Skip to content

v1.2.0: setup hard gate + guided wizard, render-time i18n, 11-locale READMEs, 5-min real-fail isolation - #16

Merged
mrzturn merged 9 commits into
mainfrom
dev
Sep 19, 2026
Merged

mrzturn merged 9 commits into
mainfrom
dev

Conversation

@mrzturn

@mrzturn mrzturn commented Sep 19, 2026

Copy link
Copy Markdown
Owner

Summary

  • Setup hard gate + guided wizard: unconfigured task pools no longer default to all models — task dispatch is denied until all 6 pools have >=1 model and the capability ranking has >=1 entry; new /switchman-setup TUI wizard + /switchman-setup-chat conversational fallback; [SETUP] directive + sidebar notice + [LIMITS] segment; mtime hot-reload opens the gate without restart
  • Render-time i18n (11 locales): sidebar, dialogs, notices, and CLI copy localize at render time (src/i18n.ts); protocol copy stays byte-stable English
  • Model rank scoped to the pool universe with pool-lane priority ordering (manual rank = absolute ordering key on pool lanes)
  • Economy lane effort preference raised to medium→high→low→minimal
  • Self-read budget defragmented: tolerance band, headroom grants, continuation merge
  • Real-fail isolation window 30 min → 5 min (rate-limit keeps 10-min self-heal; endpoint keeps 6h)
  • Docs: setup wizard documented in READMEs/quick-start/reference manuals (EN/中文); READMEs translated into 9 more locales (11 total with a unified language switcher)

Test plan

  • bun test: 551 pass / 0 fail
  • bun run typecheck: clean

… ordering

1. Narrow /modelRank to the pool universe (rankViewRows scoping, empty-universe
   guidance that gates set/add, TUI empty dialog with a jump to /poolConfig)
2. Make the manual rank the absolute ordering key on pool lanes: rankCandidates
   takes manualOrder/poolOrdered (beats tier grouping, partitions, family
   preference), and pool members are exempt from capability level floors with
   an explicit pool-config override note
3. Drop dead-key retention from the rank view and auto-prune non-pool entries
   (and their anchored scores) on every rank write via the shared persist path
4. Surface the rankable universe in the banner, update /modelRank chat template,
   rework tests around the new semantics (518 green) and regenerate shells matrix
1. Invert LANE_SPEC.economy efforts (low→minimal→medium→high becomes medium→high→low→minimal, first supported wins): a medium-less flash (e.g. glm-5.3-flash) now takes its high face as the economy chain head instead of being pinned to low; off stays the lane-tail fallback partition
2. Regenerate shells matrix (economy chain: ds-mx-v4fv-high→copilot-mx-luna-medium→copilot-mx-haiku45-off)
3. Update the activation chain assertion and the reference docs effort-order table
…s, and continuation merge

1. Allow reads within R* + deny-retry overhead (~400 tok) outright, denying them costs more than the overshoot saved
2. Size per-call grants from the remaining turn headroom (max 2x R*) so demonstrated-range reads land one-shot
3. Merge bounded continuations of the last granted window into one completion shot, even past the spent turn budget
4. Sync injected protocol section 5, config comments, and dispatch-gates tests to the new semantics
1. Add i18n core (src/i18n.ts) with 11-locale catalogs; locale chain: project lang -> global ui.lang -> terminal env -> en, with English/key fallback that never throws
2. Convert status-log notices to structured keyed entries rendered via t() at sidebar display time; sidebar chrome, quota rows, pool/rank/handover dialogs, and config CLI copy all localize
3. Localize the six-lane sidebar labels display-only (t() falls back to raw lane identifiers; runtime LANE_ORDER, pool-config keys, and protocol stay English), with CJK-aware column padding (padEndW)
4. Keep model-facing protocol copy byte-stable English (capture anchors, [ROUTES]/[WATERMARK]/[LANG] lines); extend tests for catalog completeness, placeholder parity, and locale resolution
…n-setup-chat)

- Unconfigured task pools no longer default to all models: task dispatch is
  denied in main sessions until all 6 pools have >=1 selected model AND the
  capability ranking has >=1 entry (src/setup-gate.ts; no waiver flag;
  shells/internal sessions exempt; rides on the lang gate)
- TUI wizard /switchman-setup: resumable intro jumping to the first
  unfinished step, one multi-select dialog per pool in lane order (per-pool
  confirm persists immediately, min 1), sequential strongest-first rank pick
  over the selected union, done screen with restartRequired providers
- Chat fallback /switchman-setup-chat: state overview -> per-pool
  multi-select questions -> ranking -> CLI persist -> verify
- [SETUP] directive per turn while incomplete; sidebar notice + [LIMITS]
  banner segment; mtime-cached reads so a wizard write opens the gate
  without restart; localized wizard copy across all 11 locales
- Document /switchman-setup and /switchman-setup-chat in both READMEs
  (quick start restructured: guided wizard is step 3 required-once,
  modelRank/poolConfig folded into step 4 fine-tune) and add a full
  required Step 3 with cheat-sheet rows to both quick-start guides
- Add /switchman-setup subsection to both reference manuals (manual
  overrides section)
- New README translations: 繁體中文, 日本語, 한국어, Espanol, Francais,
  Deutsch, Italiano, Portugues, Русскии — 11 languages total with a
  unified language switcher header across all READMEs
- REAL_FAIL_TTL_MS 1_800_000 -> 300_000: probe-ok-but-real-call-failed
  states are usually transient vendor blips; a 30-minute head-of-chain
  lockout starved lanes far longer than the failure warranted
- Rate-limit keeps its dedicated 10-min self-heal TTL; endpoint-class
  errors keep the 6h TTL; restart still clears all in-memory isolation
- Deny/banner copy updated (5 minutes / 5-min); TTL-separation test
  rewritten for the inverted default-vs-ratelimit hierarchy
…lake fix)

Root cause: 6 harness files boot SwitchmanPlugin without seeding
capability.json, so the startup refresh fires a REAL network fetch; on
fast CI networks the async write-back lands after SWITCHMAN_STATE has
already switched to a LATER test file's sandbox, poisoning its empty-pin
capability.json — user-overrides then reads a live index (source=api
instead of curated exact) and computeLane floors diverge. Local runs pass
purely on slower-network timing luck.

Fix: seed the empty-pin capability.json (same fixed-empty shape as the
other harnesses) in index-options, search-clarify, subagent-cap,
shell-context-advisory, ctx-pause, and expert-command — the capabilityStale
TTL gate then skips the refresh entirely: no network, no cross-file
write-back.
@mrzturn
mrzturn merged commit afe3364 into main Sep 19, 2026
3 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant