v2.1 — Workflow attribution, usage tracking, AI advice + Optimizer, dashboard settings - #39

… 5 scopes)

Phase 4 reworked: usage attribution panel modelled on the official /usage
'What's contributing to your limits usage?' screen — characteristics
(>150k context, long-running sessions, subagent-heavy, workflow share)
plus Skills/Subagents/Plugins/Models tables, scoped Day/Week/Month/
Session/Project. Compact strip in Today tab; full panel in Content tab;
Phase 5 advice consumes the signals. Documented agent-*.meta.json
agentType values verified on disk. Checkpoints now after Phases 2 and 4.

Phase 0 of docs/V2.1-WORKFLOW-SPEC.md — kills the intermittent
'terminated' / infinite-hang failures of Get AI Usage Advice:

- new shared src/httpClient.ts: fetch transport with AbortController
  timeout (clean 'timed out' error) + curl transport, factored out of
  claudeApiClient.ts (which now delegates; behaviour unchanged)
- advisor.ts: 120 s default timeout, one fetch retry, then curl
  fallback — same last-resort transport the quota client uses;
  verified against a black-holed host: clean failure in 3x timeout,
  no hang
- buildAdviceSummary: cap prompt sample at 40 prompts x 1500 chars,
  skip the section entirely above 80 KB
- prompt purity: collectPrompt no longer harvests from sub-agent
  logs, meta/sidechain lines, or agent-framework scaffolding text
  (kebab-case XML-ish wrappers, interrupt notices)
- dataLoader.ts: replace a literal NUL byte in a string key with its
  unicode-escape form so tooling stops treating the file as binary

Phase 1 of docs/V2.1-WORKFLOW-SPEC.md:

- records from files under subagents/ now carry _agentId and _agentType
  (from the sibling agent-*.meta.json, cached per refresh); files under
  subagents/workflows/ additionally carry _workflowId and _workflowName
  (resolved from workflows/scripts/<name>-wf_<id>.js)
- attribution is path-derived, never cwd-derived: worktree-isolated
  agents have a cwd pointing at a temporary git worktree
- new ClaudeDataLoader.getWorkflowBreakdown(): groups by workflow run,
  aggregates via calculateUsageData, per-agent drill-down, sorted by
  most recent activity
- new WorkflowUsage type

Verified against real data: 4 workflows with resolved names
(deep-literature-review etc.), agent counts 22/9/7/6, agentType split
workflow-subagent=412 / general-purpose=213 / Explore=143. Spec
acceptance updated: the 24.6M raw figure double-counts proxy
placeholder+final pairs; the deduped ~8.9M is the consistent number.

…rate

Phase 2 of docs/V2.1-WORKFLOW-SPEC.md:

- new tab after Branches: one sortable row per workflow run (start time,
  derived name, project, agent count, cost, token split, cache hit rate,
  duration), expanding to a per-agent sub-table via the existing
  project-group toggle pattern (sort-child rows travel with their parent)
- summary strip: workflows this month / their cost / share of month cost
- cache hit rate (cacheRead / input-side tokens) is the headline
  diagnostic; hint line under the table explains the native-vs-proxy
  story. Verified on real data: native Claude runs show 76%/72%,
  DeepSeek-routed runs 0.0% - exactly the signal the tab exists for
- i18n keys in all six languages; CHANGELOG [2.1.0]; version bump 2.1.0

Smoke-tested by rendering the tab against live ~/.claude data (vscode
stubbed): 4 workflows, 44 agent child rows, 11 sortable headers.

Phase 2 checkpoint follow-up:

- per-agent drill-down rows are now labelled by the task dispatched to
  the agent (first user message of its log, captured at load time as
  _agentTask, truncated 200 chars) with the hex id demoted to a second
  line + tooltip - expanding a workflow now tells you what each agent
  actually did
- new Model column on workflow and agent rows: distinct models used,
  most expensive first, shortened (claude-/date-suffix/[1m] stripped),
  full ids in the tooltip; sortable
- workflow-count question investigated: the disk only contains 4
  wf_* run dirs total (~/.claude/projects scanned recursively), so
  nothing is being missed locally - other remembered runs were cloud
  routines / other machines / generic Task-tool fan-out (49 generic
  agent logs found), which Phase 4's Subagents table will surface

Smoke-verified: 44/44 agent rows carry task labels, 12 sortable
headers, model cells render (deepseek-v4-pro etc.).

Phase 3 of docs/V2.1-WORKFLOW-SPEC.md:

- content analyzer now tracks estimated thinking vs. total assistant
  output per session and per local day (thinkingBySession/thinkingByDay
  on ContentAnalysis)
- Sessions tab: sortable "Thinking %" column; >60% gets a warning glyph
  and an /effort-high hint in the cell tooltip (no popups); hint line
  marks the figures as text-length estimates
- Today tab: one-line estimated thinking share for the day
- quota guard: dismissible warning banner in the dashboard when the
  5h window's remaining share drops below workflowQuotaWarnPercent
  (new setting, default 50, 0 disables); never warns on an
  already-reset window; dismissal lasts for the window-session;
  status bar untouched. Quota now also flows to the webview (only
  re-renders on change).

Smoke-verified on real data: 37 sessions with thinking data (top
64.6%), today line ~8%, banner shows at 20% remaining and hides on
enough quota / expired window / dismiss.

Phase 4 of docs/V2.1-WORKFLOW-SPEC.md (incl. mid-phase review feedback):

- content analyzer detects skill invocations (Skill tool_use blocks +
  <command-name> echo markers) with session/day/timestamp; trivial
  session-management commands (/model, /clear, ...) are excluded
- ClaudeDataLoader.getUsageAttribution(records, analysis, scope):
  scopes day/week/month/session/project; characteristics weighted by
  estimated cost (>150k-context share, 8h+-active-session share,
  subagent-heavy-session share, workflow share); tables for Skills,
  Subagents (by agentType), Plugins (namespace prefix), Models
- skill/plugin share methodology matched to the official screen after
  comparing real numbers: usage at/after the invocation in the same
  session counts toward the skill (overlapping independent signals),
  NOT the injected-prompt size (which measures ~0% and is kept only
  as a secondary column)
- Content tab: panel with scope selector (Session/Project reveal a
  target dropdown; recompute via webview message round-trip); top
  skill/plugin >=10% get official-style characteristic lines + hints
- Today tab: compact strip with >=5% characteristic lines only
- i18n: 22 new keys in all six languages

Smoke-verified on real data (week scope): 94% >150k context, 63%
8h+ sessions, 40% code-review, 14% plugin superpowers; today strip
shows 94% superpowers:executing-plans (this implementation session).

Phase 4 checkpoint review feedback:

- workflow-count gap root-caused: locally there were only 4 wf_* runs;
  the other remembered ultracode runs were ad-hoc Task-tool fan-outs
  (no wf_ dir - observed when the dynamic-workflow feature isn't
  engaged, e.g. proxy/DeepSeek routing) or mere text mentions of the
  word. getWorkflowBreakdown now also groups >=2 generic agents per
  session as an ad-hoc batch (badge "subagents", named by session
  title). Real data: 4 wf runs + 8 batches, incl. a 17-agent DeepSeek
  refactor run that was previously invisible
- agent drill-down: the agents' shared task boilerplate is hoisted
  into one pinned row per run; each agent row now shows only the part
  of its task that differs (hex id demoted, full text in tooltips)
- attribution panel: scope switcher is now tab buttons on the same
  line as the heading (Session/Project reveal a dropdown there);
  heading shortened to "Usage tracking" / 用量追踪; the four
  sub-tables re-rendered in the content-analysis bar style
- Today tab: insights are now a proper card (bar rows with short
  labels, sentences+hints in tooltips) instead of loose text lines;
  estimated thinking share merged into the same card

Phase 5 of docs/V2.1-WORKFLOW-SPEC.md (final phase):

- buildAdviceSummary extracted to src/adviceSummary.ts (testable
  outside VS Code) and extended with labelled sections: multi-agent
  runs (per-run cost, agent fan-out, cache hit rate, models),
  estimated thinking share (30d), and the week-scope usage
  attribution (characteristics + top skills/subagents/plugins/models)
- new setting claudeCodeUsage.advice.userContext: free-text background
  about the user/project; when set, the system prompt instructs the
  model to end with a "Personalised for this project" section.
  Language enforcement unchanged
- README (en + zh-CN): "What's new in 2.1" feature list

Smoke-verified on real data: summary carries all three signal
sections (10 runs with cache hit 65-92%, thinking 17%, attribution
94%/59% characteristics); prompt-purity invariant from Phase 0 holds
(no scaffolding in the sample block).

Adds a relay-execution plan (for Opus 4.8 + workflow) to the V2.1 spec
after the implemented Phases 0-5, plus the on-disk log probe findings so
the next session doesn't re-derive them:

- effort/ultracode is NOT recorded in logs; the wf_ dir is the only real
  "dynamic-workflow engaged" signal
- attributionSkill / attributionPlugin are first-class fields on usage
  lines (authoritative, replaces the <command-name> heuristic)
- why native-Claude runs "look missing" from Workflows: Opus/Fable
  orchestration stays in the main session; only cheap haiku sub-agents
  (or none) write agent files

Continuation phases: 6 integrate onto current main (incl. PR ClaudeCodeUsage#31 context
indicator) - 7 authoritative attribution + workflow-mode honesty +
main-session orchestration cost - 8 real-total calibration (daily/weekly)
- 9 advice UX + Prompt Optimizer (the headline feature) - 10 README
showcase - 11 node:test coverage.

Also gitignore docs/claude-advice-*.md (local reports, never committed).

The V2.1/V2.2 work-plan specs and the ARCHITECTURE dev reference are the
maintainer's private iteration docs, not shipped product. Gitignore them and
untrack them so they live on disk but never enter commits/pushes going
forward (same treatment as docs/claude-advice-*.md). Working copies are kept;
re-track later by removing the .gitignore lines if a doc should become public.

Note: docs/V2.1-WORKFLOW-SPEC.md and docs/V2.2-*-SPEC.md currently exist on
origin/main; this stages their removal there on the next merge of this branch.

Catches the v2.1 workflow-attribution branch up to current main (~25 commits:
v2.0.3 work, PR-attribution merges, node:test framework, CI, macOS-keychain
credentials, PR ClaudeCodeUsage#31 status-bar context-window indicator).

Conflict resolutions:
- src/claudeApiClient.ts: keep feature's httpClient-based transport AND main's
  macOS-keychain credential read; import only execFileSync (curl/fetch now
  live in httpClient.ts, so spawn is no longer needed here).
- src/dataLoader.ts: take feature's v2.1 version wholesale, then port main's
  two new context methods (getCurrentContextInfo, contextWindowFor) + the
  ContextWindowInfo import on top.
- CHANGELOG.md: fold main's [Unreleased] items (context indicator, showCost)
  into the [2.1.0] block.

Verified: clean compile; node:test green (2/2); merged loader still yields
13 workflow runs, the attribution panel, thinking share, AND the new
context-window info. version stays 2.1.0.

…stration cost

Phase 7 of the V2.1 continuation plan:

7a — authoritative skill/plugin attribution. Records now carry _skill/_plugin
from the log's attributionSkill/attributionPlugin fields (CC >= 2.1).
getUsageAttribution weights skills/plugins by the EXACT message.usage of those
stamped lines, falling back to the <command-name> heuristic only when no record
carries the fields. On real data: 426 _skill / 203 _plugin records;
superpowers:executing-plans now reports a precise 80x / real cost weight.

7b — clearer run badges: "workflow" (a wf_ run dir) vs "subagents (ad-hoc)"
(a plain Task-tool fan-out), plus a hint that the effort level (ultracode/xhigh)
is NOT recorded in the logs, so neither badge claims one.

7c — main-session orchestration cost. getWorkflowBreakdown now finds the
session's main-thread spend bracketing each run (same session, inside the run's
window, attributed to exactly one run; window capped at 3 h so multi-day ad-hoc
batches don't over-claim). Surfaced as a pinned drill-down row and folded into
the headline cost/models, so a native-Claude run's real Opus/Fable cost is
finally visible. Verified: the PPRFS run's $11.83 of claude-opus/sonnet
orchestration (previously invisible behind $1.25 of cheap agent files) is now
attributed when its window is focused.

types: WorkflowUsage.orchestration; ClaudeUsageRecord._skill/_plugin.
i18n: workflowModeBadge / workflowModeHint / orchestration in all six languages.
Compile clean; node:test green; webview renders badges + 6 orchestration rows.

Phase 8 of the V2.1 continuation plan — the "can it be more precise than
text-length estimates?" answer.

Totals were always exact (message.usage); only the within-request category
split is estimated. The Content tab now keeps the text-length estimate for
relative shares but scales each side to the real billed totals over the
analysis window:
- assistant categories (assistantText / assistantThinking / toolCalls) sum to
  the exact Σ output_tokens;
- input-side categories (userPrompts / toolResults) sum to the exact
  Σ (input_tokens + cache_creation_input_tokens).

dataLoader computes the anchors (ContentAnalysis.calibration) over the same
30-day window. The Sessions Thinking column tooltip gains a calibrated
"real thinking tokens" figure (share x the session's exact output tokens).
New setting analysis.calibrate (default on); calibrated figures are labelled
distinctly from raw estimates.

Verified on real data: calibrated assistant sum = 5.57M = exact output;
calibrated input sum = 61.85M = exact input-side (the raw text estimate was
only 5.31M total — cache-creation is invisible to character counts, which is
exactly the undercount this fixes). Toggling the setting off restores the raw
estimates; within-side shares unchanged. Compile clean; node:test green.

…d, workflow note

Review feedback on the current build:

- Context window is now per-model, not a flat 200K: Opus 4.6+/Sonnet 4.6+/
  Fable 5 = 1M, Haiku and older Claude = 200K, DeepSeek = 128K, [1m] suffix
  = 1M (verified against the model catalog). Fixes Opus showing >100% fill.
- Context tooltip rewritten as a /context-style breakdown — fresh input /
  cache read / cache write / free space table — instead of a single number;
  ContextWindowInfo carries the input-side split. Bottom note shortened and
  the /clear sentence wrapped to its own line (all six languages).
- Today "Usage tracking" card now shows only exact cost-weighted shares; the
  text-length thinking-share estimate was removed from it (it stays on the
  Sessions tab, clearly marked estimate).
- Workflows tab: honest note that native-Claude ultracode often keeps
  orchestration in the main session (no agent files) and so appears in
  Sessions / Usage tracking rather than as a workflow row; a community issue
  + V2.1.1 follow-up are queued to improve detection.

i18n: freeSpace + workflowNativeHint in all six languages; shortened
contextHint/contextCostHint. Compile clean.

The AI advice (and the upcoming Usage Optimizer) gain three backends behind
one entry point, with Anthropic as the default shape:

- 'subscription' (default, no API key): reuse the Claude Code OAuth session
  the quota indicator already reads — Bearer <token> + the
  `anthropic-beta: oauth-2025-04-20` header — to call /v1/messages with a
  cheap model (Haiku). Feasibility verified on disk: a live haiku call
  returned 200, and the end-to-end advisor path returns real advice with no
  key configured. Uses a little of the user's quota.
- 'api' + apiFormat 'anthropic' (default for a configured key): x-api-key
  against /v1/messages.
- 'api' + apiFormat 'openai': the prior DeepSeek chat-completions path, kept
  for compatibility (existing configs keep working).

advisor.ts: extracted a shared callModel(system, user, opts) the optimizer
will reuse; send() now also falls back to curl on the 403
"Request not allowed" TLS-fingerprint gate (not just on thrown errors) — the
same gate the quota client handles. claudeApiClient.getAccessToken() exposes
a refreshed OAuth token. extension.ts: the advice command no longer demands
an API key when backend=subscription. New settings advice.backend /
apiFormat / subscriptionModel. Compile clean; node:test green.

… 9a/9c/9d)

The Content tab now leads with two action cards before the analysis:

- AI advice card — the "Get AI advice" button moves out of the analysis
  header into a labelled card with a one-line note on what gets sent.
- Usage Optimizer (opt-in: advice.optimizer.enabled, default off) — paste a
  rough request, get back ONE tightened paste-ready prompt plus a recommended
  effort / thinking / model. Three optional lenses: flag ambiguous refs,
  condense long pasted material, suggest a style direction. Runs through the
  same backend as AI advice (callModel); ONLY the pasted text is sent — no
  files, never Claude Code's terminal — behind a one-time consent modal. The
  round-trip lives in extension.runOptimizer, parsing the model reply on
  ===PROMPT=== / ===SETTINGS=== markers; the webview exposes it via the
  onOptimize hook and fills a baked, hidden result skeleton (no labels in JS).
  Verified end-to-end on the subscription backend (live haiku, markers parsed).

- advice.promptWindowDays (default 30) — how many days of prompts/content the
  analysis samples. Threaded through loadUsageRecords (analysis + calibration
  cutoffs) and echoed in the advice summary prose so the model knows the
  horizon.

i18n: 16 new keys filled across all six locales (en, de, zh-TW, zh-CN, ja,
ko). Compile clean; pricing node:test green.

Co-Authored-By: Claude <noreply@anthropic.com>

README: refresh the AI-advice showcase for 2.1 — advice now works with no
API key (subscription backend) and a second card, the opt-in Usage Optimizer
(only-the-pasted-text, one-time consent). Rewrote the advice/optimizer
screenshot blurb, the "What's new in 2.1" bullets, the Configuration table
(advice.backend / apiFormat / subscriptionModel / userContext /
promptWindowDays / optimizer.enabled), the Privacy note (subscription vs api
endpoints) and the demo-mode troubleshooting entry.

Tests: extract the optimizer's two pure helpers into advisor.ts so they can be
covered without the vscode API —
- buildOptimizerSystemPrompt(language, lenses)
- parseOptimizerOutput(raw) -> {prompt, settings}
extension.runOptimizer now calls them instead of inlining the strings (and the
duplicated parser method is gone). New node:test files:
- optimizer.test.ts: marker split, missing/out-of-order/empty fallbacks,
  language + lens inclusion.
- adviceSummary.test.ts: windowDays prose threading (default 30, custom,
  rounding/flooring).
11/11 green; compile clean.

Co-Authored-By: Claude <noreply@anthropic.com>

…ettings

VS Code's Settings UI listed ~27 options for this extension. Move the bulk of
them into a new ⚙ Settings tab in the dashboard and keep only a syncable core
in package.json.

- src/settings.ts (new): a single catalog of every setting (type, default,
  storage, group, control hints) plus SettingsStore — one read/write layer over
  two backings. Core settings (language, dataDirectory, advice.apiKey) stay in
  VS Code config so they remain in settings.json and travel with Settings Sync;
  the other 22 move to the extension's globalState. coerce() clamps numbers,
  validates enums and casts booleans so the panel can't store garbage.
- Migration: migrateOnce() copies any explicit settings.json value for a moved
  key into globalState on first launch (via config.inspect, which still reports
  user values for now-undeclared keys), guarded by a one-shot flag — upgrades
  keep their configuration.
- extension.ts: getConfiguration() now reads through SettingsStore; the store is
  shared with the webview and an onSettingsChanged callback re-applies config
  after dashboard edits (globalState changes don't fire onDidChangeConfiguration).
- webview.ts: new ⚙ Settings tab renders the catalog grouped (General / Status
  bar / Data & refresh / AI advice & Optimizer) with the right control per type;
  updateSetting / resetSetting / resetAllSettings handlers. The four in-webview
  config reads (workflow quota %, pause refresh, optimizer.enabled, calibrate)
  and the pause write now go through the store. Header gear + the optimizer
  "enable" button point at the new tab (the latter's setting left VS Code).
- package.json: trimmed contributes.configuration to the three core keys.
- i18n: panel chrome (tab title, intro, reset, four group headers) in all six
  locales. Setting labels/help are intentionally English (technical identifiers).

Verified: tsc clean; 11/11 node:test green; a mocked-vscode smoke run confirms
defaults, number clamping, enum guard, boolean coercion, snapshot integrity
(25 settings, unique keys), the core trio, and reset.

Co-Authored-By: Claude <noreply@anthropic.com>

Subscription backend dropped (this version): a re-test confirms Anthropic 403s
"Request not allowed" for calling /v1/messages with the Claude Code OAuth token
(it only succeeds by routing around the TLS-fingerprint gate via curl) — too
fragile/inappropriate for a public extension. Advice + Optimizer are API-only;
backend hardcoded to 'api'; advice.backend / advice.subscriptionModel settings
removed. The subscription transport stays in advisor.ts, dormant, to re-enable
if direct calls become permitted.

Context-window indicator:
- Honesty: contextWindowFor() now returns { tokens, estimated }; unrecognised /
  proxied models flag estimated=true and the % shows a "~" prefix instead of a
  confident figure. New setting contextWindowOverride (0=auto) lets users on
  custom/proxied models pin their real window — fixes the ClaudeCodeUsage#31-class wrong %.
- Correctness: getCurrentContextInfo() now prefers main-thread records (excludes
  _agentId/_workflowId) so a running sub-agent/workflow no longer hijacks the
  "current context" with its own window.
- Startup visibility: staleness guard widened 5h→24h so opening a window in the
  morning still shows the last session's context instead of blanking it.
- Default OFF + flagged experimental in Settings (the tooltip can only show the
  input-side total, not /context's category breakdown, which are CC internals
  not on disk — so it's approximate by nature; opt in if useful).
- Tooltip decluttered: dropped the verbose estimate/compaction notes (the
  explanation moved to the setting help), added a quota-style progress bar, and
  renamed the misleading "Free space"/"剩余空间" to "Context left"/"上下文余量".

Header: removed the apple-style auto-refresh toggle (auto-refresh is managed in
the ⚙ Settings tab now); the manual ↻ refresh button still appears top-right
when auto-refresh is paused. Two shortcut buttons remain — ✨ AI advice (jumps to
the Content tab) and ⚙ Settings (jumps to the Settings tab). The gear icon now
lives on the header button; the tab label drops it (cleaner, conventional).

Compile clean; 11/11 node:test green.

Co-Authored-By: Claude <noreply@anthropic.com>

Optimizer:
- The rewritten prompt is now PLAIN TEXT (the system prompt forbids Markdown —
  no bold/headings/backticks/bullets) so it pastes cleanly into a terminal.
- Clearer, task-framed copy: optimizerDesc says what it is; a new optimizerHowto
  line says how to use it (type/paste, pick lenses, Optimise). The consent text
  no longer references the removed advice.backend setting.
- Marked experimental (a pill on the card title); still default-off.

Cards: the AI advice + Usage Optimizer blocks are redesigned as one cohesive
"action card" treatment (frontend-design pass) — an accent rail + icon badge +
title/purpose/action — so they read as the two tools, distinct from the data
panels, while staying quiet and fully theme-driven (VS Code vars only). Replaces
the old left-tinted .advice-card.

i18n: optimizerDesc reworded, new optimizerHowto + experimentalBadge across all
six locales; consent text updated.

Compile clean; 11/11 node:test green.

Co-Authored-By: Claude <noreply@anthropic.com>

…r switch

Symptom: switching to another folder in the SAME VS Code window left the 5h/weekly
quota indicator blank — and it stayed blank even while Claude Code was actively
running. Opening the folder in a NEW window was fine.

Cause: the quota fetch falls back to the system curl binary (Anthropic's edge
rejects Node's TLS fingerprint). spawn() inherited the extension host's cwd, and
after a same-window folder switch that inherited cwd can point at a directory
that no longer exists — so spawn fails with ENOENT, fetchUsageLimits returns null,
and the indicator hides. A fresh window starts curl with a valid cwd, so it works
there.

Fix:
- httpClient.requestViaCurl spawns curl with cwd pinned to os.homedir() (curl
  needs no particular cwd) — immune to a stale/invalid inherited cwd.
- extension listens for onDidChangeWorkspaceFolders and forces a fresh quota
  fetch + refresh (resets the fetch throttle and the cold-start retry guard), so
  the indicator reappears even when the host persists across the change.

Compile clean.

Co-Authored-By: Claude <noreply@anthropic.com>

- Upstream moved to an organization: all upstream PR/issue URLs now point to
  ClaudeCodeUsage/ClaudeCodeUsage (numbers preserved on transfer) across the
  README, the translated READMEs, CONTRIBUTING and the CHANGELOG. The original
  author (@jack21) stays credited by name.
- Credits refreshed beyond the 2.0 baseline: added @ScherbakovAl (PR ClaudeCodeUsage#31, the
  original context indicator + showCost) and @nickearnshaw's PR ClaudeCodeUsage#24 (quota
  rollover), plus the closed issues (ClaudeCodeUsage#25 test seed, ClaudeCodeUsage#31 context detection).
- Reconciled the docs with the final 2.1 behaviour: AI advice/Optimizer are
  API-only (the keyless subscription backend is documented as prototyped-but-
  not-shipped), the context indicator and Optimizer are experimental/off, the
  Optimizer returns plain text, settings live in the dashboard, and the header
  was trimmed. Added CHANGELOG Changed/Fixed bullets for the header rework,
  plain-text optimizer, redesigned cards and the workspace-switch quota fix.

Docs only. Compile clean; 11/11 node:test green.

Co-Authored-By: Claude <noreply@anthropic.com>

Fixes:
- **F5 launch failed every time** ("preLaunchTask npm: compile terminated with
  exit code 1"). On Windows the npm task type runs through powershell.exe →
  npm.ps1, which the default execution policy blocks. .vscode/tasks.json now runs
  tsc directly via node (`type: process`, no shell); launch.json points at it
  (and its missing-comma JSON error is fixed).
- **Optimizer result wiped by auto-refresh** — the result lived only in the DOM,
  so any dashboard refresh erased a prompt you were still reading. The webview
  now persists the last optimizer interaction (draft + lens choices + result) and
  restores it on every re-render.
- **No way back to the dashboard when cost+quota+context are all hidden** — the
  first status-bar item now stays as an icon-only entry point ($(graph), opens
  the dashboard) when all three are off, instead of leaving zero clickable items.

Features / polish:
- **Status-bar metric** setting (`statusBarMetric`, default cost) — switch the
  first item from today's cost to today's total token count, shown compact (k/M)
  with the configured decimal places. New `I18n.formatTokensCompact`.
- **Optimizer copy de-jargoned** — dropped the "lens/透镜" wording for plain
  "optional tweaks"; each of the three options now has a tooltip explaining it
  (new *Hint keys, all six locales).
- **Settings recommendation rendered for reading, not copying** — it's no longer
  a monospace code block; clean UI-font block with a soft accent rail. The
  optimised prompt stays monospace/copyable.
- **Model recommendation grounded in reality** — the optimizer is told to pick
  only from the models you actually use (Claude reduced to family names
  haiku/sonnet/opus/fable, third-party kept as-is), so it stops naming models you
  don't have or stale versions a third-party model wouldn't know.
- **Context tooltip** widened (40-segment bar) so its right edge lines up better
  with the figures below.

Compile clean; 11/11 node:test green.

Co-Authored-By: Claude <noreply@anthropic.com>

The dashboard felt heavy in the Extension Development Host. Two safe mitigations:

- updateWebview() now skips assigning panel.webview.html when the rendered HTML
  is byte-identical to what's already shown. Assigning .html forces a full
  webview reload (re-parse + re-run the inline script + reset scroll) — the
  heaviest recurring cost on refresh ticks; skipping no-op renders avoids it.
  lastHtml is cleared on panel dispose so a reopened panel always paints.
- Removed the two highest-frequency webview console.log("[DEBUG] …") lines (one
  per message received, one per document click) that piled up in the webview
  console during long sessions.

The live full-reload during active sessions is still inherent to the current
"regenerate the whole HTML" model — `pauseDashboardRefresh` (now in the ⚙
Settings tab) remains the lever for it; an incremental-update refactor is a V2.2
candidate.

Compile clean.

Co-Authored-By: Claude <noreply@anthropic.com>

…icon

- README-zh-CN.md fully rewritten to match the current English README and the
  2.1 final state (settings-in-dashboard, API-only advice + dormant subscription
  note, experimental Optimizer with plain-text output, experimental context
  indicator with window override, status-bar cost/tokens metric, the all-off
  entry icon, the workspace-switch quota fix). Other locale READMEs stay slim by
  design — only their upstream-org URLs were updated earlier.
- README.md: added the status-bar metric (cost ↔ tokens) and the all-off entry
  icon to "What's new in 2.1".

Docs only.

Co-Authored-By: Claude <noreply@anthropic.com>

…eUsage#38)

Integrate the leftover upstream PR ClaudeCodeUsage#38 (@wheelbarrel00) that never shipped with
the unreleased v2.0.3. It adds `showOpusWeekly` (opt-in, default off): when on,
the quota item appends `opus:NN%` after the 5h / weekly figures, so heavy Opus
users get an at-a-glance weekly Opus signal.

This branch diverged from upstream at PR ClaudeCodeUsage#31, so ClaudeCodeUsage#38 wasn't present; cherry-pick
would conflict with the status-bar changes made here (setVisibility signature,
cost/tokens metric, all-off entry icon). Re-applied faithfully instead, adapted
to the dashboard-managed settings:
- settings.ts: showOpusWeekly in the Status-bar group (state-stored, default off).
- statusBar.ts: 5th setVisibility param; updateQuota appends opus:NN% and keeps
  the item visible when only the Opus window is present.
- extension.ts / types.ts: thread the setting through.
- README / README-zh-CN / CHANGELOG: documented and credited @wheelbarrel00 (ClaudeCodeUsage#38).

Compile clean; 11/11 node:test green.

Co-Authored-By: Claude <noreply@anthropic.com>
Co-Authored-By: wheelbarrel00 <103789673+wheelbarrel00@users.noreply.github.com>

…t tooltip

Usage Optimizer:
- The "copy prompt" control is now a compact floating icon button (⧉ + label)
  in the prompt block's top-right, themed to match the extension, with a green
  "copied" state — instead of a generic button below the text.
- "Recommended run settings" now renders each line as "Label  [value chip]  reason",
  so the choice that matters (high / on / opus) stands out as a filled chip
  instead of plain text. The optimizer is told to emit exactly three parseable
  lines (Effort/Thinking/Model) with English value tokens; a webview formatter
  (formatOptSettings) turns them into chips and re-applies after a re-render.

Context tooltip:
- Rebuilt as a single HTML table with the progress bar in the right-hand value
  column, so the bar's right edge lines up with the cache/figure numbers below
  it (the requested alignment).
- The footer hint is now two short lines: "New task → /clear" and
  "Same task → /compact" (contextHintCompact added, all six locales).

Compile clean; 11/11 node:test green; the settings parser verified on sample
model output (values isolated correctly for the chips).

Co-Authored-By: Claude <noreply@anthropic.com>

Ran the zh-CN README through DeepSeek V4 Pro (the user's own key) with strict
"only smooth the wording — keep every heading, table, link, code span, command
and technical term verbatim" instructions, then validated the result before
applying: same 10 ## / 8 ### headings, 5 images, 26 links, tables and section
order unchanged; tables / privacy / troubleshooting / credits untouched. It
removes the most translated-sounding bits (the opening sentence, clichés like
"尽收一处" / "着力点", over-used em-dashes) so it reads less machine-generated.

Docs only (zh-CN README).

Co-Authored-By: Claude <noreply@anthropic.com>

…iers

- **Settings panel is fully localized.** The ⚙ Settings tab showed translated
  group headers/chrome but English per-setting labels and help. Added a
  SETTINGS_I18N map (de-DE / zh-TW / zh-CN / ja / ko) covering every setting's
  label + help, resolved via I18n.settingText() with the English catalog as the
  en fallback. Code/command tokens (token, /v1/messages, opus:NN%, k/M,
  git/folder/flat …) kept verbatim. (Translations drafted with DeepSeek V4 Pro
  and validated for completeness across all 26 settings × 5 locales.)
- **Context tooltip:** swapped the progress bar to the left and the percentage
  to the right, matching the tooltip's left-label / right-value convention.
- **Usage Optimizer effort tiers:** the run-settings recommendation may now
  suggest `max` and `ultracode` (in addition to low/medium/high), with a note
  that ultracode means splitting across sub-agents — same chip UI.

(A local-only roadmap note in docs/V2.2-FEATURES-SPEC.md — gitignored — records
that the Optimizer / advice suggestions do NOT read local files; that stays a
future opt-in.)

Compile clean; 11/11 node:test green; settingText verified resolving per locale.

Co-Authored-By: Claude <noreply@anthropic.com>

…mented on the new dashboard-settings architecture, original PR ClaudeCodeUsage#38 by @wheelbarrel00 retained as a merged ancestor)