook, then screenshots four surfaces restored by PR #1412: - 01-prompt-context-inline: collapsed inline prompt-context block - 02-system-prompt-title: system prompt with title visible in feed - 03-permission-approved: permission row with Approved (allow_once) outcome - 04-permission-cancelled: permission row with Cancelled outcome Supporting changes: - injectObserverEventsForE2E exported from observerRelayStore.ts - __BUZZ_E2E_SEED_OBSERVER_EVENTS__ Window hook wired in e2eBridge.ts - spec added to smoke project in playwright.config.ts Co-authored-by: Will Pfleger <pfleger.will@gmail.com> Signed-off-by: Will Pfleger <pfleger.will@gmail.com>

Three regressions introduced by #1381's activity-feed rebuild: 1. Per-turn prompt context sections are now visible inline in the feed body. Previously they were hidden behind a modal accessible only via a small CheckCheck toggle inside the user-message bubble footer. Now the sections render directly below the user bubble with each section title visible and body expandable on click. A "View full" button opens the scrollable modal for focused reading. 2. Metadata items (system prompt, prompt context) now display their semantic title instead of the anonymous "Captured N raw sections" label. RawRailActivity now uses item.title as the verb, so "System prompt" and "Prompt context" render with their real names. 3. Permission rows now show the decided outcome. The permission response arrives as an acp_write event with result.outcome and the same JSON-RPC id as the request. A pendingPermissions map on TranscriptState indexes request id → lifecycle item id + option kind map. On the response the outcome is appended: Approved (allow_once) / Denied (reject_once) / Cancelled. Also restores metadata participation in isMeaningfulItem() for the Now summary bar. Raw JSON-RPC frames remain excluded (acpSource=raw_json_rpc); all semantic metadata (system prompt, prompt context) now contribute again, matching pre-#1381 behavior. System-prompt key (system-prompt:${ch}) is intentionally unchanged — the frame predates session creation and correctly reflects current channel setup by replacing on each session/new. Co-authored-by: Will Pfleger <pfleger.will@gmail.com> Signed-off-by: Will Pfleger <pfleger.will@gmail.com>

…lation JSON-RPC 2.0 allows both string and numeric ids. The ACP runtime preserves permission request ids as serde_json::Value for exactly this reason. The prior commit used asString(payload.id) for both the request index and response lookup in pendingPermissions, silently dropping any numeric id and leaving the outcome unattached. Add a jsonRpcId() helper that accepts string or finite-number values and keys them via JSON.stringify (preventing collisions between the number 1 and the string "1"). Use it at both the request-index site and the response-lookup site instead of asString. Regression tests: numeric id with selected allow_once, numeric id with cancelled, and a collision test verifying number 1 and string "1" attach to their respective items independently. Co-authored-by: Will Pfleger <pfleger.will@gmail.com> Signed-off-by: Will Pfleger <pfleger.will@gmail.com>

Inside the acp_write response branch, `responseId` is typed `string | null`. The `if (pending && outcomeKind)` guard only narrowed `pending`, not `responseId`, so `d.pendingPermissions.delete(responseId)` was rejected by tsc (argument of type 'string | null' not assignable to 'string'). Add `&& responseId` to the guard. This is semantically a no-op — `pending` can only be truthy when the earlier `responseId ?` branch took the non-null path — but TS requires the explicit narrowing. Co-authored-by: Will Pfleger <pfleger.will@gmail.com> Signed-off-by: Will Pfleger <pfleger.will@gmail.com>

…sions Adds a Playwright smoke spec that seeds observer relay events directly into the production appendAgentEvent → processTranscriptEvent ingestion path via a new __BUZZ_E2E_SEED_OBSERVER_EVENTS__ e2e bridge hook, then screenshots four surfaces restored by PR #1412: - 01-prompt-context-inline: collapsed inline prompt-context block - 02-system-prompt-title: system prompt with title visible in feed - 03-permission-approved: permission row with Approved (allow_once) outcome - 04-permission-cancelled: permission row with Cancelled outcome Supporting changes: - injectObserverEventsForE2E exported from observerRelayStore.ts - __BUZZ_E2E_SEED_OBSERVER_EVENTS__ Window hook wired in e2eBridge.ts - spec added to smoke project in playwright.config.ts Co-authored-by: Will Pfleger <pfleger.will@gmail.com> Signed-off-by: Will Pfleger <pfleger.will@gmail.com>

Store the resolved permission outcome on a new `outcome?` field of the lifecycle item type instead of appending it to `text` via joinLifecycleText. This separates the data model cleanly: `text` holds request detail + options; `outcome` holds the final decision. LifecycleActivity.tsx renders permission items in three visually distinct parts: 1. Request row: shield icon + title + request description 2. Options sub-line: muted indented 'Options: ...' text 3. Decision row (when resolved): divider + colored icon + outcome label - green + CheckCircle2 for 'Approved (...)' - destructive + XCircle for 'Denied (...)' - muted + XCircle for 'Cancelled' Also adds expanded E2E screenshot variants: - 05-prompt-context-expanded.png: all three accordion sections open - 06-system-prompt-expanded.png: outer <details> + inner section <details> open Unit tests updated to assert outcome is on item.outcome (not in item.text). Co-authored-by: Will Pfleger <pfleger.will@gmail.com> Signed-off-by: Will Pfleger <pfleger.will@gmail.com>

Adds isSpineItem() predicate — true for tools, messages, thoughts, plans, and meaningful lifecycle events; false for metadata items (system prompt, prompt context). These are 'reads' that should recede when real work is present, per VISION_ACTIVITY.md:49,61 (failures rise; reads recede; suppression is what makes signal legible). BotActivityBar now uses a two-tier headline scan: first pass collects spine-only headlines; if none found (session start / idle), falls back to all meaningful items via isMeaningfulItem so the bar isn't left empty. Feed visibility for metadata items is unaffected — AgentSessionTranscriptList renders them independently of this predicate. Also fixes the isMeaningfulItem docstring, which incorrectly claimed the function feeds a 'Now summary' (it only feeds the headline scan). Co-authored-by: Will Pfleger <pfleger.will@gmail.com> Signed-off-by: Will Pfleger <pfleger.will@gmail.com>

…dion RawRailActivity now branches on acpSource: raw_json_rpc items keep the existing <pre>/details raw treatment (the ambient safety net), while all other named metadata items (system prompt, steer-turn prompt context) render via the shared PromptSectionList accordion — the same polished rounded-2xl card used for per-turn prompt context. Extract PromptContextSections + PromptContextSectionAccordion from AgentSessionTranscriptList into PromptSectionAccordion.tsx and import the shared component back in both AgentSessionTranscriptList and RawRailActivity. No visual change to prompt-context; the raw-rail treatment is now reserved for genuine raw payloads only. Co-authored-by: Will Pfleger <pfleger.will@gmail.com> Signed-off-by: Will Pfleger <pfleger.will@gmail.com>

…x shot 06 Replace the three acpSource boolean tests in agentSessionTranscriptPresentation.test.mjs (which only asserted item.acpSource === 'raw_json_rpc' on hand-built objects — they would pass with the isRawPayload branch deleted) with proper renderToStaticMarkup render tests in RawRailActivity.render.test.mjs. The new tests render RawRailActivity for three metadata items and assert the HTML: raw_json_rpc includes <pre> and no rounded-2xl; system prompt and steer-turn prompt context include rounded-2xl and no <pre>. Fix shot 06 in observer-feed-screenshots.spec.ts: after the render unification, system-prompt sections are React button accordions inside a native ActivityRow <details>. Shot 06 now opens the outer <details> first, then clicks each inner section button to expand it — matching the interaction pattern in shot 05. Shot 06 confirmed to show both Base and System sections fully expanded in the polished card UI. Co-authored-by: Will Pfleger <pfleger.will@gmail.com> Signed-off-by: Will Pfleger <pfleger.will@gmail.com>

Add count assertion before the section-button click loop in shot 06 so a future testid/nesting break fails loudly instead of silently producing a collapsed screenshot. Swap the render-test accordion marker from the style token 'rounded-2xl' to 'data-testid="transcript-prompt-context-sections"' — the semantic shared-list marker emitted by PromptSectionList. Style tokens can drift without breaking behavior; the testid is stable and specific. Co-authored-by: Will Pfleger <pfleger.will@gmail.com> Signed-off-by: Will Pfleger <pfleger.will@gmail.com>

…lable_commands_update, config_option_update Add 4 named session/update classifier cases before the existing else-drop safety net in agentSessionTranscript.ts: - current_mode_update: lifecycle 'Mode / {currentModeId}'; suppressed when currentModeId is absent. - usage_update: lifecycle 'Usage / Tokens: {used}/{size}' (+ cost if present); uses replaceLifecycleItem so repeated frames per turn coalesce to the latest value rather than accumulating. - available_commands_update: lifecycle 'Commands / Commands available: N'. - config_option_update: lifecycle 'Config / {name} = {val}, ...' joined from configOptions array; suppressed when array is empty. The else branch is left unchanged — unknown/future types still drop, preserving the firehose safety net. keepalive stays dropped. Add replaceLifecycleItem helper (parallel to upsertLifecycleItem but replaces text on update instead of appending via joinLifecycleText). Tests: 12 new unit tests covering each case, edge paths (missing fields, zero-length arrays), usage coalescing (3 frames → 1 item with last value), keepalive drop, and unknown-type drop. File-size override bumped to 1162 with a narrowly scoped justification note. Co-authored-by: Will Pfleger <pfleger.will@gmail.com> Signed-off-by: Will Pfleger <pfleger.will@gmail.com>

…pdate types Co-authored-by: Will Pfleger <pfleger.will@gmail.com> Signed-off-by: Will Pfleger <pfleger.will@gmail.com>

…re per-turn prompt-context The session/new arm in processTranscriptEvent was passing ctx (which includes the active turnId) to upsertMetadata for the system-prompt item. When session/new and session/prompt share the same turnId — as the harness always does on the first turn — the system-prompt metadata item landed in the same turn bucket as the prompt-context item. buildTranscriptDisplayBlocks / classifyTurnItems treats unrecognized turn-bucket items as activity, which renders after the prompt segment (user message + setup + prompt-context). Result: system-prompt appeared below prompt-context in the feed. Fix: pass { ...ctx, turnId: null } so the item is always treated as a standalone single entry by the display grouper. This is semantically correct — the system prompt is per-channel, not per-turn. The upsert update path handles null via ctx.turnId ?? existing.turnId, so re-created sessions also keep turnId=null. Two new tests added to agentSessionTranscript.test.mjs covering: - First-turn shared-turnId ordering (the bug case) - Multi-turn ordering (system-prompt stays before all prompt-context items) Co-authored-by: Will Pfleger <pfleger.will@gmail.com> Signed-off-by: Will Pfleger <pfleger.will@gmail.com>

The two system-prompt ordering tests were asserting on raw buildTranscript output (d.items insertion order). Since session/new always fires before session/prompt on the wire, system-prompt was always at a lower index in d.items regardless of the fix — making the ordering assertions tautological. Fix: route through buildTranscriptDisplayBlocks + flattenDisplayBlocks, the layer that actually contained the bug (classifyTurnItems placed system-prompt as an 'activity' item after prompt-context when both shared the same turnId). The turnId=null assertions remain on raw items — they verify the fix input that makes the display grouper behave correctly. Verified: both tests fail when the { ...ctx, turnId: null } one-liner is reverted, pass when it is present. Co-authored-by: Will Pfleger <pfleger.will@gmail.com> Signed-off-by: Will Pfleger <pfleger.will@gmail.com>

…isplay order pool.rs emits turn_started before session/new on the first turn, so turn_started creates the turn bucket before session/new fires. With turnId=null, system-prompt lands as a standalone single AFTER the turn block in displayOrder — rendering after prompt-context. Fix: tag system-prompt items with acpSource "session/new". In buildTranscriptDisplayBlocks, hold system-prompt singles until the first turn with a user-prompt item, then pass them to classifyTurnItems as externalSystemPrompt. classifyTurnItems slots the item into the prompt segment between the user message and prompt-context. The flattenDisplayBlocks order is: user → systemPrompt → context. Tests updated to use the full realistic wire sequence (turn_started → session/new → session_resolved → session/prompt). Fails-on-revert confirmed: removing acpSource "session/new" makes both ordering tests fail (1473 pass / 2 fail); restored → 1475/1475. Co-authored-by: Will Pfleger <pfleger.will@gmail.com> Signed-off-by: Will Pfleger <pfleger.will@gmail.com>

PromptContextInline hard-coded "Prompt context" for both the inline heading and the dialog title, so the system-prompt metadata block (title: "System prompt") rendered with the wrong visible label. Use context.title in both places so system-prompt shows "System prompt" and prompt-context shows "Prompt context" as expected. Co-authored-by: Will Pfleger <pfleger.will@gmail.com> Signed-off-by: Will Pfleger <pfleger.will@gmail.com>

… session/new without session/prompt When session/new arrives without a following session/prompt (e.g. isolated test scenarios or incomplete streams), the pending system-prompt item was silently dropped. Add a fallback: if pendingSystemPrompt is unconsumed after processing all display-order entries, emit it as a standalone single. Add E2E spec shot 11 (first-turn ordering) showing the full pool.rs wire sequence: turn_started → session/new → session_resolved → session/prompt. Asserts user bubble → System prompt → Prompt context render order is visible in the observer feed. Update shot 02 comment to document that it exercises the fallback path. Co-authored-by: Will Pfleger <pfleger.will@gmail.com> Signed-off-by: Will Pfleger <pfleger.will@gmail.com>

wpfleger96 force-pushed the duncan/observer-feed-1381-restore branch from 787713f to ce9a03f Compare June 30, 2026 23:49

wpfleger96 pushed a commit that referenced this pull request Jul 1, 2026

screenshots: PR #1412

4e9a773

wpfleger96 pushed a commit that referenced this pull request Jul 1, 2026

screenshots: PR #1412

21c4af7

wpfleger96 pushed a commit that referenced this pull request Jul 1, 2026

screenshots: PR #1412

baaf806

wpfleger96 pushed a commit that referenced this pull request Jul 1, 2026

screenshots: PR #1412

5e83530

wpfleger96 mentioned this pull request Jul 1, 2026

feat(buzz-agent): emit agent_thought_chunk for reasoning content #1436

Merged

wpfleger96 changed the title ~~fix(desktop): restore observer-feed regressions from #1381~~ feat(desktop): restore observer-feed regressions from #1381 and classify 4 new session/update types Jul 1, 2026

wpfleger96 pushed a commit that referenced this pull request Jul 1, 2026

screenshots: PR #1412

32b9792

wpfleger96 force-pushed the duncan/observer-feed-1381-restore branch from 31579f2 to d791a25 Compare July 1, 2026 19:06

wpfleger96 pushed a commit that referenced this pull request Jul 1, 2026

screenshots: PR #1412

a811588

wpfleger96 force-pushed the duncan/observer-feed-1381-restore branch from 7a3838d to 9fa77da Compare July 1, 2026 20:20

npub1mn7jgtj4w2pd0g0zeuhxsa6jy6p0rewxz4kujt98my82ahfmp72sxjexk7 and others added 15 commits July 1, 2026 16:30

test(observer-feed): add E2E screenshot specs 07-10 for new session/u…

b4f8557

…pdate types Co-authored-by: Will Pfleger <pfleger.will@gmail.com> Signed-off-by: Will Pfleger <pfleger.will@gmail.com>

wpfleger96 force-pushed the duncan/observer-feed-1381-restore branch from 9fa77da to 0ffd40f Compare July 1, 2026 20:30

wpfleger96 merged commit fec7684 into main Jul 1, 2026
20 checks passed

wpfleger96 deleted the duncan/observer-feed-1381-restore branch July 1, 2026 20:33

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

feat(desktop): restore observer-feed regressions from #1381 and classify 4 new session/update types#1412

feat(desktop): restore observer-feed regressions from #1381 and classify 4 new session/update types#1412
wpfleger96 merged 16 commits into
mainfrom
duncan/observer-feed-1381-restore

wpfleger96 commented Jun 30, 2026 •

edited

Loading

Uh oh!

wpfleger96 commented Jul 1, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Uh oh!

Conversation

wpfleger96 commented Jun 30, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

What

1. Restored observer-feed behaviors (6 items)

2. New lifecycle types (4 items)

3. System-prompt display ordering

Tests

Uh oh!

wpfleger96 commented Jul 1, 2026

Shots 01–10: restored behaviors and new lifecycle types

Shot 11: first-turn ordering fix (new)

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

wpfleger96 commented Jun 30, 2026 •

edited

Loading