[Lightheartdevs]

Problem

OpenClaw 2026.6.33 drops before_agent_finalize revisions after a plugin tool call, so an inspection fix can reach the model only in the tool result. On main d4a61f3 (fleet round 107), website_create_artifact_preview_interaction failed on laptop, strixy and mac-mini, and the failure cascaded into website_followup_edit and conversation_reload_continuity. All three towers passed.

Every failing host used up the progress budget (4 consecutive or 12 total failures) on the inspector itself:

• strixy r107: flattened steps were rejected with "call tool_describe ... then retry through tool_call". The model then called tool_call {id:"tool_describe"} twice and was stopped.
• laptop r107:
  • Call 4 had no exact:true.
  • Call 5 used a state-qualified locator on a page that never hid the card.
  • The model then read the receipt's before as a pre-script snapshot, and later sent {}.
• mac r106 ran into four locator failures on a page that works.
• mac r107 said a broken click handler was "a timing issue".

Evidence is in tower2 ods-main-qualification-20260924-r049, round-10{5,6,7}.

Change

Every non-passing inspection result now names exactly one of these:

• one ready, validated call on the same snapshot, when the site needs no change;
• a measured diagnosis with one repair, plus the steps to run on the new snapshot;
• or, when this failure ends the response's tool use, no call at all.

isError stays true. details stays the capsule receipt, or the invalid_request shape. The Evidence JSON, the palette line, page-error precedence and the INCOMPLETE (#6740) path are unchanged. Proof binding is unchanged.

Rejected arguments (invalid_request)

• The text names the step, the key and the reason.
• A lossless repair is offered as "call ... with exactly these args". It moves a flattened selector or role/name into locator, and adds exact:true to a role/name locator.
• If the repaired steps do not test a change and an owner requirement is bound, the requirement plan is offered instead. For laptop call 4, that plan asserts .event-card.sold-out hidden, clicks, then asserts it visible.
• Otherwise the text shows one placeholder shape example.
• The tool_describe sentence is gone. The guard's coaching now says tool_call takes the tool's id directly, which is how OpenClaw 2026.6.33 tool-search findEntry works.

Locator failures (no_match, selector_not_unique, invalid_selector)

• The owner-requirement plan is offered, and never reuses the failing locator.
• Failure paths prefer a CSS target for the owner-named item over a role/name heading. That target never contains another heading or the control. It is measured whether or not the element is rendered, so a broken toggle gets a measurement instead of a no_match dead end (laptop call 13, site-c54d9d87).
• A role/name click that matched nothing gets the published CSS locator of the control whose authored name matches ignoring case. The outline now records control roles and names. This covers laptop call 24, a visible text-transform: uppercase button.
• fix(pixel): verify owner-requested exact control names after the page scripts run #6741's load-time names take precedence:
  • When they show the exact name on a rendered control, the same steps are offered with that control's CSS locator.
  • When they show a different name, the text keeps fix(pixel): verify owner-requested exact control names after the page scripts run #6741's repair-or-rename wording and offers no call.

Measured failures (visibility_mismatch, click_failed, unstable)

• The text states the step's measured display, visibility, opacity and rectCount, which clicks ran before it, and the inspector facts.
• The next step depends on the owner's stated direction. That is statedVisibilityDirection: an initially-hidden phrase or a click that only reveals means hidden first, and a click that only hides means visible first. A toggle, both or neither leaves it unstated.
  • Stated, and the page contradicts it: repair.
  • Stated, and only the plan's order is reversed: the owner-order plan on the same snapshot, with "the site needs no change".
  • Not stated: both orders are offered, and the owner is never said to have asked for one.
• After a click:
  • If passing earlier steps already measured one locator hidden, then clicked, then visible, and the failing step checks another element, those passing steps are offered. This is tower2 r107 call 20, a passing session that previously would have been told to repair a working site.
  • If the failing step is another element, the target plan is offered.
  • If the element changed where the plan expected no change, the corrected step is offered.
  • If the target did not change, repair.
• After a repair, the steps for the new snapshot are the target plan without identifiers. They never reuse "the same steps". The target is an id, or the item's other classes when the class the repair may change is the locator (laptop call 40).

Final failure

• The guard's previewInspectionGuidance(toolCallId, params) is bound to exactly the pending call, like previewInspectionTransition. It carries the stated direction and finalFailure from the read-only budget.failureEnds(lane).
• On the failure that exhausts the budget or its lane, every next step is replaced by "do not call any tool; answer the owner now". The progress-finalization pending phase accepts that answer.

Review fixes (round 1, d775fbd)

Finding	Fix
Before-click text treated the hidden-first default as the owner's words (P4, P5, P9)	Stated direction only; neutral otherwise; the owner-order plan when only the plan is reversed
Post-click repair of a working site (tower2 r107 #20)	Proven-prefix and off-target branches offer ready calls; repair only for an unchanged target
Laptop r107 not rescued	Requirement plan for call 4; a CSS control for role/name click misses; CSS item targets; repair-stable steps
Heading dead end on a broken toggle	CSS item target; the role/name text after "matched hidden before the click" names both causes
"The same steps" defeat themselves	Target-plan steps with a repair-stable target
Ready args at the budget limit	finalFailure replaces the next step
No ready call when hidden at load	Owner-order or page-order plan on the same snapshot
Coaching still said "call tool_describe"	Reworded

Replays of rounds 105 to 107 through the real guard and real Chromium:

• laptop r107 (counterfactual, following each offered call): call 4 → requirement plan → "visible at load, owner has it hidden" repair with its steps → after the repair, the role/name click miss gets the #showSoldOutBtn plan → passes on the repaired page. Three failures, with a reset in between.
• tower2 r107: call 20 → the passing steps 1–7 pass and verification passes.
• strixy r107 and mac r106: unchanged ready calls. The final failure now says to answer.
• mac r107: the repair steps (heading around .reveal-btn) pass once the handler toggles the class.

Review fixes (round 2, f67b0ab)

Finding

Fix

A passing change of another element was offered in place of the owner's target. On a page whose click reveals an unrelated #note but leaves the requested #midnight-card hidden, the plan #note hidden, click, #note visible, #midnight-card visible failed at step 4 and was offered steps 1 to 3. Sending them passed, bound the visibility proof and set the run's verification to passed (reviewer probes P1 and P1c, direct and tool_call). The guard's own instruction says not to substitute an unrelated passing interaction.

With a target named from the published source, the passing steps are offered only when they changed that target, and never when the failing step is on the target itself. That step then gets its measured repair. A failing step on another element gets the target plan, which measures the real defect.

• Tower2 r107 call 20 keeps its ready call (steps 1 to 7). This was checked with the PR's fixture and also with the guard's outline, rebuilt from that session's own writes and edits (digest-verified against publication 19's sha256).
• New test in inspection_recovery_replay.test.mjs: "another element's passing change is never offered in place of the owner's target". It fails on 447fd80 and passes here.

Composition notes (for whichever of these lands second)

• Defer a host-guided inspection failure that would stop the run for exactly one corrected attempt, with every hard bound kept #6749: failureEnds does not know about Defer a host-guided inspection failure that would stop the run for exactly one corrected attempt, with every hard bound kept #6749's waiting charge. Composed as merged, a correctable failure at the fuse is told "do not call any tool" even though the budget admits its corrected call. The strixy test "the failure that ends tool use names no further call" fails on main + fix(pixel): every non-passing preview inspection returns a ready, validated next call or a measured diagnosis #6748 + Defer a host-guided inspection failure that would stop the run for exactly one corrected attempt, with every hard bound kept #6749. A reviewer sweep of 300 plans found 83 results worded as final whose next call the guard admitted. The integration needs failureEnds(lane, {correctable}) to follow Defer a host-guided inspection failure that would stop the run for exactly one corrected attempt, with every hard bound kept #6749's failure path, plus a correctableWaits hint. With that, the same sweep finds 0 mismatches and the full suite passes.
• fix(pixel): keep a passing show/hide proof across later passing checks of the same snapshot #6754: "round 108: a later passing check with a click keeps the transition proof of the same snapshot" asserts byte-identical recorded failure texts from before this PR. On main + fix(pixel): every non-passing preview inspection returns a ready, validated next call or a measured diagnosis #6748 + fix(pixel): keep a passing show/hide proof across later passing checks of the same snapshot #6754 it fails on that text assertion only, so that assertion needs updating.

Merge with main (#6741)

Main moved: #6741 added load-time control names. This branch merges 6cba572. The three conflicts are resolved as described in the merge commit:

• fix(pixel): verify owner-requested exact control names after the page scripts run #6741's texts and tests are kept.
• Its "address that control with a CSS selector" case now carries the ready call.
• The evidence copy omits controls.

The branch merges cleanly with current main 321f82a (#6746, playground paths only).

Not changed, and why

• The proof accepts a transition in either direction (boundVisibilityInspection). This behavior predates this PR. Gating proof on direction would rest on the wording heuristic, and "Clicking shows it; clicking again hides it" already derives visible-first, so correct toggles would fail. This PR only removes the steering: it never offers a plan reversed against a stated direction.
• Heading targets still appear when no CSS item target exists. After a heading matched hidden and then nothing rendered after the click, only a CSS target is offered. Otherwise the text names both causes and offers no call.

Deliberate test edits (owner sign-off)

• inspection_transition_coverage.test.mjs: one assertion (from round 1). The tower2 r102 create-turn no_match text is kept byte for byte, with one ready call inserted. This is the same pattern as laptop call 5. Nothing else in that file changed.
• preview_interaction_assurance.test.mjs: the finalize coaching assertion /tool_describe/ is inverted, and the new tool_call sentence is asserted.

Tests

• Full Linux suite (remote-js-test.sh, tower2):

  Ref	tests	pass	fail
  Round-2 head f67b0ab	3209	3171	0
  Round-2 head composed with main 0a63258 (local, not pushed)	3237	3199	0
  Merged head 447fd80	3208	3170	0
  Main control 6cba572	3177	3139	0
  Composed with current main 321f82a (local, not pushed)	3223	3185	0
  Round-1 head d775fbd	3168	3130	0

• tower2 Linux:

  • node --test ods/tests/test-preview-inspection.mjs: 30/30.
  • python3 -m unittest tests/test_preview_inspection.py: OK (89 tests, 46 opt-in skipped).

• Opt-in real Chromium (ODS_PREVIEW_BROWSER_TESTS=1): all 39 BrowserTests pass, including fix(pixel): verify owner-requested exact control names after the page scripts run #6741's. They ran locally on Windows, with the Playwright headless shell and a scratch shim for os.killpg. tower2 has no Python Playwright. The new tests cover:

  • the uppercase control needing its CSS locator;
  • the CSS item target measuring the broken toggle on site-c54d9d87;
  • call 24's id click reaching the measurement;
  • the mac r107 repair steps failing on the recorded page and passing after the repair.

• New or updated JS tests:

  • inspection_guidance.test.mjs: stated direction for exactly the pending call, inheritance on a preserve turn, toggle, finalFailure, and outline control names.
  • inspection_recovery_replay.test.mjs: laptop calls 4, 5, 13, 24 and 26; strixy at the limit; tower2 r107 calls 20 and 21; mac r107 with and without the requirement; P4, P5 and P9; and load-time names with the published control.
  • run_progress_budget.test.mjs: failureEnds.
  • ods/tests/test-preview-inspection.mjs: direction-aware texts, and the final failure across every failing path.
  • New fixture inspection-recovery-tower2-round107.json, byte for byte from J/round-107/tower2/serving-logs/openclaw-pixel-a674292a-….jsonl.gz.
  • Laptop fixture: calls 13 and 24 added.

🤖 Generated with Claude Code