feat(ai): better prompts, console-error context, and idle-session recovery#360
Merged
Conversation
dobrinyonkov
force-pushed
the
refactor/ai-assistant-architecture-v1
branch
from
July 2, 2026 09:18
2145a54 to
285be09
Compare
Rewrite PromptBuilder.buildSystemPrompt around a small-model-friendly shape: a one-sentence Role, a numbered Rules block (English-only reply, grounding to the Current UI5 Control Context, prescribed uncertainty phrase, runtime-data preference), a short Style block, and a copy-me bad/good example instantiating the uncertainty phrase. Enrich Current Application Context with three additional fields: UI locale (configurationComputed.data.language), sap-ui-debug when present in URL parameters, and Application entry point from common.data.Application. Each line is omitted when its data is unavailable, and the section is dropped entirely when appInfo has no recognized fields. Assertions cover: zone order, rule order and content, uncertainty phrase template, the bad/good example, per-field presence/omission, and four golden-file snapshots (no info, framework only, framework+theme+libraries, all six fields). No production behavior change beyond the emitted system prompt string.
…pe-driven curation Replace the raw-JSON dumps of properties, bindings, and aggregations with curated, one-line renderings, and wrap the user's question in a sandwich so the small local model keeps focus on it after the context block. New user-prompt shape when Inspection Context is attached: User asked: <msg> Current UI5 Control Context: - Type: ... - ID: ... Properties: - key: value Bindings: - prop <- "path" = value (model: ..., type: ..., formatter: yes) Aggregations: - name: N children (Type x N, Other x M) Now answer: <msg> Per-section rules: - Properties render as key: value lines; capped at 800 chars. - Bindings render one line each with the resolved value inline when the snapshot carries it; null and undefined print literally; values truncate at ~100 chars; composite bindings (parts: []) collapse to a degenerate <composite> line; section capped at 800 chars. - Aggregations render as: empty (empty), N children -- id1, id2 (<=3), or N children (Type x N, ...) (>3); section capped at 400 chars. - Circular binding graphs still bail out to the pre-existing 'cannot serialize' placeholder. No-context / no-control calls return the raw user message unchanged. Existing PromptBuilder.spec.js invariant tests are updated to the new shape; new golden-file tests cover identity-only, properties-only, bindings, and mixed aggregations. Two AssistantController tests are updated to assert 'Now answer:' instead of 'User Question:'. Refs .scratch/ai-prompt-context-quality/issues/02-rewrite-user-prompt.md
The helper functions and module-scope constants added in 16e273b used 'var' — inconsistent with the rest of the file (and the codebase's ES2020+ baseline). Switch to const/let. No behavior change.
Pure state machine over an error-event stream backing the AI Assistant's Recent Console Errors signal. Bounded FIFO of capacity 3 with dedup by (normalized message, top-shown stack frame) — duplicates increment the count on the existing entry and do not re-promote. Stack-frame selection skips up to 3 framework frames matching 'sap-ui-core.js' or 'resources/sap/', then falls back to whatever frame we landed on. No browser dependencies — testable without a running page. Part of .scratch/ai-prompt-context-quality/issues/03.
Thin adapter around consoleErrorBuffer. Monkey-patches console.error / console.warn (originals continue to run) and subscribes to window.onerror and unhandledrejection, funneling every event into the buffer. An optional onRecord callback lets the injected main.js push a fresh snapshot to the panel on every recorded event. install() is idempotent across a re-injected content script. Part of .scratch/ai-prompt-context-quality/issues/03.
buildUserPrompt gains a third argument for the recent-console-errors snapshot. Non-empty renders a 'Recent Console Errors:' block inside the sandwich, immediately after Current UI5 Control Context (or in its place when no control is attached). Empty or missing snapshot omits the section entirely. Backwards-compatible: existing 2-argument callers keep working. Each entry renders the message with a '(×N)' annotation when count > 1 and — when a frame is present — an indented 'at <frame>' line beneath. The section is capped at ~400 characters with the standard [truncated] marker on overflow. Errors render newest-first (reverse of the buffer's natural FIFO arrival order). Part of .scratch/ai-prompt-context-quality/issues/03.
…ontroller AssistantController accepts two new optional injected callbacks mirroring the existing getAppInfo pattern: - getConsoleErrors — called on every sendUserMessage; the snapshot is forwarded to PromptBuilder.buildUserPrompt as the third argument. - clearConsoleErrors — called from clearConversation and setUrl(differentUrl) so buffered errors reset in lock-step with Conversation Memory and Inspection Context. Both seams tolerate missing options, undefined return values, and throws from broken panel wiring — the send flow and clear paths never break because of a downstream integration. Part of .scratch/ai-prompt-context-quality/issues/03.
…tant
End-to-end wiring for the recent-console-errors signal:
- The injected script installs consoleErrorCapture on load and pushes a
fresh snapshot to the panel via the existing devtools port protocol
('on-console-errors-updated') on every recorded event.
- Panel main.js caches the snapshot in frameData[frameId].consoleErrors
and wires getConsoleErrors / clearConsoleErrors to the AI Assistant.
- clearConsoleErrors sends 'do-clear-console-errors' to the injected
script so the panel's cache and the page-side buffer stay in lock-step
when the developer invokes Clear Conversation.
- AIChat forwards the two new callbacks into AssistantController.
Closes .scratch/ai-prompt-context-quality/issues/03.
…work Delete comments that restate the next line, narrate the PRD, or repeat what a constant name already says. Keep comments that explain non-obvious 'why' — races, external-quirk workarounds, historical constraints. Touches only files from the previous 9 commits: - PromptBuilder.js: -117 lines - consoleErrorBuffer.js: -46 lines - consoleErrorCapture.js: -25 lines - AssistantController.js: -14 lines - injected/main.js: -4 lines - devtools/panel/ui5/main.js: -5 lines No behavior change. 571/571 tests pass, lint clean.
Rewrite remaining comments using plain, direct language: no jargon (sandwich, curated, funnel, lock-step, glue, adversarial), no editorializing about what the model sees, no filler doc-block prose. Shorten sentences and drop redundant qualifiers. Same 6 files as the previous docs commit.
Enrich JSDoc and inline comments for the console-errors and PromptBuilder work with more rationale (why the caps, why the sandwich, why the install- once guard). Behavior unchanged.
This reverts commit 193ffab, restoring the deliberately-lean comment state that d92920e had established across the six AI-work files. Docs history on ai-improvements-v1: 0c2d409 docs(ai): remove obvious/verbose comments (initial cull) d92920e docs(ai): tighten surviving comments (concise pass) 193ffab docs(ai): expand doc comments (re-expansion) this revert of 193ffab (restore d92920e state) The 193ffab re-expansion added multi-paragraph JSDoc rationale, rendered example prompts in docstrings, and forward references to follow-up issues. The d92920e state carried one-sentence summaries on public functions and bare @private/@param/@returns on helpers, which matches the density used elsewhere in this codebase. Docs-only. No code logic touched. See .scratch/ai-prune-comments/PRD.md for context.
Review-and-prune pass over the six files touched by the reverted 193ffab. Heuristic: if a JSDoc block has more sentences than its function has statements, cut prose. Four files needed no changes; two had cuts: - AssistantController#on: drop "in-process event bus" restatement — the class docblock already declares no Chrome deps. - AssistantController#sendUserMessage: shrink 3-sentence race-guard essay to the one-line invariant; the "do not overwrite session-failed" clause is enforced mechanically at the rejection handler below. - AssistantController#updateInspectionContext: collapse numbered lifecycle list to prose. Same info, one sentence. - PromptBuilder#_renderAggregationLine: drop rendered example prompt from docstring — tests cover output shape. - PromptBuilder#_renderBindingLine: same. No code logic changed. npm test: 582/582 pass.
…one method
Collapse the two AssistantController reseed-tracking methods into a single
_trackReseed(rawSeed, { announceReady = false } = {}). The pair only
differed in whether 'ready' was re-emitted on the Assistant Capability
State stream after a successful reseed.
Call sites:
- initialize, downloadModel: _trackReseed(seed) (unchanged)
- clearConversation, setUrl reseed: _trackReseed(seed, { announceReady: true })
The _pendingReseed write, the session-failed translation on rejection,
and the 'resurface ready when capability is currently ready' logic all
stay in one place now. Pure internal refactor; no public API, event, or
test assertion changed. All 582 tests pass.
When Chrome tears down the extension's background service worker during idle (~30-60s of port inactivity), the Prompt API session dies with it. Prior to this fix the next Send tripped PromptClient's 'No active session' guard and surfaced as a streaming-failed banner; a second Send recovered. AssistantController now checks PromptClient.hasActiveSession() at the top of sendUserMessage. If false and no reseed is already in flight, it kicks off a reseed via _trackReseed(_seedSession()) before awaiting the pending-reseed gate. The recovery reseed carries the current Conversation Memory so follow-ups still reference earlier turns, and the sticky Inspection Context (ADR-0002) applies to the recovered prompt as usual. UI is silent — the first Send after idle just takes slightly longer to start streaming. Prefactor: _pendingReseed is now null when no reseed is in flight and a Promise while one is (previously initialized to Promise.resolve() as a sentinel that was never reset). All existing reseed origins funnel through _trackReseed, which clears the field back to null on both fulfillment and rejection. sendUserMessage reads it defensively via (this._pendingReseed || Promise.resolve()).then(...). PromptClient gains a hasActiveSession() getter over the private _hasActiveSession flag. The existing onDisconnect handler that synthesises a mid-stream error for in-flight streams stays untouched. Tests: six new cases in tests/modules/ai/AssistantController.spec.js exercise session-dead-on-Send, memory carried through recovery, coalescing with an in-flight clearConversation reseed, recovery reseed failure surfacing as session-failed (not streaming-failed), unchanged alive-path behaviour, and Inspection Context surviving recovery. The fake PromptClient gains a controllable hasActiveSession() (default true). Refs .scratch/ai-session-dies-when-idle/PRD.md Refs .scratch/ai-session-dies-when-idle/issues/01-recover-from-idle-killed-session.md
dobrinyonkov
force-pushed
the
ai-improvements-v1
branch
from
July 3, 2026 12:12
cbea3d3 to
b88ae43
Compare
kgogov
self-requested a review
July 7, 2026 09:16
kgogov
approved these changes
Jul 7, 2026
Contributor
|
🎉 This PR is included in version 1.10.0 🎉 The release is available on GitHub release Your semantic-release bot 📦🚀 |
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.
Three improvements to the AI Assistant, plus a bug fix.
Better prompts
The system prompt is now structured into Role / Rules / Style zones with a
bad/good example for the uncertainty phrase. Application context also picks up
UI locale,
sap-ui-debug, and the app entry point when available.The user prompt uses a sandwich structure — the user's question is repeated
after the context block so the small local model stays focused on it:
Properties, bindings, and aggregations are rendered as curated one-liners
instead of raw JSON dumps. Bindings show the resolved value inline. Aggregations
with more than 3 children collapse to a type histogram. Each section has a
character cap so a huge control can't blow out the context window.
Recent console errors as context
The injected script now captures
console.error,console.warn,window.onerror, andunhandledrejectioninto a small bounded buffer(capacity 3, deduped by message + top frame). The buffer is pushed to the
panel and rendered into the user prompt as a
Recent Console Errors:blockso the assistant can see what the app just complained about.
Clearing the conversation also clears the buffer, on both sides.
Idle-session recovery
Chrome tears down the extension's background service worker after ~30-60s of
port inactivity, which killed the Prompt API session with it. The next Send
would fail with a streaming banner and a second Send was needed to recover.
sendUserMessagenow checks whether the session is still alive and reseedstransparently if it isn't. The first Send after idle just takes slightly
longer to start streaming. Conversation memory and Inspection Context survive
the recovery.
Tests
582/582 pass. New tests cover the console-error buffer/capture modules, the
new prompt shape, and the idle-recovery paths.