chad/sweatshop-legacy
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

0051 — Context-window meter per chat #131

Merged
chad merged 7 commits from glen/context-meter into main 2026-10-18 22:30:28 -04:00
Conversation 1 Commits 7 Files changed 16 +642 -16
glen commented 2026-10-18 22:17:59 -04:00
Collaborator
Copy link

Implements docs/0051-context-meter.md (approved plan v1 in the ticket thread): ROADMAP item 7, the per-chat context-window meter — a percentage-first chip in the chat header ("oops, context is at 80% — wrap this up", per Chad).

What lands

  • Capture — migration 0056_context_meter.sql adds last_input_tokens/last_cached_tokens/last_output_tokens/last_reasoning_tokens (nullable) to usage_events; turn.ts keeps the turn's LAST slice's counts beside the summed usage and recordUsageStep persists them. The last slice's input re-sent the whole conversation, so lastInput+lastOutput is the conversation's size in tokens. Reasoning is NOT added on top (schema convention: it's a display-only slice of output).
  • Read — readContextUsage (SQL ORDER BY ended_at DESC LIMIT 1, pre-0051 rows excluded outright); readConversationStats gains context {usedTokens, windowTokens, pct} for THIS conversation (never the group), window resolved through the model catalogs. Null → hidden when unmetered or the model is unknown.
  • UI — components/context-meter.tsx: mono percent + thin fraction bar in the chat header, warning ≥85%, destructive ≥95%, exact numbers in the tooltip; refetches on turn settle with a delayed follow-up (the usage row lands after the stream finishes). The stats drawer gains a fuller readout section.
  • Docs — 0051 design doc; AGENTS.md UI bullet line; ROADMAP item 7 marked done.

Verification

  • Gate: npm run typecheck && npm run lint && npm run test — tsc clean, eslint at exactly the 2 pre-existing warnings, 774 passed / 13 skipped (11 new tests: store latest-row/pre-0051/harness exclusion, turn single- vs multi-slice last-slice pin via the steer mock, API present/absent/clamp-at-1/zero-token/unknown-model/tie-break).
  • Migration: fresh-db and main(0055)→branch(0056) upgrade paths both verified, PRAGMA foreign_key_check clean. The upgrade-path check caught a real bug pre-commit: the 0056 journal entry needs when > 0055's future-dated timestamp (drizzle applies only entries newer than the last applied) — the first version silently skipped.
  • Visual QA: preview with seeded metered rows — 92% warning-zone chip, drawer readout 117,600 / 128,000 tokens, post-fix 50% neutral chip; screenshots in the ticket thread.

Fresh-eyes review (fix-first verdict, no blockers)

Both should-fixes taken:

  1. Settle-race — the chip's settle-edge fetch raced the usage write (stream finish fires before recordUsageStep); a one-shot 1.5s follow-up fetch after a turn ran closes it. (Lint then caught my ref-during-render — moved into an effect.)
  2. Doc double-count trap — 0051's semantics paragraph read as adding reasoningTokens; fixed the wording (code was already right).

Nits taken: SQL pick with isNotNull (dead tie-break code gone), pct comment matches reality, drawer copy says "that turn's model's window", plus 3 new guard tests (clamp/zero/latest-row-vs-newer-pre-0051). Nits declined: shared formatTokens helper (two components, different display intents — refactor when a third appears).

Deliberately out of scope (0051 records the decisions): compaction/fresh-thread affordances (warn-only v1), backfill (forward-only per 0047), per-harness-call meters.

Implements docs/0051-context-meter.md (approved plan v1 in the ticket thread): ROADMAP item 7, the per-chat context-window meter — a percentage-first chip in the chat header ("oops, context is at 80% — wrap this up", per Chad). ## What lands - **Capture** — migration `0056_context_meter.sql` adds `last_input_tokens/last_cached_tokens/last_output_tokens/last_reasoning_tokens` (nullable) to `usage_events`; `turn.ts` keeps the turn's LAST slice's counts beside the summed usage and `recordUsageStep` persists them. The last slice's input re-sent the whole conversation, so lastInput+lastOutput is the conversation's size in tokens. Reasoning is NOT added on top (schema convention: it's a display-only slice of output). - **Read** — `readContextUsage` (SQL `ORDER BY ended_at DESC LIMIT 1`, pre-0051 rows excluded outright); `readConversationStats` gains `context {usedTokens, windowTokens, pct}` for THIS conversation (never the group), window resolved through the model catalogs. Null → hidden when unmetered or the model is unknown. - **UI** — `components/context-meter.tsx`: mono percent + thin fraction bar in the chat header, warning ≥85%, destructive ≥95%, exact numbers in the tooltip; refetches on turn settle with a delayed follow-up (the usage row lands after the stream finishes). The stats drawer gains a fuller readout section. - **Docs** — 0051 design doc; AGENTS.md UI bullet line; ROADMAP item 7 marked done. ## Verification - Gate: `npm run typecheck && npm run lint && npm run test` — tsc clean, eslint at exactly the 2 pre-existing warnings, **774 passed / 13 skipped** (11 new tests: store latest-row/pre-0051/harness exclusion, turn single- vs multi-slice last-slice pin via the steer mock, API present/absent/clamp-at-1/zero-token/unknown-model/tie-break). - Migration: fresh-db and main(0055)→branch(0056) upgrade paths both verified, `PRAGMA foreign_key_check` clean. **The upgrade-path check caught a real bug pre-commit**: the 0056 journal entry needs `when` > 0055's future-dated timestamp (drizzle applies only entries newer than the last applied) — the first version silently skipped. - Visual QA: preview with seeded metered rows — 92% warning-zone chip, drawer readout `117,600 / 128,000 tokens`, post-fix 50% neutral chip; screenshots in the ticket thread. ## Fresh-eyes review (fix-first verdict, no blockers) Both should-fixes taken: 1. **Settle-race** — the chip's settle-edge fetch raced the usage write (stream `finish` fires before `recordUsageStep`); a one-shot 1.5s follow-up fetch after a turn ran closes it. (Lint then caught my ref-during-render — moved into an effect.) 2. **Doc double-count trap** — 0051's semantics paragraph read as adding reasoningTokens; fixed the wording (code was already right). Nits taken: SQL pick with `isNotNull` (dead tie-break code gone), pct comment matches reality, drawer copy says "that turn's model's window", plus 3 new guard tests (clamp/zero/latest-row-vs-newer-pre-0051). Nits declined: shared `formatTokens` helper (two components, different display intents — refactor when a third appears). Deliberately out of scope (0051 records the decisions): compaction/fresh-thread affordances (warn-only v1), backfill (forward-only per 0047), per-harness-call meters.
ROADMAP item 7. Last-slice tokens on the existing usage_events row,
context block on the stats read path, percentage-first chip in the
chat header. Warn-only near the limit; forward-only metering.
Migration 0056 adds last_input/cached/output/reasoning_tokens
(nullable, forward-only per 0047). turn.ts keeps the last slice's
counts beside the summed turnUsage; recordUsageStep persists them.
readContextUsage reads the conversation's latest metered chat row.
readConversationStats gains context {usedTokens, windowTokens, pct}
from the conversation's own latest metered chat row (never the
group's) against the catalog window. Null when unmetered/unknown.
The percentage leads (Chad's call): mono % with a thin fraction bar
in the chat header, warning tone at 85%, destructive at 95%. The
drawer gains a fuller readout. Hidden until metered turns exist.
Store: readContextUsage latest-row pick, pre-0051 rows skipped,
harness rows never answer. Turn: single-slice equals last-slice;
multi-slice records the LAST slice, not the sum (steer mock, two
slices, one executor per file — the queue targets one base URL).
API: 0.8 at 102.4k/128k; null when unmetered/pre-0051/unknown model.

Journal fix caught by the upgrade-path check: the 0056 entry needed
drizzle's real shape (breakpoints:true, no sql key) and a `when`
above 0055's future-dated timestamp — the migrator applies only
entries with when > last applied. Verified main→branch applies.
fix(usage): review fixes — settle-race follow-up, SQL pick, doc double-count trap
All checks were successful
deploy / checks (pull_request) Successful in 42s
Details
deploy / deploy (pull_request) Has been skipped
Details
c96fb3cae7 
Fresh-eyes review (fix-first verdict; no blockers):

1. Settle-edge fetch raced the usage write (stream finish fires BEFORE
   recordUsageStep): one delayed 1.5s follow-up fetch after a turn ran.
2. Doc said to ADD reasoningTokens; per the schema's own convention
   reasoning is a display-only slice of outputTokens (already inside) —
   doc fixed, code unchanged.
3. readContextUsage: SQL ORDER BY ended_at DESC LIMIT 1 with
   isNotNull(lastInputTokens) — pre-0051 rows excluded outright, no
   dead tie-break code.
4. pct comment matches reality (clamped, unrounded); drawer copy says
   'that turn's model's window' (accurate after a route switch).
5. Tests: clamp at 1 over window, zero-token null, latest-row pick
   with a newer pre-0051 row never winning.

Lint caught the ref-during-render my first fix had — now in an effect.
Gate: tsc clean, eslint at the 2 pre-existing warnings, 774 passed.
chad approved these changes 2026-10-18 22:30:22 -04:00
chad merged commit 5f24ec5746 into main 2026-10-18 22:30:28 -04:00
Sign in to join this conversation.
Reviewers
No reviewers
chad
Labels
Clear labels
No items
No labels
Milestone
Clear milestone
No items
No milestone
Projects
Clear projects
No items
No project
Assignees
Clear assignees
No assignees
2 participants
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference
chad/sweatshop-legacy!131
No description provided.