Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@docs/implementation_plans/2026-06-02_executor_idempotency_fencing_plan.md`:
- Line 128: The phrase "synced lease state + fence-rejection-on-reconnect,
degrades to this design" uses nonstandard/awkward wording; change the defer
rationale to clearly phrase that the system "needs lease state to be synced" (or
equivalent like "requires the lease state to be synced") and keep the rest of
the clause intact so it reads e.g. "needs lease state to be synced +
fence-rejection-on-reconnect, degrades to this design." Update the sentence
containing "synced lease state + fence-rejection-on-reconnect, degrades to this
design" accordingly to improve clarity.

In `@lib/features/agents/projection/compaction_summary.dart`:
- Around line 155-159: The body text normalization currently uses rawBody.trim()
which preserves embedded newlines; update the normalization in the compaction
logic (references: rawBody, body, editedTag and the similar block around lines
166-168) to collapse all newline and excessive whitespace into single spaces and
then trim so each entry is a single line (e.g., replace sequences of
newline/whitespace with a single space, then .trim()) to maintain the compacted
tail's line-oriented contract.

In `@lib/features/agents/projection/decision_events.dart`:
- Around line 27-33: The summary and reason may still contain embedded
newlines/tabs because only trim() is used; normalize internal whitespace for
entry.humanSummary and entry.reason so they become single-line strings (e.g.,
collapse all runs of whitespace to a single space) before using them to build
summary, trimmedReason and reasonSuffix in decision_events.dart; update the
variables summary and trimmedReason to perform this normalization (split/join or
regex replace) so the returned ledger line always stays on one line.

In `@lib/features/agents/README.md`:
- Around line 442-465: The README currently contradicts itself about whether the
read-path and summary-checkpoint flow are active behind the
enable_agent_compaction flag; reconcile the two sections so there's a single
source of truth by either (A) marking the earlier block as "Planned
Improvements" and noting that AgentLogCompactor.maybeCompact,
AgentLogCompactor.assembleContext, AgentLogLlmSummarizer and
buildTaskStateMarkdown are not yet wired into the live wake read-path, or (B)
updating the later "Planned Improvements" lines to state that these features are
already live when enable_agent_compaction is true (with
TaskAgentWorkflow.compactionEnabled remaining a test override); pick one
approach and make both sections consistent about the runtime behavior and flag
semantics.

In `@lib/features/agents/tools/agent_tool_registry.dart`:
- Around line 228-235: The item schema currently requires "id" but allows `{
"id": "..." }`; update the JSON schema for the checklist item (the schema object
inside the parameters for the checklist-update tool) to require "id" and also
add an anyOf constraint so at least one of isChecked, title, or isArchived is
present (e.g. keep required: ["id"] and add anyOf:
[{required:["isChecked"]},{required:["title"]},{required:["isArchived"]}]).
Apply the same change to the other identical item schema occurrence mentioned in
the comment so the validator rejects no-op items.

In `@lib/features/agents/workflow/task_agent_workflow.dart`:
- Around line 360-363: The DB read for enableAgentCompactionFlag can throw and
currently will abort wake execution; change the compactionOn initialization to
catch failures from journalDb.getConfigFlag and default to false so compaction
remains optional: when computing compactionOn (which currently uses
compactionEnabled ?? await journalDb.getConfigFlag(enableAgentCompactionFlag)),
wrap the await call in a try/catch (or use a safe helper) to return false on any
error, then continue constructing compactor as before (referencing compactionOn,
compactionEnabled, journalDb.getConfigFlag, enableAgentCompactionFlag, and
compactor).

In `@test/features/agents/workflow/task_state_markdown_test.dart`:
- Around line 62-65: Replace the brittle .singleWhere lookup in the test with a
safer .firstWhere (or .where(...).first) and include the closing parenthesis in
the match so you match the full "(id: ...)" token; specifically change the
expression on text.split('\n') that currently uses .singleWhere((l) =>
l.contains('(id: ${item.id}')) to e.g. .firstWhere((l) => l.contains('(id:
${item.id})'), orElse: () => throw StateError('no matching line for id
${item.id}')) and keep the existing expect assertion using item.completed.

---

Outside diff comments:
In `@lib/features/ai/functions/lotti_checklist_update_handler.dart`:
- Around line 141-146: Update the user-facing validation/help strings in
lotti_checklist_update_handler.dart to include `isArchived` as a valid update
field; specifically change the message passed to _createErrorResult (the
"Provide a JSON array..." hint) and the other similar guidance string later in
the same handler so they list {"id": "...", "isChecked": true, "isArchived":
true} (and mention title if present) — search for the current strings that
mention only "isChecked" / "title" in the handler and add `isArchived` to both
messages.

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Duplicate comments:
In `@lib/features/agents/projection/compaction_summary.dart`:
- Around line 185-189: The code builds rawBody/body which can contain embedded
newlines and thus violate the single-line compacted tail; update the
normalization so rawBody collapses all whitespace/newlines into a single space
before trimming (e.g. replace runs of whitespace like \s+ with a single space on
rawBody) so that body = normalizedRawBody.trim() yields a single-line string;
apply this change around the rawBody/body construction (referencing rawBody,
body, and editedTag) so entries never contain embedded newlines.

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Outside diff comments:
In `@test/features/agents/projection/compaction_summary_test.dart`:
- Around line 127-145: The oracle's isValid(SummaryCheckpoint s) only checks
retractions but ignores the completeness check performed by selectActiveSummary;
update isValid to also validate that every non-suppressed event at or before
s.cutoff is covered by s.coveredSources by inspecting log.events and
latestRetraction: for each event where event.position <= s.cutoff, skip if
suppressed (latestRetraction[event.contentEntryId] exists and
isAfter(event.position)), otherwise require s.coveredSources contains
event.contentEntryId; only return true when both the retraction check and this
completeness check pass so the property test exercises the same logic as
selectActiveSummary.

In `@test/features/agents/workflow/task_agent_workflow_test.dart`:
- Around line 819-831: The test stub for mockAgentRepository.getEntity returns a
hardcoded id ('sha256-v1:existing') instead of echoing the requested digest;
update the when(...).thenAnswer(...) to capture the incoming digest argument and
use it as the returned entity's id so the mocked content-addressing contract
holds. Specifically, change the thenAnswer closure for
mockAgentRepository.getEntity to read the digest from the invocation (e.g.,
invocation.positionalArguments.first or invocation.positionalArguments[0]) and
pass that value as the id to AgentDomainEntity.agentMessagePayload (while
keeping the other fields the same).
- Around line 2784-2801: parseLinkedTasks currently replaces a missing id with
the string 'linked-task', which hides the "null id" test case; change the id
assignment to preserve null by using id = row['id'] as String? (no default) and
update the title fallback to (row['title'] as String?) ?? id ?? 'linked-task' so
empty-string and missing-id cases are handled differently; apply the same change
to the other parseLinkedTasks occurrences that construct AiLinkedTaskContext to
ensure tests for null-id remain valid (refer to parseLinkedTasks and
AiLinkedTaskContext usages).

---

Nitpick comments:
In `@lib/features/agents/README.md`:
- Around line 319-323: Reword the ambiguous sentence that describes behavior
when enable_agent_compaction is false: update the sentence around "with the flag
off, only capture runs and the wake reads the journal as before" to explicitly
state that capture still runs in shadow (building the event log) while the
legacy wake path continues to read the journal unchanged; reference the flag
name enable_agent_compaction and ADR 0017/ADR 0020 if helpful and produce a
clarified line such as "with the flag off, capture still runs (building the
event log in shadow), but the wake reads the journal as before — byte-identical
behavior."

In `@test/features/agents/workflow/task_agent_workflow_test.dart`:
- Around line 761-763: Replace the brittle numeric assertion on
mockSyncService.upsertEntity with assertions that the expected persisted
entities (e.g., the system prompt entity, the user message entity, and the state
update entity) were upserted: capture the arguments passed to
mockSyncService.upsertEntity (using the mock's capture mechanism or an
ArgumentCaptor equivalent), then assert the captured list contains entities with
the expected identifiers/types/contents (inspect properties such as
role/messageId/payload) rather than asserting .called(5); focus assertions on
upsertEntity and the specific entity properties to make the test resilient to
unrelated persistence changes.