On Windows the CPU-time and wall clocks both advance in about 15.6 ms
ticks, but not in phase. The test waited only for the CPU clock and then
asserted both had advanced, so a loaded runner could see the CPU tick
land before the wall clock's and fail. Wait for both, still bounded by
5 s.
@opentui/core's isRenderable was an inert stand-in, whose call returns a
truthy proxy, so isRenderable(undefined) passed and the plugin read a
field of undefined next (CT-40: recap's app slot, on anchor.parent once
findDescendantById found nothing). It now answers whether the value is
one of the compat layer's renderables.
The JS host's renderer.root stub holds none of the host's renderables
but lacked findDescendantById, so a plugin that searches the root for a
message's renderable crashed its app slot (CT-40: recap). It now finds
nothing, as OpenTUI does for an id it does not hold. Placing a node in
the transcript would need a contract mount; F16 says so.
OpenTUI clamps any scrollTo offset to the content, and plugins scroll to
the bottom with scrollTo(Number.MAX_SAFE_INTEGER). The compat layer sent
the number as is, the host could not read it as the verb's i32, and the
JS host was ended (CT-42: opencode-omeglecode). The host clamps to the
content anyway, so a number past i32 is now its edge.
A terminal resize rewraps every row, and the scroll offset stayed on the
same content row, so the rows above the reader's slid it down or up. The
width change also rebuilds every row a frame later, a second unanchored
reflow.
Before the layout, a resize away from the bottom now pins the rows above
the reader's row. Through the resize's frame and the rebuild's, each of
them that changes height scrolls the view by as much, as the reveal pin
does for rows mounted above. The content view's resize cannot do it: its
rows' geometry is not current yet in its handler. At the bottom the view
follows as before. A row taller than the view keeps its top; its own
content may still move as it rewraps.
streams_an_admitted_turn_and_resolves_with_usage_after_its_terminal_event
already drives the same tool, reasoning and terminal events. Check the
remaining thought texts and pending tool metadata in its transcript.
maps_failures_to_acp_errors checks the client-visible RPC error for
step authentication, execution authentication and rate limiting. Drop
the duplicate translation rows, keeping retry and cancellation cases.
alert_matches_typescript already draws a real dialog through the native
library. builtin_themes_match_the_typescript_resolver and
copied_tokens_shift_like_the_typescript_tokens check token lookup and
shifts without relying on how two visitors walk pointers.
oracle_key_sequences_run_the_commands_typescript_runs exercises the
generated defaults and Windows undo and suspend. Drop the hard-coded
definition counts and intermediate binding representation.
Replace old model and deviation labels with the actual conditions:
a cursor can become stale, a read can be overtaken by an event, and an
unlocked writer can change the file being replaced. Remove the same
tracking references from comments on UI behavior.
Clicking the open tab, cycling a single tab, or picking the current
session in the session list, the open dialog, a session link or a
tui.session.select event called Tui::navigate with the route already
shown. navigate treats that as a remount (theme and plugin changes rely
on it), so the transcript mounted again at its end and then restored
its scroll anchor: a flicker on every click.
TS's route.navigate reconciles the route store, so the same session is
a no-op. Tui::show_session does the same for the user's picks; the
deliberate remounts keep calling navigate.
appended_and_loaded_histories_keep_the_same_recent_prompts exercises
both public entry points. Whitespace and adjacent duplicates do not
consume history slots; nonadjacent repeats survive. Check the actual
sequence below, at and above 200 entries for both append and load.
requests_drive_a_headless_app checks typed text, modifiers, arrows,
resize and unsupported requests against a running app. Check the
string response ID, success and error envelopes, fractional resize
and unsupported click there instead of comparing request enums or
calling response constructors. malformed_requests_are_rejected keeps
the bad-schema cases.
reports_overlapping_contender_failures_without_recruiting_replacements
now kills the second child so the first child's status and stderr must
survive a distinguishable failure. Keep
recovers_when_its_own_unresponsive_contender_is_evicted_after_a_prior_failure
for eviction and reports_a_contender_terminated_by_a_signal for exit
status. Keep the deterministic clock checks for backoff boundaries.
reuses_a_registered_service, waits_for_a_registered_service_to_finish_starting
and reports_a_failed_registered_service_without_spawning cover the
normal probe decisions through ensure.
Each row still gets its own store and the same request schedule.
One test per invariant replaces the one-line wrappers:
echoes_resolve_without_duplicate_compactions
settled_compactions_stay_settled_after_the_response
rejected_rpcs_release_the_next_prompt
acknowledged_compactions_survive_http_errors
durable_items_survive_delayed_sse_and_http_failure
failed_preparation_releases_following_work
bulk_reads_settle_all_joiners
late_snapshots_preserve_live_activity
late_snapshots_cannot_resurrect_consumed_items
hydration_retains_concurrent_inbox_events
inflight_reads_cannot_resurrect_deleted_sessions
The previous length bound allowed dropping an ASCII character even
when it fit. tail_is_the_longest_suffix_within_the_utf16_budget checks
that the preceding scalar, if any, would overflow the budget. It
covers the surrogate-pair examples as well as empty, combining and
mixed-width strings.
usage_errors_match_the_typescript_cli already checks both output
streams for these exact arguments. Drop the parser tests' substring
checks for the same errors. help_matches_the_typescript_cli covers
hidden flags and visits all help pages; table_matches_the_typescript_cli
still compares the command tree without counting the fixture itself.
boolean_takes_a_literal keeps the positive boolean parsing cases.
cli_config_files_match_typescript covers ordered edits, merging,
legacy tabs and preservation of neighboring settings in the saved file.
decode_and_resolve_match_typescript checks default values. Keep
reloading_keeps_the_initial_tabs_environment for the reload-only case.
jsonc_edits_and_parse_match_typescript covers parsing and number
formatting with an independent implementation. Exercise escaped
surrogates, nested values, and both sides of the decimal/exponent
boundary through the edit result instead of a second formatter table.
a_session_panel_toggles_focuses_and_closes_itself and a_v2_plugin_reads_a_sessions_pending_inbox observe the actions and inbox from the real pipe. a_reload_gives_the_focus_back_only_to_a_field_that_held_it covers both focus outcomes across reload. a_field_shown_and_focused_in_one_handler_takes_the_focus_once_shown and keys_typed_as_a_dialog_opens_go_to_its_field check that input reaches the right mount. Remove the four corresponding fake-host rows. All 74 remaining JS tests passed.
empty_scripts_add_no_geometry and alphabets_and_styles compare rendered layouts instead of assembling parallel syntax trees. fractions_braces_matrices_and_tall_scripts now checks spaced array column rules on screen; the same table covers both brace directions. All 20 renderer properties and the layout oracle passed.
replacing_a_path_suppresses_its_contents_regardless_of_registration_order covers previous replacement winners, nested replacement suppression, preserved siblings and the strict dot boundary together. a_growing_status_claim_fills_the_footer_row covers flex growth in the real footer. Keep the distinct shrinking-root case. Four slot tests passed.
the_schema_takes_what_serde_takes now checks value equality after decoding its node corpus, replacing the single tree wire example. the_schema_takes_every_message_of_the_transcripts and a_claimed_part_is_equal_on_every_carrier exercise claim messages. the_envelopes_take_what_serde_takes retains the absent-node default. Seven schema and memo tests passed.
complete_frames_survive_every_stream_cut replaces the single round-trip and truncation examples and the random-byte loop. It checks all three kinds at every cut of a short payload, including an empty frame and a preceding intact frame. caps_and_kinds_are_checked_before_the_payload checks both length boundaries and every unknown kind. Both properties passed.
The private toggle-row examples duplicate the observable manager:
a_failed_built_in_shows_and_a_plugin_is_found_by_its_target checks the
failed active row, plugin_manager_toggles_checks_and_updates exercises
enabling and disabling, a_configured_typescript_plugin_is_listed_as_skipped_with_its_reason
opens the skipped row's error, and the_manager_asks_before_a_project_plugin_runs
opens the pending row's approval question.
Move the version abbreviation assertion to the_server_rows_follow_typescript
on both carriers. locale_compare_matches_bun already checks the exact
three sort inputs. Keep the independent slow-report cases.
transcript::the_footer_measures_throughput_from_the_first_token now
drives the two-step turn for modern, mixed and legacy timestamp data.
Assert the displayed rate including reasoning tokens and the absence
of a separate first-token label. This replaces
measures_decode_throughput_from_the_first_streamed_token and its
private typed-message builders. The cross-page footer test still
checks turn duration from the user's prompt.
standalone_mentions_survive_unicode_and_reject_word_and_path_contexts
checks both bare and prefixed IDs across mixed-width separators and
word and path contexts. It replaces the hand-picked ASCII and Unicode
neighbor examples and checks both the reported span and decoded ID.
Keep the seeded arbitrary-piece test as a separate no-panic property.
session_links::code_spans_link_only_when_they_are_exactly_an_id now
renders padded double backticks and a punctuated ID in both Markdown
modes. It replaces a_code_span_is_a_mention_only_when_it_is_exactly_one;
loaded_sessions_show_their_titles_and_the_rest_stay_ids covers bare IDs.
navigate::message_navigation_aligns_messages_with_the_top now includes
a reasoning-only response between two navigable messages. Check the
rendered row after stepping across it in both directions. The same
test already covers user-only jumps, viewport alignment, repeated
jumps at the head, and slack at the tail.
Remove boundaries_step_from_the_view_or_the_last_jump, which builds
another set of messages and calls the geometry helpers directly.
the_tab_window_fits_the_strip_and_tracks_selection covers the crowded,
roomy, active-tab and new-session-slot examples across strip widths
and selections. moving_inside_the_window_or_returning_home_keeps_its_start
keeps the sticky-window and home cases. The rendered strip remains
covered by tabs and the app tab-strip recording.
closing_and_reopening_any_tab_restores_its_position_and_title covers
closing, losing a tab to another client and reopening the newest tab.
Check the round trip at every position, including the last tab.
A camelCase name above its Rust counterpart adds no contract. Remove
those labels from prompts, footers, menu rows and theme code. Keep
actual protocol keys and comments that explain behavior.
Keep input precedence, immutable scrollback, reconnect handling,
palette fallbacks and terminal limits. Remove migration inventories,
review labels and comments that narrate a test or name its TS file.
Keep the CLI's non-obvious contracts: open log handles during
compaction, restart ownership, child signal handling and quoting.
Drop file coordinates, checklist labels and test narration.
Document what a reconnect invalidates, how writers share config files,
and when a lock can be broken. Remove review labels, port inventories
and implementation history from the modules and their tests.
Keep discovery and replacement policy, event delivery order, PTY frame
layout and the no-retry rule for writes. Remove port ownership notes,
review labels and source coordinates from the service, API, ACP and
drive crates. No test is removed.
Keep the dispatch, loading and resource-limit rules in terms of what the host does. Remove section numbers that no longer resolve to maintained documentation. The native rollback test and scoped clippy for both plugin crates passed.
Document the separate depth, backlog and paint budgets, including why an idle interval can reset the count without a repaint. Both effect-queue regression tests passed against the landed frame-effects implementation.
Replace command-by-command narration with each example’s purpose. Keep fixture behavior, actual feature names and the SDK quick start intact. Both C and Rust transcript replays passed, as did the Luau prelude test and check:luau-types.
Keep ownership, watch lifetimes, input ordering and backpressure where they explain the implementation. Drop catalog findings and test narration. Regenerating the bundle produced identical bytes; all 78 compatibility tests passed.
Describe message ordering, bounded delivery and VM table charging directly. Preserve the safety preconditions and the observable protocol. The pipe queue and Luau decoder comparison passed.
Keep the constraints that govern mount generations, retained trees, input and resource lifetime. Remove rollout narratives and reviewer annotations. The frame-effects region was excluded. The layout equivalence and storage tests passed.
Keep the file-descriptor, quarantine, dependency and crash-record constraints beside the code they govern. Drop step names, reviewer annotations and TS line numbers. Nine loader and record tests passed.
Put allocator ownership, table negotiation, effect ordering and panic handling first. Describe the current contracts for surfaces, geometry and worker threads. Schema and task checks passed (31 tests).
Keep the source and cell budgets, streaming fallback and styling rules. Drop migration notes and descriptions that repeat the layout tests. The layout oracle and 21 renderer tests passed.
The update notifier needs an updater and remembers dismissed notices.
State that directly. Remove the remaining project labels and test
prologues found by scanning the combined TUI comments.
Replace handoff history and reviewer IDs with the distinctions that
matter to the fixture: ownership, carrier behavior, resource bounds,
and observable delivery order. Keep model descriptions and public
test-helper documentation.
Describe the setup that affects what a test observes: local time, frame
delivery, dynamic registration, file writes, and reference inputs.
Drop issue inventories and assertion walkthroughs from the fixtures.
Keep the contracts governing approval, storage limits, stale replies,
and focus across remounts. Drop implementation milestones, reviewer
names and source coordinates from plugin and diff viewer modules.
Keep the reasons drafts survive remounts, edits cancel pending submits,
and writes run in order. Replace implementation history and TypeScript
line references with short descriptions of the behavior.
Describe what owns focus, how transcript windows and drafts survive
updates, and when terminal output can resume. Remove porting row IDs
and source coordinates from the view modules.
Document input precedence, connection ordering, terminal restoration and
shared state where they constrain the implementation. Remove porting
milestones, review credits and source-line citations from the shell,
feature services, client MCP and panel modules.
Remove what a reader does not need to use, build, test or extend the client, and merge what two documents said twice: the plugin limits, request ownership, Bun lookup and authoring commands now live in one place each. The README links to the docs for parity commands and the crate layout instead of repeating them, and the manual Windows checklist becomes a section of testing.md.