# Changelog

## 0.2.0-beta.4 - 2026-07-28

- Made the unqualified `pi install npm:tokenomy-pi` command the documented
  install and update path, matching pi.dev and the supported npm default
  channel.
- Added a standalone standard usage scenario covering setup, route
  explanations, quality feedback, reports, budgets, memory, compaction,
  project-local data controls, and updates.
- Corrected stale scope and language guidance: generated data remains
  project-local, supported routing dictionaries cover seven languages, and
  runtime control commands do not persist configuration changes.
- Included the new usage guide in the npm package and added regression coverage
  for public installation guidance.

## 0.2.0-beta.3 - 2026-07-27

- Kept npm `beta` as the prerelease channel while also moving `latest` to the
  current prerelease until Tokenomy publishes its first stable version. This
  keeps pi.dev, unqualified Pi installs, and normal package updates aligned with
  the supported beta without allowing future betas to replace a stable default.
- Updated the pinned GitHub installation example to this release.

## 0.2.0-beta.2 - 2026-07-27

- Made statistics, routing history, and telemetry read-modify-write updates
  transactional across concurrent Pi processes; added a four-process race test.
- Replaced the partial schema with a generated, exhaustive, strict schema and
  runtime leaf validation from the same source of truth.
- Expanded the economic benchmark to a 30-case corpus, repeatable
  counterbalanced arms, classifier-inclusive measurement, real-repository
  manifests, Wilson intervals, trend evidence, and a quality non-inferiority
  gate.
- Added authenticated Plus catalog/rate freshness checks, a weekly self-hosted
  workflow, and the observed `gpt-5.3-codex-spark` catalog entry without routing
  to it before a verified rate exists.
- Added Windows packed-install CI, a Pi 0.82.1/latest-0.82.x matrix, and legacy
  stats/v2 rollup migration coverage.
- Split defaults, schema, budget policy, and evaluation logic into focused
  modules.
- Added reserved credits, per-tier session budgets, one/two-tier downshifts,
  per-tier dashboard spend, and estimated remaining turns.
- Added `/tokenomy data` inventory plus selective cache, telemetry, memory,
  debug, and confirmed all-data purge commands.
- Restricted supported subscription wording to the live-tested ChatGPT Plus
  setup; Pro, non-Codex providers, and native Codex integration are explicitly
  unsupported or experimental rather than implied compatible.

## 0.2.0-beta.1 - 2026-07-27

- Updated Plus-tested routing defaults for the current Codex catalog:
  `gpt-5.4-mini` for classifier/cheapest work, `gpt-5.6-terra` for medium
  work, and `gpt-5.6-sol` for complex work, with Luna and prior-model
  fallbacks.
- Added an opt-in paired economic benchmark that runs identical fresh-workspace
  tasks through Tokenomy and a fixed model, verifies both outcomes, and records
  measured token and estimated-credit deltas.
- Documented the counterbalanced Plus run: all three task pairs verified and
  39.6% fewer estimated credits in aggregate versus fixed GPT-5.6 Sol; the
  small-sample limitation is explicit.
- Added strict pre-merge config type checking, unknown-key warnings, a bundled
  JSON Schema, and `/tokenomy doctor`.
- Added actionable `warn`, `save`, and `ask` budget policies while preserving
  stronger routing for high-risk work.
- Made local JSON persistence crash-safe with private directories/files,
  per-file locking, temporary files, atomic rename, and stale-lock recovery.
- Added private, default-redacted debug traces, retention cleanup, and
  `/tokenomy debug purge`; raw capture now requires `debug.redact: false`.
- Split config, current model policy, and local persistence out of the main
  extension module.
- Added installed-tarball smoke tests and Linux/macOS CI coverage.
- Fixed release channels: prereleases publish only under npm `beta` and become
  GitHub prereleases; stable releases alone move `latest`. npm propagation is
  retried before creating a tag and release.
- Clarified throughout the package that Tokenomy is a Plus-tested Pi extension,
  not a native Codex CLI extension, and documented explicit beta exit criteria.
- Expanded the mocked integration suite from 73 to 83 tests.

## 0.1.25-beta - 2026-07-27

- Added verified `success|partial|failure` feedback, multilingual correction
  detection, and an opt-in independent quality evaluator.
- Added deterministic economy-mode experiments, no-cost shadow tiers, and
  per-mode prompt/completion/success/credit rollups.
- Added local routing support for Ukrainian, Russian, Spanish, French, German,
  and Portuguese instructions.
- Added provider allowlists, provider-qualified multi-provider routing,
  optional live model discovery, and validated external rate-card loading.
- Added a validated account-quota snapshot adapter and `/tokenomy quota`;
  unavailable personal ChatGPT Plus totals are never guessed.
- Added tool-result token/character measurement, duplicate detection, and
  opt-in head/tail truncation for oversized results.
- Added measured compaction before/after/saved token counters.
- Added `/tokenomy dashboard`, 7/30-day trends, session/daily credit alerts,
  quality evidence, and mode comparisons.
- Added an opt-in signed-in end-to-end evaluation runner and manual
  self-hosted workflow with uploaded evidence.
- Upgraded telemetry rollups to schema version 3 while retaining v2 data.
- Automated npm verification, matching `v<version>` tag creation, and
  idempotent latest GitHub Release creation in the publish workflow.
- Clarified that ChatGPT Plus is Tokenomy's live-tested subscription target;
  ChatGPT Pro is expected to be compatible but remains unvalidated.
- Expanded the integration suite from 60 to 73 tests.

## 0.1.24-beta - 2026-07-27

- Added `save`, `balanced`, and `quality` economy modes, configurable
  plan-credit rate cards, live-classifier session budgets, and conservative
  classifier break-even gating.
- Added completion/tool-error/retry telemetry and clearly labeled
  cost-per-completed-turn reporting without claiming verified task success.
- Added sanitized provider rate-limit header snapshots and `/tokenomy limits`
  with explicit project/process scope.
- Added `/tokenomy compact`, opt-in threshold compaction with cooldown and
  task-preservation instructions, and compaction rollups.
- Replaced model-rank “token-equivalent” claims for new turns with
  provider-reported input, cached-input, cache-write, output, reasoning,
  total-token, request, and cost accounting from Pi `agent_end` messages.
- Added telemetry rollup schema version 2, per-turn measured/unavailable status,
  cache-read ratio reporting, classifier overhead, and versioned ChatGPT
  plan-credit estimates.
- Migrated lifecycle handling to current Pi `agent_end` and `agent_settled`
  events. Model restoration now waits until retries and continuations settle.
- Preserved the user's startup model and added guarded restoration of the
  pre-route thinking level.
- Made project-memory and routing-digest prompt injection opt-in by default and
  removed changing savings/context counters from prompt-discipline text to
  improve exact-prefix cache stability.
- Kept older telemetry readable while labeling historical savings/cost-unit
  values as pre-v2 model-rank proxies.
- Expanded integration coverage for current lifecycle registration, measured
  usage, plan-credit conversion, unavailable usage, startup model preservation,
  state restoration, and stable default prompt additions.
- Added strict TypeScript checking against pinned Pi `0.82.1` development
  packages and constrained supported peers to the `0.82.x` API line.

## 0.1.23-beta - 2026-07-09
- Preserved the previous routing context for short continuation prompts such
  as `continue`, so follow-up turns do not accidentally downshift mid-task.
- Added a session/config snapshot when debug tracing is enabled by command and
  made debug telemetry snapshots show the same `updatedAt` value persisted to
  disk.

## 0.1.22-beta - 2026-07-09
- Added opt-in local debug session traces with `/tokenomy debug on`,
  `/tokenomy debug path`, `/tokenomy debug off`, and `debug.trace` config.
  Traces are JSONL files intended for diagnosing routing decisions and include
  a clear warning because they can contain raw session data.

## 0.1.21-beta - 2026-07-09
- Added default-on model restoration after each prompt so temporary routing
  choices do not leave Pi on the cheaper or stronger selected model.
- Added local prompt-shape analysis for question/action/mixed prompts,
  action-count signals, multi-step routing, and prompt-shape telemetry.
- Adopted `compromise` for local sentence, question, and verb analysis in the
  prompt-shape router.
- Replaced Tokenomy's local classifier-token estimate with
  `tokenshrink.countTokens()`.

## 0.1.20-beta - 2026-07-09
- Added durable local telemetry rollups with daily, monthly, and lifetime
  prompt-safe aggregates for estimated savings, route distribution, memory,
  compression, cache, fallback, and guard activity.
- Added `/tokenomy report`, period-specific report commands, and
  `/tokenomy export-report` for local evidence of Tokenomy efficiency over
  time.

## 0.1.19-beta - 2026-07-08
- Kept trivial general prompts and single-command local info questions such as
  `how time is it?` on the cheapest model even in large project contexts.

## 0.1.18-beta - 2026-07-08
- Removed Tokenomy's main Pi footer/status entry to avoid crowding or
  conflicting with other extension footers; `/tokenomy status`, notifications,
  history, and telemetry remain the stable visibility paths.

## 0.1.17-beta - 2026-07-08
- Made the Tokenomy status/footer text visibly labeled and refreshed by
  `/tokenomy status`.
- Routed broad review/refactor prompts such as `please do an audit`,
  `please review`, and `please refactor` to the complex tier instead of the
  cheapest mini model.
- Added transparent bypass for prompts written primarily outside English while
  still allowing English instructions that contain non-English payload text.

## 0.1.16-beta - 2026-07-08
- Added a compact Tokenomy status/footer formatter that shows tier, source,
  confidence, session savings, and lifetime savings without replacing other
  plugin status entries.
- Added local workflow routing for state-changing git/tool chains such as
  `commit & push`, while keeping read-only git inspection prompts cheap.

## 0.1.15-beta - 2026-07-08
- Fixed short config-audit prompts such as nvim/tmux final scans routing to
  the cheapest mini model.

## 0.1.14-beta - 2026-07-08
- Audited docs for npm install, compatibility, memory, compression, telemetry,
  and release guidance alignment.

## 0.1.13-beta - 2026-07-08
- Improved the npm/GitHub package description and README introduction to state
  Tokenomy's user-facing goal, default benefits, Codex subscription scope,
  local memory, compression, safety guards, and telemetry.

## 0.1.12-beta - 2026-07-08
- Added default-on local project memory with automatic safe fact discovery and advisory injection.
- Added `/tokenomy memory`, `/tokenomy memory show`, `/tokenomy memory refresh`, `/tokenomy memory clear`, `/tokenomy memory on`, and `/tokenomy memory off`.
- Added memory injection telemetry and tests for value, privacy, disabling, stale facts, and simple-prompt skips.

## 0.1.11-beta - 2026-07-08
- Added prompt-safe local routing telemetry with `/tokenomy history`, `/tokenomy export-history`, and `/tokenomy reset-history`.
- Added compression guard telemetry and lifetime stats for rejected classifier prompt compression.

## 0.1.10-beta - 2026-07-08
- Added local TokenShrink-based classifier prompt compression.
- Added `promptSimplification.compressionEnabled` to disable TokenShrink compression while keeping it enabled by default.
- Added local prompt simplification for large classifier prompts and long-output condensation guidance.
- Fixed simple shell/listing prompts such as `ls -l` routing to the medium tier in large contexts.

## 0.1.1-beta through 0.1.9-beta - 2026-07-07 to 2026-07-08
- Reduced duplicate stats writes during digest injection and added regression coverage for corrupted classifier cache handling.
- Added intent/risk-aware routing, classifier decision caching, adaptive fallback tiers, and compact project digest injection for large/repeated contexts.
- Changed the default medium model preference to `gpt-5.4` before `gpt-5.4-mini`.
- Added the installed Tokenomy package version to `/tokenomy status`.
- Synced npm's default `latest` dist-tag during publish so `pi install npm:tokenomy-pi` gets the current beta.
- Fixed stats persistence for globally installed Tokenomy in projects that do not already have a `.pi` directory.
- Public-readiness documentation pass.
- Added installation, configuration, security, contributing, and limitations docs.
- Expanded integration tests for classifier acceptance/rejection, missing model fallback, and invalid config warnings.
- Added `/tokenomy explain`, `/tokenomy reset-stats`, and `/tokenomy dry-run` commands.
- Added Pi package manifest support for `pi install`.
- Added GitHub Actions CI for JSON validation and integration tests.
- Added npm package metadata and automated npm publish workflow.
- Added manual npm publish workflow dispatch and graceful skip when `NPM_TOKEN` is not configured.
- Updated GitHub Actions to Node-24-compatible action versions and bumped beta package for npm publish flow testing.

## 0.1.0-beta - 2026-07-07
- Added config validation and warnings for invalid/empty model lists.
- Added optional dry-run/debug routing support.
- Made tool management opt-in by default for safer sharing.
- Added compatibility notes for public use.