GeezerVison

chatgpt-codex-connector Bot left a comment

chatgpt-codex-connector Bot left a comment

chatgpt-codex-connector Bot left a comment

chatgpt-codex-connector Bot left a comment

Adds groupWorkingBlocks, which folds a turn's consecutive tool-step rows
into WorkingBlock descriptors the timeline can render as one disclosure.
A step row is assistant output whose last visible block is a tool call or
reasoning; answers, prompts, hook notices, and rows carrying
ask_user_question, propose_plan, or compaction markers never fold.

Durations are the wall-clock span of part timestamps taken from the raw
messages, so hidden tool-result messages count and parallel tools are not
summed. The stream state now records tool part timestamps and the earliest
part time so a live block can start its clock at the first reasoning delta.

Drop the unread createdAt fields on stream tool calls and results, the
rowKey accessor, the unreachable NaN branch in formatWorkingDuration, and
the binary search over per-turn spans. Name getStepRowContent for what it
returns. The grouping test reuses the shared fixture clock and stream
builder instead of redeclaring them; the stream start test moves next to
the code it covers.

A user message made only of context files or skills has no timeline row,
so steps on either side of it folded into one block and the block before
it counted as live. Grouping now reads user boundaries from the raw
entries, anchoring the next turn's live key on the hidden message.

…and partial blocks

A metadata-only user message now flushes the read_file merge so reads
from two turns never share a row, and a block whose loaded page starts
with such a prompt is no longer reported as partial.

A turn that follows a metadata-only prompt can open with provider-executed
parts, which carry timestamps but render no row, so the preceding block's
duration ran into the next turn. Cap the span at the next user message.

Collect part timestamps into one list per block instead of two running
min/max closures, trim restating docs, drop grouping tests that repeat
coverage, and leave formatWorkingDuration to the disclosure that uses it.

…e the visible row content filter

Working block grouping reuses the messageHelpers filter instead of a
drifted copy, and drops two no-op guards in its timestamp math.

…rking block comments

jaaydenh left a comment

…nk lines between working block steps

…rking block type docs

…ext-only user messages inside the turn

Users cannot send context-file or skill-only messages; the only ones in
history were inserted mid-turn after create_workspace by the removed
persistInstructionFiles. Treating them as prompts split one turn into
two working blocks, so they are invisible to grouping again.

…working block timestamps in one pass

Each block filtered every loaded entry, which is O(blocks x entries).
Entries arrive sorted by message ID and block spans are disjoint and
ascending, so one forward cursor collects every span in a single walk.

…lify stream startedAt

chatd stamps stream parts at publish time in publish order, so the first
timestamped reasoning or tool part is the earliest. Set startedAt from it
in place instead of taking a running minimum in a wrapper.

… undefined timestamps at the source

Optional part and stream timestamps are filtered where they are read, so
the parse step only sees strings. Empty live rows now share the single
empty-blocks check in getStepRowContent.

…nt-steps-grouping

… counting failed steps in working blocks

The summary no longer shows a failed step badge, so blocks track only the
distinct tool IDs they hold.