Skip to content

STATE.md

Latest commit

 

History

History
127 lines (107 loc) · 6.27 KB

STATE.md

File metadata and controls

127 lines (107 loc) · 6.27 KB

STATE — agentic_rc_cli

Frozen: 2026-05-21 16:55 (Europe/Berlin) Branch: develop Last commit: 8ea1000 · feat: real-world hardening v0.6.0 Dirty: uncommitted v0.6.1 eval-target-library fix — staging now

Last work-unit

Shipped v0.6.1 — the actual fundamental fix for Flutter Web that v0.6.0's diagnostic surfaced. Flutter Web's rootLib is a generated web_entrypoint.dart that doesn't import the framework → eval against it can't resolve Element / WidgetsBinding / *Button. Every gesture / inspector tool failed silently on Web (would have, before v0.6.0 diagnostic; failed loudly with eval_kind:"@Error", eval_error:"RPC 113 Expression compilation error" after).

Fix: FlutterService.evalTargetLibraryId() probes each candidate library with the bare identifier Element. First one that compiles wins, cached for the session. Order: rootLib → material.dart → widgets.dart → cupertino.dart. macOS keeps using rootLib (1 probe call), Web transparently falls back to material.dart (2-4 probe calls).

Captured the new constraint as Constraint #5 + #4 (Dart 3 records ban) in docs/learnings/vm-service-eval-quirks.md. 64/64 unit tests + tap-demo + login-demo all green on macOS — no regression. Ready for re-test against the user's Flutter Web app.

Previous v0.6.0 changes also still in this commit chain:

Shipped v0.6.0 real-world hardening, based on a Flutter Web session where four blockers surfaced. Top-4 prio agreed with user, implemented:

  1. Universal eval diagnostic — new shared safeEval helper in src/tools/flutter/_eval_diagnostic.ts. Every tool that calls svc.evaluate(...) now surfaces eval_ok / eval_kind / eval_error / expression_preview instead of the v0.5 opaque reason: "empty", raw_eval: null. Migrated tap, widget_geometry, wait_for_widget; enter_text already had its own diagnostic. Convention captured in docs/learnings/eval-diagnostic-discipline.md.
  2. Tap walker: self → descendants → ancestors (default; opt-out descend:false). Real apps wrap built-ins in custom widgets (TPKButton → TextButton); old walker missed them. Multiple descendants → reason:"ambiguous_descendants" + structured target list so the agent can disambiguate via by:"key" instead of guessing.
  3. by: "text" matcher on tap/geometry/wait_for_widget. Match a Text widget whose data contains the value (case-insensitive substring). Pairs with the descendant-first walker: by:"text", value:"Sign In" finds the wrapping button via ancestor walk.
  4. widget_tree filtering & flat mode. Default include_framework:false — framework subtrees collapse to {_elided:true, framework_node_count:N} markers (kills the 200 KB tree explosion). source_prefix:"…" for strict path filter. flat:true returns a flat list with path strings instead of a nested tree (~70% token saving combined with source_prefix).

One bug rediscovered: Dart 3 record types ({void Function() cb, …}) are rejected by the VM-service eval frontend with RPC 113. Workaround: use List<dynamic> 2-tuples for the (cb, name) pairs. Added to the permanent list in docs/learnings/vm-service-eval-quirks.md (next edit).

22 MCP tools, 64/64 unit tests (was 57; +7 for by:text / descend / ambiguous), tap-demo and login-demo both green end-to-end. Skill + README + CLAUDE.md trigger updated for the new learning.

Next intended step

Two open paths, user's preference unclear:

  1. Real Flutter Web test session — the user reported the original findings while testing on Flutter Web. v0.6 fixes should make Web workable; the diagnostic in particular will reveal any new Web-specific issues. If new findings emerge → new learning file docs/learnings/flutter-web-quirks.md + trigger row in CLAUDE.md.
  2. More gesture coverage — rc_flutter_long_press, rc_flutter_swipe, rc_flutter_scroll, rc_flutter_dropdown_select. Same pattern as tap. ~30-60 min per primitive incl. live demo.
  3. Inspector key-matching fix — Push widget_find by=key to the Dart-eval path by default. Open since v0.5.
  4. CHANGELOG.md — git log has it but no human-readable summary. 15 min.

User flow last expressed: "lass uns alles umsetzen" → Top-4 done, now they'll likely want the real-world re-test (path 1).

Open friction

  • rc_flutter_screenshot extension_not_registered on macOS desktop.
  • Login demo patches flutter_example/lib/main.dart; SIGKILL skips cleanup (very rare; SIGINT/SIGTERM safe).
  • widget_find by=key still uses the cached inspector path (drops keys on Text leaves) — workaround documented in docs/learnings/inspector-tree-keys.md.

Live context for the agent

  • Active spec areas: src/flutter/gesture_dart.ts (Dart expressions; use List<dynamic> for tuples, never records), src/tools/flutter/_eval_diagnostic.ts (every new eval-driven tool must go through safeEval).
  • Empirical Dart-eval constraints captured in docs/learnings/vm-service-eval-quirks.md: single-line only, no @visibleForTesting, no Dart 3 records.
  • User-code-only widget tree by default — pass include_framework:true only when debugging framework wrappers.
  • Demo discipline: every new tool gets a live-driven script under scripts/. Live verification is the truth, not the unit tests.
  • User mood: real-world feedback driven, expects rapid iteration — not over-design. Surface gaps with diagnostics; fix in subsequent passes.

How to resume

  1. Read this file.
  2. git log -5 --oneline and git status -s — detect any drift since 2026-05-21 16:05.
  3. If clean and the user said "weiter": offer the four Next-intended-step paths in 2 sentences each, wait for them to pick. Most likely they want path 1 (Flutter Web real-world re-test) since v0.6 was built for exactly that.
  4. Recent reflexion: v0.6 hardening came in ~50 min (4 features + tests
    • docs + 1 bug fix on Dart records). Naive 60 min was a tight upper bound; minute-unit calibration continues to hold.