The week strip: the 7d period as its 5h windows - #12
Closed
lroolle wants to merge 14 commits into
Closed
Conversation
…input jq's `catch .` binds the error MESSAGE as the value, not the original input. Snapshots with an empty or unparseable five_hour.resets_at all normalized to the same error string — one shared fake window identity — so the pct_per_window pair walk could pair samples across real windows. Observed on real data: phantom window transitions and a skewed ratio (9.91 -> 10.06 after the fix on the same log). Normalize via a named def that treats empty as empty and falls back to the raw string on parse failure. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
`statusline.sh report [--days N]` replays usage.jsonl and ledgers every closed window: 7d closes as used%/expired% (converted to 5h-windows-worth via the learned pct_per_window ratio), 5h closes as count/avg/capped, plus the week in progress projected with the same learned walk the advisor uses — the surfaces cannot disagree. A window closes when consecutive samples disagree on resets_at, normalized to the minute (the ratio learner's identity rule). The advisor prevents waste prospectively; this proves it retroactively. Every usage snapshot now also logs limits[] (scoped weekly caps, verbatim), model (id active in the logging session), and predicted_end — the learned walk's projection at sample time, the calibration seed that lets closed windows score the forecast later. Learning lags logging by weeks; a field absent today is a pattern that can't be learned next month. Subcommands dispatch after all function definitions, take no stdin, and are read-only against the state dir. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
CHANGELOG entry, README section with a real ledger example, state-dir contract gains the widened usage.jsonl fields (limits, model, predicted_end) and names report as the reference consumer, counts synced to 327 tests across README / llms.txt / site. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Stacked PRs (base = a feature branch) got no checks under the branches:[main] filter; judgment should not depend on where a PR points. Push triggers stay main-only. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The statusline never runs when you're away — exactly when expiring capacity needs a voice. Instead of a daemon, `check` exposes the advisor's judgment as an exit code (0 calm / 1 opportunity / 2 pressure / 3 unknown-or-stale) with the plain verdict on stdout, for tmux segments, cron notifiers, and scripts. Model context for the scoped clauses comes from the last logged snapshot — the first consumer of the widened `model` field. `session-summary` gives one-line session retrospectives from the usage log, designed as a SessionEnd hook (hook JSON on stdin, only session_id read; bare runs fall back to the last logged session). Window deltas are positive-delta sums — the profile builder's rule — so a session straddling a 5h reset still reports what it consumed. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
README scripting section (cron + tmux + SessionEnd hook recipes), CHANGELOG entry, counts synced to 334 tests. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Three layers, one source of truth: line 1 shows the numbers, the advisor row says the one sentence, and this skill carries the full conversation — should I start a heavy task, which account has headroom, what did I waste this week. It encodes the state-dir contract, the learned-forecast semantics, and the advisor's judgment rules (feasibility before advice, facts only while unlearned, staleness said out loud), and runs the same report/check subcommands instead of re-mining — so the layers cannot disagree. Install: cp -r skills/usage-insight ~/.claude/skills/ Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Two taste fixes for the two-line render. A session traced by cctrace (deva --trace, or cctrace directly) now wears a dim [cctrace:PORT] chip on the LEFT, next to path and branch: being recorded is session identity, so it lives in the structure lane — red stays reserved for pressure. Detection takes the strongest signal available: the trace env cctrace exports into its child (CCTRACE_SERVER_PORT), the capture's CA plumbing (NODE_EXTRA_CA_CERTS under a cctrace dir), or deva's DEVA_TRACE=1. Plumbing-only captures (older cctrace) resolve the port through the live-instance registry — session id first, then project path, then only-live-capture; tombstones never lend their port. No resolvable port still shows a bare [cctrace]. The advisor row anchored on term_width - 5, but under a pipe `tput cols` answers a flat 80 whatever the terminal is — line 1 overflowed the phantom edge while the advice dangled mid-line beneath it. The row now anchors on line 1's ACTUAL rendered width (line1_cols, recorded by format_output), so the second line's right edge meets the first's even when the width guess is wrong. Width detection itself got honest too: the controlling tty (stty size </dev/tty) and an inherited COLUMNS both beat tput's default. Tests: 12 new (8 chip units, 2 chip integrations, 2 edge-alignment integrations); helpers now scrub the trace env the way they scrub account identity — this repo's own dev loop runs under deva --trace, which is exactly how the leak would happen. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Live testing against this repo's own traced dev session caught two
registry realities the first cut missed. cctrace REDACTS session ids
before they land on disk ("c3a6e0f3-****-…"), so an exact-id match
never fires against a real registry — match on the sid8 prefix, the
same join key cctrace's own UI uses. And crashed captures leave
non-tombstoned "live" entries behind for up to a day, so the project-
path and only-live fallbacks now trust heartbeat-fresh files only
(<2min against a 30s heartbeat); the sid8 match keeps trusting any
non-tombstone entry — if OUR capture died, this session's proxy died
with it.
Tests 346 -> 348: the sid test now stores a redacted id, plus
stale-entry coverage both ways (sid8 claims its own stale entry;
fallbacks refuse a stranger's).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The 7d window becomes a `week` subcommand: a 56-cell timeline (8 cells/day) where fill is budget consumed on a time axis, `│` marks now, `▒` is headroom to the clock, and `▓` is usage running ahead of it. The day ruler anchors to the account's own reset weekday — no pretending every week is Mon–Sun — and shares the reset formatters' zone, so the labels can't drift a weekday from the boundaries they describe. The advisor renders underneath in always-mode; the strip is the prospective glance beside `report`'s retrospective ledger, and the same one claudex's watch mode draws. State dir only; stale data renders but says so; no cache or no active window exits 3 like `check`. The calm advisor line now speaks the shared budget frame: `- budget ~19x5h left · even 1.1%/win · heading ~52%`. "budget" names the frame, "pace" stops doing double duty (it means usage/elapsed everywhere else), and in the last window — where per-window math just restates the headroom — it degrades to `- budget last window · 61% left · heading ~40%`. Tests 348 -> 352: strip geometry against a fixed clock, ruler anchoring, last-window degradation, and the exit-3 contract. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The 56-cell time axis told you when the week ends; it could not tell you where the week went. Redrawn: the 7d period as its own 5h windows, one cell each (34, the last a 3h stub). ▁▂▃▄▅▆▇█ a window that ran, height = the 7d points it burned; · ran but under 1%; ░ unknown, no samples on record; ▮ the window you're in; ▫ still ahead; × a window the pool won't cover at the current pace — the same wall claude.py draws, learned forecast when trained, linear projection otherwise. Past cells are reconstructed from usage.jsonl: a window instance is keyed by its 5h resets_at rounded to 5min (the API jitters it, and 05:59:59/06:00:00 are one window), costed by the 7d movement observed inside it. ░ and · stay different glyphs because a gap in the record is not an idle session, and drawing it as one is the lie this row must not tell. Count ▮ and what follows and you get the budget line's own ~Nx5h left, so the picture and the sentence under it are one number. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Member
Author
|
Absorbed: feat/shared-store-contract-v2 (#15) carried the week subcommand, tests, and changelog; the remaining README budget-row wording landed directly on main. Nothing left in this diff. |
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.
What
A new
weeksubcommand: the prospective glance besidereport'sretrospective ledger. First landed as a 56-cell time-axis strip, then
redrawn in this same PR as what it wanted to be: the 7d period as its
own 5h windows, one cell each (34 per period, the last a 3h stub).
▁▂▃▄▅▆▇█·░▮▫×Past cells are reconstructed from
usage.jsonl: a window instance iskeyed by its 5h reset (rounded 5min — the API jitters it), costed by
the 7d movement observed inside it.
░and·stay different glyphson purpose: a gap in the record is not an idle session, and drawing it
as one is the lie this row must not tell.
Count
▮and what follows and you get the budget line's own~Nx5h left— the picture and the sentence under it are the same number. The×wall uses the learned forecast when trained, linear projectionotherwise — the same wall claudex's
claude.pydraws, so bothsurfaces tell one story.
Also: the calm advisor line now speaks the shared budget frame
(
- budget ~19x5h left · even 1.1%/win · heading ~52%), withlast-window degradation where per-window math stops meaning anything.
Contract
Reads the state dir only; stale data renders but says so; no cache or
no active window exits 3 like
check.Verify
npm exec --yes bats -- t/— 354/354.Stacked on #9 (base
feat/trace-chip).🤖 Generated with Claude Code