Repository navigation
feat(statusline): context row options and a session token totals row - #541
Merged
Merged
Conversation
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
|
This security review couldn’t start because the paying account has exhausted its Codex allowance and has insufficient available credits. The payer can check their balance and add credits in the Codex usage dashboard, then try again. If you do not manage the paying account, contact this repository's admins. |
A context settings group chooses whether the row draws its bar or only its figures, the figure (tokens held against the window and the percentage, either alone, or the percentage left), the warning and critical thresholds of the context bar, a window to measure the tokens held against, what to draw when the client reports no percentage, and a breakdown of the last request once use reaches the critical threshold. A threshold left at 0 follows the bands, and the estimate defaults to never, so the default output is unchanged.
A transcript further ahead than one read took in was read from its end, so every fact folded from it (compaction count, name, start, prompt cache clock) was missing the part skipped. A refresh now reads one window from where the cache ends and folds its lines in order, so the facts are exact once the cache reaches the end of the file and marked incomplete until then, with the work of each refresh still bounded by one window. A line bigger than a window is passed a window at a time without being folded, which leaves the facts incomplete for good and the session's start unknown if no timestamp had been seen.
The tokens row sums the usage of every response in the session's transcript and draws the total with its input, output and cache shares. It is a row a layout asks for, not one of the defaults. The payload carries only the context window's current load, so the totals come from the transcript, through the reader the session rows share: a new key in its cache beside the session facts, folded from the same lines. A response is written as several records that repeat its usage, so only the last response is remembered and each later record adds its growth. While the reader is catching up with a transcript too big to read at once the total ends in a plus, as the compaction count does.
… exact A reported 0 beside tokens already held is kept when no estimate can be made, instead of making the row vanish. A window set in the settings draws nothing until the client reports the tokens held, rather than the percentage against the client's own window. A tokens figure with no bar and nothing to show draws nothing instead of a bare label. The context thresholds and the window are now absent until set, not 0, so a layer unsets them with --unset-statusline and they follow what they fall back on. Optional keys are listed as having no default. The estimate description states exactly when it applies.
…ayer set An optional statusline setting (the context thresholds and window) may be written as null in a layer, which removes it while layers merge, so a deeper layer can return it to unset. profile set takes none for such a setting, as against --unset-statusline, which only removes it from that one layer. The layer schema, the key listing, the generated JSON schemas and the docs state which settings have no default.
Mearman
force-pushed
the
feat/statusline-context
branch
from
October 20, 2026 11:35
e9beb5c to
e014cf6
Compare
…e catching up While the reader catches up with a transcript bigger than a window, the facts that are the newest value of something (the latest response, the names, the prompt cache's clock) came from the first window and stayed stale for as many refreshes as the catching up took, with nothing to say so. The session state is now two parts: the cumulative facts (start, compactions), which are exact only once every record is read, and the latest-state facts, which the end of the transcript settles. While the cumulative part is behind, the latest-state part is also folded from the last window of the transcript and carried in the cache, advanced over what is appended, and its set facts replace the first part's in what the reader reports. A refresh reads at most two windows.
…defaults behave The compactions row says what is unread while the reader catches up and what a skipped line leaves for good, and the statusline paragraph names the keys that have no default without counting them.
The colour of the figure and the breakdown are checked at a percentage exactly on the warning and critical thresholds and a hair under them, and the bar glyph is written as an escape so the source stays ASCII.
…hing up When the end of a long transcript holds a request but no cache write that names a tier, the tier the first window showed is kept, as a full read of the transcript keeps it, instead of falling back to the default lifetime. The unused helper that listed the end's set facts is removed.
… at once The latest-state facts are right at once only when the last window of the transcript holds a record that sets them, and otherwise stay as the part read so far has them until the reader reaches one. A refresh reads at most two windows and a byte, beside the check of the first line.
The tier an early cache write showed, a title set only in the middle, a last window that starts exactly on a line, and a transcript cut to a length between the cumulative fold and the fold of the end are read as a full scan would read them.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Gives the built-in statusline's context row claude-hud's context options. A new
contextsettings group sets whether the row draws its bar or only its figures, the figure itself (tokens held over the window and the percentage, either alone, or the percentage left), separate warning and critical thresholds for the context bar, a window in tokens to measure against instead of the model's, what to estimate when the client reports no percentage (nothing, the raw share, or the share with the compaction reserve added), and the token breakdown of the last request once use reaches the critical threshold. A newtokensrow sums the session's usage from its transcript.Defaults draw byte for byte what they did. That shaped three choices that differ from claude-hud's own defaults:
context.warnAtandcontext.criticalAtdefault to 0, which followsbands.warnAtandbands.criticalAt, so a layer that only setsbandsstill colours the context bar as before;context.estimatedefaults tonever, because claude-hud's estimate would draw a bar before the first response of a session where the line has always drawn nothing; andcontext.breakdownand thetokensrow are off until asked for. The schema has no boolean kind and I did not add one here, so the two switches are named choices (shown/hidden,never/critical), and a 0 stands for "not set" in the three integer keys that need it.claude-hud takes session token totals from its transcript parser, and the payload has no cumulative figures (
total_input_tokensis the context window's current load, as the Claude Code statusline reference says). So thetokensrow reads the transcript named bytranscript_paththrough a new port. A response is written as several lines that repeat its usage, adjacent among the lines that record usage in every transcript I checked, so only the last response is remembered and each later line adds its growth. The reader is incremental: where the last read ended, and the open response, are kept in a small file per transcript under<home>/statusline-cache/(written atomically, removed after the usage log's retention when a new session's file is first written), so a refresh reads only the bytes appended since. A transcript that shrinks is read again from the start. I compared the first read of a large transcript against a separate script that sums the same fields and the totals matched. Other groups that need transcript data (activity lines, cost) will want the same reader; I kept it behindStatuslinePorts.sessionTokensso they can fold it into whatever shape they share.realStatuslinePortsnow takes the cache directory, which the command derives from agent-shim's root.