chore: rewritten changelog

AGENTS.md

@@ -282,6 +282,7 @@ Location: `packages/*/CHANGELOG.md` (per package).
 **Rules:**
 
 - New entries always go under `## [Unreleased]`.
+- Entries are one line, brief, and user-facing: lead with what the user will see or can now do. Root-cause narration and implementation detail belong in the commit/PR, not the changelog.
 - Never modify already-released sections (e.g., `## [0.12.2]`) — they are immutable.
 - Don't flag changelog section order or formatting in reviews or PRs — `bun run release` runs `fix-changelogs` which normalizes everything automatically.
 

packages/agent/CHANGELOG.md

@@ -4,18 +4,18 @@
 
 ### Breaking Changes
 
-- Local token counting is now an immutable, model-scoped `Tokenizer` instance instead of process-global functions. The free `countTokens`/`countTokensConservatively` exports and the `setTokenizerModel()` global setter are gone; construct a `Tokenizer(modelId)` — the encoding is fixed at construction, there is no `setModel` — and call `tokenizer.countTokens(text, mode?)`. An `Agent` owns one for its active model (exposed as `agent.tokenizer`) and replaces the instance when the active model's encoding changes, so don't cache it across model switches. `countTokensConservatively` collapsed into that one method as `mode: "upperbound"`; the modes are `"strict"` (always exact native), `"approximate"` (default, fast byte estimate when no exact tokenizer applies), and `"upperbound"` (raw byte length, never undercounts).
+- Replaced global token counting functions (`countTokens`, `countTokensConservatively`, `setTokenizerModel`, and `estimateTokens`) with model-scoped, immutable `Tokenizer` instances (`agent.tokenizer`). Use `tokenizer.countTokens(text, mode?)`, `tokenizer.countMessage(message)`, or `tokenizer.countMessages(messages)`.
-- The free `estimateTokens(message, tokenizer, options?)` is gone; use `tokenizer.countMessage(message, options?)` or `tokenizer.countMessages(messages, options?)`. Per-message estimates are memoized per tokenizer instance (keyed by message identity) and invalidated across every instance by `invalidateMessageCache(message)`, which bumps a shared version tag. `findCutPoint`, `prepareBranchEntries`, `collectShakeRegions`, `pruneToolOutputs`, `pruneSupersededToolResults`, and `trimRemoteCompactionInputToContextWindow` take an explicit `Tokenizer`, and `prepareCompaction` accepts the caller's (warm) tokenizer as an optional trailing parameter. Token math is scoped to the model that will be billed for it rather than to whichever `Agent` last constructed itself.
+- Updated context management functions (`findCutPoint`, `prepareBranchEntries`, `collectShakeRegions`, `pruneToolOutputs`, `pruneSupersededToolResults`, and `trimRemoteCompactionInputToContextWindow`) to require an explicit `Tokenizer` instance.
 
 ### Added
 
-- Added `Tokenizer.checkTokenBudget(text, budget)`: a cheap-first budget probe. Byte length is a hard upper bound on token count, so text whose raw bytes already fit answers "fits" without tokenizing at all; only text that busts the bound pays for an exact count (and that count is returned, so a proportional clamp gets the denominator it needs). Since the bound overshoots ~4x on prose, the common "comfortably under budget" answer is free. Compaction's summary-window fit check and OpenAI remote-compaction trimming now route through it.
+- Added `Tokenizer.checkTokenBudget(text, budget)` to efficiently verify if text fits within a token limit using fast byte-bound checks before falling back to full tokenization.
-- Added provider-anchored transcript accounting (`findTranscriptUsageAnchor`, `isTranscriptUsageAnchor`, `estimateTranscriptTokens`). Every settled assistant turn carries `usage` covering the exact prompt it was sent, so transcript sizing charges that report for the prefix and tokenizes only the tail appended after it — counting proportional to one turn instead of the whole history, every turn. The four hand-rolled copies of the anchor trust rules (session stats ×3, shake) now share one predicate, and the deliberately provider-independent compaction floor counts every message locally via `tokenizer.countMessages`.
+- Added provider-anchored transcript token estimation (`findTranscriptUsageAnchor`, `isTranscriptUsageAnchor`, `estimateTranscriptTokens`) to calculate transcript token counts incrementally from the latest reported assistant turn usage.
-- Exported `remotePreserveReusable(preserveData, activeModel, settings)` — whether a prior remote compaction's provider-native replay payload is still readable by the active model — so hosts can validate speculatively produced compaction results before committing them.
+- Added `remotePreserveReusable()` to check whether a previous remote compaction payload remains reusable with the active model.
 
 ### Changed
 
-- Catalog-resolved tokenizer families now drive exact native counts: Claude, Qwen 3.5+, DeepSeek V3/V4/R1, Kimi K2/K3, and GLM-5+ use their matching embedded tokenizer; unknown models retain the fast estimate (or o200k with `PI_TOKENIZER_ACCURATE=1`). `Tokenizer` now takes the resolved catalog `Model`, never a raw model id.
+- Expanded native tokenizer support across catalog models, adding exact embedded token counting for Claude, Qwen 3.5+, DeepSeek V3/V4/R1, Kimi K2/K3, and GLM-5+ models. `Tokenizer` now constructs from a resolved catalog `Model`.
 
 ## [17.3.8] - 2026-08-19
 

packages/ai/CHANGELOG.md

@@ -8,8 +8,8 @@
 
 ### Fixed
 
-- Fixed the tool-argument repair layer applying lossy repairs on union-branch diagnoses: when a value failed every `anyOf`/`oneOf` variant, the first failing branch's issues were treated as authoritative, so object payloads got JSON-stringified into string-typed fields and unrecognized keys were silently deleted — corrupting subagent `yield` payloads (validation "passed" and parents received `summary: "{\"purge\":13,…}"` instead of a retryable error). Issues from a failed union are now marked at every depth and only receive lossless repairs (JSON-string parsing, boolean spellings, scalar coercion) — unless a `const`/`enum` discriminator uniquely tag-selects the intended variant, whose diagnosis stays fully repairable (and is now the surfaced one, rather than the first failing branch's).
+- Fixed tool-argument repair applying lossy transformations (such as stringifying objects or stripping unrecognized keys) when validating union schemas (`anyOf`/`oneOf`), preventing corrupted tool call and subagent payloads
-- Fixed local OpenAI-compatible servers with strict `chat_template_kwargs` whitelists (e.g. NInfer) failing every Qwen 3.8+ turn with `400 chat_template_kwargs.reasoning_effort is not supported` after the effort routing fix: the reasoning-effort fallback now recognizes a rejection of the kwargs spelling itself, retries with the kwarg stripped while keeping the effort on the standard top-level `reasoning_effort` field (hoisting it there for the kwargs-only vLLM dialect), and remembers the shape for the rest of the session. Value-level rejections and drops now also update the `chat_template_kwargs.reasoning_effort` twin instead of leaving a stale effort for kwargs-reading renderers, and unknown-parameter 400s naming `reasoning_effort` are recognized as effort rejections.
+- Fixed 400 errors when communicating with local OpenAI-compatible inference servers that reject `chat_template_kwargs.reasoning_effort` by improving reasoning effort parameter fallback and compatibility handling
 
 ## [17.3.8] - 2026-08-19
 

packages/catalog/CHANGELOG.md

@@ -4,13 +4,12 @@
 
 ### Added
 
-- Models now materialize an optional `tokenizer` family in the catalog (`claude-v3`/`v47`/`v5`, Qwen 3.5+, DeepSeek V3/V4/R1, Kimi K2/K3, and GLM-5+). The field follows `requestModelId`, applies to bundled, discovered, and custom models, and can be explicitly overridden in model configuration.
+- Models now include an optional `tokenizer` family field across bundled, discovered, and custom models (supporting Claude, Qwen, DeepSeek, Kimi, and GLM families), with support for explicit overrides in model configuration.
-- Subscription Codex GPT-5.6 Sol/Terra/Luna now carry the same `cost.longContext` tier as their first-party API siblings (2x input / 1.5x output above 272K input tokens, [openai/codex#32486](https://github.com/openai/codex/issues/32486)), so cost attribution reflects the higher rating above the threshold and downstream consumers can locate the standard-pricing boundary.
+- Added long-context cost tiers (`cost.longContext`) to subscription Codex GPT-5.6 models (Sol, Terra, Luna) matching first-party API pricing above 272K input tokens.
 
 ### Fixed
 
-- Fixed `opencode-go/muse-spark-1.2` and `muse-spark-1.2-contributor` still failing every tool-call turn with `OpenAI completions stream closed before a finish_reason was received` on 17.3.8. The earlier pin only covered the models.dev resolver, but models.dev omits these ids under `opencode-go` entirely, so live `/zen/go/v1/models` discovery had no bundled reference and defaulted them to chat completions. The per-id API pins now also apply inside the discovery mapper, and pinned ids invalidate cached routes written before the pin ([#8957](https://github.com/can1357/oh-my-pi/issues/8957)).
+- Fixed tool-call turn failures for `opencode-go/muse-spark-1.2` and related variants by ensuring API transport pins apply to live discovery and automatically inferring response routes for gateway-first OpenCode models ([#8957](https://github.com/can1357/oh-my-pi/issues/8957)).
-- Future gateway-first OpenCode models (ids the gateway serves before models.dev lists them, like muse-spark-1.2 was) no longer default to chat completions blindly: discovery now borrows the `openai-responses` route from the sibling gateway's catalog or the billing-variant base id (`-free`/`-contributor`). Only the responses signal is borrowed — anthropic transports genuinely diverge across the gateways (e.g. `minimax-m2.5`) and are never inferred.
 
 ## [17.3.8] - 2026-08-19
 

packages/coding-agent/CHANGELOG.md

@@ -4,39 +4,34 @@
 
 ### Added
 
-- Added `composer.shape` setting (`/settings` → Appearance → Composer) to customize the editor's visual layout — rounded box (default), Claude Code rules with the right status group chipped onto the top rule, upstream-pi rules, or borderless — with live layout previews in settings and a new setup-wizard scene. Non-box shapes render the status bar as a plain bottom line (no powerline caps or background) that yields its row to the autocomplete menu.
+- `/cleanse` (and `omp cleanse`) — run the checker/repair loop in-session, with a live status board of running checkers, repair subagents, and token/cost totals.
-- Added `statusLine.contextLine` (Context-Reactive Line): the gauge line between the status groups tracks context usage — `off` (solid accent), `percentage` (used/unused split), `annotated` (default; adds boundary markers where speculative compaction starts and auto-compaction fires), or `embedded` (annotated plus in-gauge percentage/window labels).
+- `omp ps` — interactive monitor for daemon-supervised background processes.
-- Added `omp ps` for inspecting and controlling daemon-broker supervised processes from outside the harness: an interactive alt-screen monitor on TTYs (live table, info/logs views, stop/kill/restart, all-scopes toggle) plus static `--plain`/`--json` listings and `info`/`logs`/`stop`/`kill`/`restart` subactions with `--all`, `--dir`, and `--global` scope selectors. Brokers now record their project directory in `scope.json` so runtime scopes can be mapped back to projects offline.
+- Composer layouts — `composer.shape` picks the editor frame (rounded box, Claude Code rules, upstream-pi rules, borderless), with live previews in `/settings` and the setup wizard.
-- Added `qwenTemplateReasoningEffort` to the `models.yml` `compat` schema, so the auto-enabled Qwen 3.8+ template effort dialect (`chat_template_kwargs.reasoning_effort`) can be switched off per provider/model for strict local servers that reject unknown `chat_template_kwargs`.
+- Context line — `statusLine.contextLine` gauge (`percentage`, `annotated`, `embedded`) showing context usage and compaction boundaries.
-- Added `tokenizer` to custom model and `modelOverrides` configuration. It overrides the catalog-resolved local tokenizer family for a model when a proxy serves a known model id with a different tokenizer.
+- Backgroundable Python — `eval` cells can run async and auto-background like `bash`, with configurable thresholds.
-- Added `extendedContext` setting (`/settings` → Context → General, default on). When off, models with a premium long-context price tier (OpenAI GPT-5.6 Sol/Terra/Luna bill 2x input / 1.5x output above 272K input tokens, on both the API and subscription Codex) are capped at the standard-pricing threshold — they appear as 272K again and compaction fires before a request crosses into premium billing. Toggling mid-session re-clamps or restores the active model's window immediately. Anthropic Claude 4.6+ serves its full 1M window at standard pricing, so no Anthropic model is affected.
+- Local Claude token counting — Anthropic-family tokens now count via a native local tokenizer, and every counter (session maintenance, advisor, stats, context tools) uses the active model's own tokenizer.
-- Added click-to-toggle and drag-to-reorder controls for list-valued `/settings` editors.
+- `extendedContext` setting — pick whether models with premium long-context pricing (272K/1M tiers on Codex-class models) use the extended window or compact early and stay on standard pricing.
-- Added `compaction.asyncEnabled` (Async Compaction, default on): when context enters the band just below the compaction threshold, maintenance speculatively summarizes in the background off a branch snapshot (first configured LLM-backed method — remote, handoff, or soft — isolated from the live turn by a side session id) and holds the armed result; crossing the threshold then splices it in instantly instead of blocking on a summarization round-trip. Armed results are invalidated by branch changes, reset boundaries, model switches that strand provider-native replay payloads, and context growth past `keepRecentTokens` (which re-speculates). The status line pulses the auto-compact icon while a speculation runs and holds it in accent once a result is armed.
+- Speculative compaction — with `compaction.asyncEnabled`, all compaction modes compact in parallel while the session continues, then splice the result in instantly.
-- Added `icon.subscription` and `icon.advisor` symbol theme tokens. In Nerd Font mode, subscription spend renders with `󰙺` (`\u{f067a}`) and advisor spend renders with `` (`\uea70`); in Unicode mode, advisor spend renders with `👁` (e.g. `󰙺 2.67 +  󰙺 0.41` in Nerd Font mode, `S2.67 + 👁 S0.41` in Unicode mode, and `S2.67 + S0.41 (adv)` in ASCII mode).
+- `tokenizer` property on custom models and `modelOverrides` to pin the tokenizer family for proxy models.
+- `qwenTemplateReasoningEffort` in `models.yml` `compat` to disable the Qwen 3.8+ reasoning-effort template parameter for strict local servers.
+- Click-to-toggle and drag-to-reorder for list-valued editors in `/settings`.
+- `icon.subscription` and `icon.advisor` symbol-theme tokens (Nerd Font, Unicode, ASCII).
+
 ### Changed
 
-- The context-reactive status line gained an Embedded mode that absorbs configured context segments into in-gauge percentage/window labels. Annotated and Embedded gauges use `󰕝` for the async-speculation boundary and `󰁨` for the compaction boundary under the Nerd Font symbol preset; Unicode and ASCII keep their existing boundary ticks.
+- Revamped the todo HUD — overall progress renders along the tree-spine connector with smooth completion transitions.
-
+- `/handoff` (and automatic handoff compaction) now compacts in place, replacing the session context instead of forking a new session.
-- Unified inline overlay chrome on the rounded-box style used by the model picker/hub and `/settings`: selectors (theme, thinking, queue mode, show images, login/logout, reset usage, session account, plugins, MCP wizard, history search, branch-from-message, sessions, session tree, debug tools) and the `/cleanse`, `/omfg`, `/btw` run panels now render inside a titled `╭─╮│ │╰─╯` box (new shared `OverlayPanel` container) instead of two full-width horizontal rules.
+- Replaced `compaction.strategy`/`compaction.remoteEnabled` with the ordered `compaction.methodOrder` fallback list.
-- `omp cleanse` and the `/cleanse` slash command now render a live interactive status board with running checkers, repair subagents, tool counts, token/cost totals, and live scrollback in both the CLI and interactive terminal modes
+- Unified inline overlays and selectors (model picker, settings, `/cleanse`) into one titled rounded-box panel style.
-- Replaced the single `compaction.strategy` / `compaction.remoteEnabled` policy with ordered `compaction.methodOrder` preferences. The default now tries OpenAI-compatible server compaction, snapcompact, handoff, shake, then soft compaction; unavailable or failed methods advance through that list.
+- Risk badges and warnings on `/settings` rows, starting with External Thinking.
-- `/settings` rows can now carry a risk note: a warning glyph on the row plus a warning-colored line above the description. `External Thinking` (`externalThinking`, `--external-thinking`) is the first user — providers have flagged the request shape it produces as abuse, up to account-level enforcement, so both the settings entry and `--help` now say so.
-- The todo HUD header now draws a summed progress bar counting closed/total tasks across every stage. Once all tasks close, the bar smoothly collapses before the row disappears.
-- The todo HUD now carries overall progress in the tree spine instead of a horizontal header bar: the top-level connector column runs unbroken from the `TODO` header down the left edge and closes with a short `└────` elbow tail under the block. The closed/total fraction (summed across every stage) fills that path in accent — down the spine, around the bend, out along the tail. Once all tasks close, the accent smoothly drains back up before the header row disappears.
-- The todo HUD now carries overall progress in the tree spine instead of a horizontal header bar: the top-level connector column runs unbroken from the `TODO` header to the last row, and the closed/total fraction (summed across every stage) colors it top-down in accent. Once all tasks close, the accent smoothly drains out of the spine before the header row disappears.
-- Token counting is now scoped to the model being billed rather than to a process-global tokenizer: session maintenance, stats, advisors, `/context`, snapcompact inline imaging, and `compress` each count through the owning agent's `Tokenizer` (`agent.tokenizer`). Message counting is `Tokenizer.countMessage`/`countMessages` (replacing the free `estimateTokens(message, tokenizer)` helper; the legacy shim keeps a compat `estimateTokens` export for legacy pi extensions). `estimateToolSchemaTokens`, `estimateSkillsTokens`, `computeNonMessageTokens`, and `computeNonMessageBreakdown` take an explicit tokenizer; standalone prompt inspection intentionally keeps the default estimate because it has no resolved catalog model.
-- The advisor runtime's `maintainContext` hook now receives the pending update as a message instead of a pre-computed token count — sizing it needs the advisor model's tokenizer, which the host owns.
-- Handoff no longer starts a new session: `/handoff` and the auto-maintenance `handoff` method now commit the generated document as a regular compaction entry on the current session (document becomes the summary, recent history is kept per `compaction.keepRecentTokens`, session id/transcript/cache key unchanged). The `session_before_switch`/`session_switch` extension events no longer fire with reason `"handoff"`, mid-turn maintenance no longer skips the handoff preference, and overflow recovery can now apply a pre-armed handoff result.
 
 ### Fixed
 
-- Fixed GitHub `file_read` failing on image and binary responses that GitHub CLI attempted to transform; it now requests uncompressed JSON content, returns supported images as image blocks, and identifies unavailable or non-text files with a view URL.
+- Subagent `yield` structured results no longer get corrupted by lossy argument repairs; prompt guidance improved for weak callers.
-- Fixed subagent structured returns being silently corrupted by the tool-argument repair layer: the `yield` tool's parameters embed the caller's output schema under an `anyOf` wrapper, and lossy repairs fired on union-branch guesses — JSON-stringifying object payloads into string-typed fields (parents received `summary: "{\"purge\":13,…}"` instead of prose) and deleting unrecognized keys — bypassing yield's own validate-and-retry loop. The repair layer (pi-ai) now restricts union-branch diagnoses to lossless repairs, so mismatches surface as retryable schema errors and accepted payloads arrive verbatim.
+- GitHub `file_read` returns proper image blocks and direct view URLs for image/binary files.
-- Fixed the `yield` tool bouncing common weak-caller envelope shapes with `result must be an object containing either data or error` retries (the dominant structured-output failure in Gemini-flash subagent traces): `type: "result"` with the `result` wrapper omitted entirely now finalizes as the documented last-turn yield, top-level `data`/`error` payloads missing the wrapper are salvaged, and `result` or `data` sent as a JSON-encoded string is parsed losslessly (mirroring executor finalization) before consuming a schema retry.
+- Cancelled prompts during pre-stream turn setup restore the text and image attachments to the editor.
-- Fixed the `yield` tool's instructions teaching weak callers the wrong call shape for structured tasks: the description led with `Pass type:"result" to finalize; when data is omitted, your last assistant turn becomes the raw final result` before ever stating the `result: { data }` wrapper, producing wrapper-less `{type:"result"}` punts and top-level `data:` payloads in Gemini-flash traces. The description (now `prompts/tools/yield.md`) and the subagent system prompt lead with the wrapper contract and only advertise last-turn extraction when no output schema is declared; a schema-bound last-turn finalize with no accumulated incremental sections is rejected in-band as a retryable error instead of terminating the child into an uncorrectable post-mortem `schema_violation`.
+- `top` builtin accepts single-dash macOS flags such as `-pid` and `-stats`.
-- Fixed a prompt cancelled during turn setup (Esc while the pre-stream spinner is up, after dispatch had started) vanishing entirely: it was never persisted to the session — so the `/tree` and `/branch` selectors had nothing to rewind to — and was not returned to the editor either, while its optimistic transcript row kept lingering. A prompt dropped before reaching the agent (abort or usage-preflight denial racing setup) is now handed back: the stale transcript row is removed and the typed text and image attachments are restored to the editor for editing.
+- GNU/BSD compat sweep across built-in shell utilities (`timeout`, `diff`, `find`, `date`, `tail`, `head`, `rg`, `stat`, `truncate`, `cksum`, `sleep`, `which`, `nohup`, `kill`).
-- Fixed macOS `top`-style single-dash long options in the `top` shell builtin: `top -l 2 -pid 56943 -stats pid,cpu,th,mem,pstate` previously failed with `invalid value 'id' for '--pid <PIDS>'` because clap read `-pid` as `-p id`. Single-dash long spellings now parse, and `-stats` selects and orders output columns using macOS stat keys (`pid`, `cpu`, `th`, `mem`, `pstate`, ...).
-- Fixed a sweep of GNU/BSD compatibility gaps in the built-in shell utilities, found by auditing every builtin against its real counterpart: `timeout` gained `-s`/`-k`/`--preserve-status`/`--foreground`/`-v`, GNU exit codes (124/125/137), signal delivery to the child process group with `-k` SIGKILL escalation, and `timeout 0` disabling the limit; `diff` gained normal-format default output, `-w`/`-b`/`-B`/`-i`/`-x`/`-L`/`-s`/`--strip-trailing-cr`, context format (`-c`/`-C`), bundled flags (`-ru`, `-urN`), timestamped unified headers, and no longer recurses directories without `-r`; `find` fixed inverted `-newerXY` timestamp comparisons, anchored `-regex` to whole paths, and gained BSD `-perm +mode`, `-type f,d` lists, `-size` `T`/`P` suffixes, ISO dates in `-newermt`, and BSD leading flags `-E`/`-x`/`-s`; `date` gained BSD `-r <epoch>`, `-v` adjustments, and `-j -f` strptime parsing, and `-I` no longer swallows a following `+FORMAT`; `tail`/`head` accept obsolete `-N`/`+N` counts at any argv position with any file count, `tail -r -n N` works, and `head` continues past per-file I/O errors with GNU header/separator placement; `rg` resolves `-s`/`-i`/`-S` by last occurrence, accepts `--no-config`/`-j`/`--threads`/`--no-column`, implements `--path-separator`, and emits clean NUL-delimited output under `-0`/`-l0`; `stat` prints integer epochs for `%X`/`%Y`/`%Z` (bash arithmetic on `stat -c %Y` works) and gained BSD `-s`/`-x` output modes plus `-t` time formatting; `cksum` is now registered (multi-algorithm `cksum -a sha256`); `truncate` implements `-o`/`--io-blocks` (previously silently truncated to the raw byte count), accepts `b` (512-byte) suffix and BSD `=` prefix; `sleep`/`timeout` accept `infinity` and keep sub-millisecond precision; `yes` and `errno` accept hyphen-prefixed operands (`yes -n`, `errno -2`); `nohup -- cmd` no longer tries to run `--` (including backgrounded via the brush wrapper); `which` gained BSD `-s` and errors on zero operands; `kill` accepts attached values (`-s9`, `-sKILL`, `-l9`) and maps exit statuses above 128 (`kill -l 137` → `KILL`).
 
 ## [17.3.8] - 2026-08-19
 

packages/hashline/CHANGELOG.md

@@ -4,12 +4,12 @@
 
 ### Added
 
-- Added an opener-escape landing correction: a plain `PUT >N:` anchored on a construct's opening line (per tree-sitter) with a body that parses as one self-contained construct claiming a strictly shallower column depth (tab bodies in space files included) is landed after the innermost enclosing construct whose own depth admits the body as a sibling, verified by the syntax probe. Previously such an insert silently split the opener from its body — and could still parse (items are legal inside Rust fn bodies), so no advisory fired.
+- Added an opener-escape landing correction for insertions anchored on a construct's opening line to place shallower sibling constructs after the enclosing block rather than splitting the opener from its body.
 
 ### Fixed
 
-- Dropped one-sided boundary echoes on single-line replacement ranges when every echoed row is an attribute/decorator/annotation node per the tree-sitter grammar (`#[napi]`, `@Injectable()`): the authored result parses, so the syntax probe never fired and the attribute was silently duplicated. Under-filled annotation echoes are now rejected instead of applied.
+- Fixed an issue where single-line replacements echoing attributes or decorators (such as `#[napi]` or `@Injectable()`) could lead to silently duplicated annotations.
-- Raised the default snapshot-store path capacity from 30 to 256 so tags minted early in a wide session no longer age out of the LRU and degrade a recoverable stale-tag mismatch into the misleading "hash is not from this session" rejection.
+- Increased the default snapshot-store path capacity from 30 to 256 to prevent early tags in wide sessions from aging out and triggering misleading "hash is not from this session" errors.
 
 ## [17.3.3] - 2026-08-14
 

packages/natives/CHANGELOG.md

@@ -4,14 +4,14 @@
 
 ### Added
 
-- Added `ClaudeV3`/`ClaudeV47`/`ClaudeV5` encodings to `countTokens`: a Rust rewrite of [ctok](https://github.com/sanderland/ctok) by Sander Land (MIT), reconstructing Anthropic's `count_tokens` offline. Counts are exact on ctok's ~3.4M-response measurement corpora; the port is validated against 493 Python-ctok reference fixtures covering all three families. The pipeline is byte-level throughout — markers occupy one byte, normalization borrows text no rule touches, ASCII and ideographs skip the Unicode tables, and pieces are matched with one Aho-Corasick transition per byte instead of a per-position vocabulary descent — which counts English prose at 64 MiB/s, markdown at 73 MiB/s, source code at 35 MiB/s and CJK at 49 MiB/s per core: 1.5× (CJK, already cheap per byte) to 5.5× (prose, markdown, digits) a straightforward character-level implementation of the same model, which is held to byte-for-byte identical counts across 2.4M randomized differential comparisons.
+- Added offline `countTokens` support for Anthropic Claude families (`ClaudeV3`, `ClaudeV47`, `ClaudeV5`) via a high-performance native port of `ctok`.
-- Added zstd-embedded exact content tokenizers for Qwen 3.5+/3.6+/3.8, DeepSeek V3/V4/R1, Kimi K2/K3, and GLM-5 alongside the rebuilt OpenAI o200k/cl100k and Claude reconstructions. `countTokens` now reads JavaScript strings through a reusable UTF-16 buffer, so native counting does not allocate a UTF-8 temporary.
+- Added exact offline token counting support for Qwen (3.5+, 3.6+, 3.8), DeepSeek (V3, V4, R1), Kimi (K2, K3), and GLM-5 models alongside rebuilt OpenAI encodings, with optimized zero-allocation string passing from JavaScript.
-- Added `nodeChainAt`: the named tree-sitter node chain containing a line, innermost-first, with grammar kind and line span per node. Powers hashline's structural edit repairs (annotation-row classification, opener-anchored insert relocation) without lexical heuristics.
+- Added `nodeChainAt` native API to retrieve innermost-first tree-sitter node chains with grammar kinds and line spans for structural syntax analysis.
 
 ### Changed
 
-- Shell builtin utilities now stream their output. Utilities that emit progressively (`grep`, `rg`, `sed`, `cat`, `head`, `tail`, `cut`, `date`, `uniq`, `comm`, `jq`, `ls`, `fd`) write through a destination-aware buffer: line-buffered to pipes/terminals so each completed line reaches the TUI's live tool output (or the next pipeline stage) as it is produced, block-buffered to regular files for throughput. Previously they held everything in an exit-flushed 8–32 KiB `BufWriter`. `rg --line-buffered`/`--no-line-buffered` still force a policy. Builtin stderr goes through the same policy, and when fd 1 and fd 2 share a destination (`2>&1`, or the default merged capture pipe) both streams funnel through one serialized writer, so diagnostics interleave with output in exact write order.
+- Improved shell builtins (`grep`, `rg`, `sed`, `cat`, `head`, `tail`, `jq`, `ls`, etc.) to stream output progressively with destination-aware line buffering for pipes, terminals, and live TUI output, while maintaining block buffering for file writes.
-- Compound (`{ …; }`, `(…)`) and shell-function pipeline stages now run concurrently with the rest of the pipeline, like builtin and external stages already did. They previously executed inline while the pipeline was being spawned, which delayed all downstream output until the stage exited and deadlocked the shell when a stage produced more than a pipe buffer with no reader running (e.g. `{ seq 1 200000; } | head -n 1`).
+- Updated compound blocks (`{ ...; }`, `(...)`) and shell-function pipeline stages to run concurrently with other pipeline stages, preventing head-of-line blocking and pipe buffer deadlocks.
 
 ## [17.3.8] - 2026-08-19
 

packages/stats/CHANGELOG.md

@@ -4,11 +4,11 @@
 
 ### Changed
 
-- Window token estimates now use broker-held fleet token burn (per-client observed-usage reports) when an auth broker is configured, matching the fleet-wide window fractions instead of undercounting with local-only message stats.
+- Window token estimates now incorporate broker-reported fleet token burn when an auth broker is configured, accurately tracking fleet-wide usage instead of undercounting with local-only statistics.
 
 ### Fixed
 
-- Fixed subscription-window insights merging distinct limits that share a duration label (Anthropic's overall vs model-scoped 7-day windows, Codex base vs Spark weeklies); interleaved fractions inflated window-equivalents consumed by orders of magnitude and broke tokens-per-window estimates. Windows now group by provider limit id.
+- Fixed an issue in subscription-window insights where distinct limits sharing a duration label (such as Anthropic overall vs. model-scoped 7-day windows) were incorrectly merged, which inflated window-equivalents and skewed tokens-per-window estimates. Windows are now grouped by provider limit ID.
 
 ## [17.3.6] - 2026-08-17
 

packages/tui/CHANGELOG.md

@@ -4,8 +4,8 @@
 
 ### Added
 
-- Added composer border styles (box, claude, pi, borderless) as per-style `ComposerStyle` objects (`getComposerStyle`) owning chrome geometry, top/bottom/row rendering, and status-bar attachment metadata — shared by the editor and host previews so they cannot drift
+- Added composer border styles (`box`, `claude`, `pi`, `borderless`) via `ComposerStyle` objects and `getComposerStyle`, unifying chrome geometry and rendering across the editor and previews.
-- Added support for warning risk notes and row markers in settings lists
+- Added support for warning risk notes and row markers in settings lists.
 
 ## [17.3.8] - 2026-08-19