The `/tree` "Summarize" option ignored the session's thinking level.
`AgentSession` already passes `thinkingLevel` into the
`generateBranchSummary` options, but `generateBranchSummary` never read
it, so the request was built without `reasoning`.
For reasoning models on OpenAI Responses and Azure Responses, a missing
`reasoning` becomes `effort: thinkingLevelMap.off ?? "none"`. Models
that don't accept `"none"` then reject the branch summary with a 400
error:
```
Unsupported value: 'none' is not supported with the 'gpt-6-astra-…' model.
```
This change adds `thinkingLevel` to `GenerateBranchSummaryOptions`. It
sets `reasoning` on the request when the model supports reasoning and
the level is not `off`, the same rule compaction uses in
`createSummarizationOptions`.
**Testing**
- Added a test to
`packages/coding-agent/test/branch-summarization.test.ts` that records
the stream options for three cases: a reasoning model at `xhigh`, a
reasoning model at `off`, and a non-reasoning model. The expected
`reasoning` values are `["xhigh", undefined, undefined]`. Without the
fix, the test fails and every case gets `undefined`. With the fix, it
passes.
- Ran `branch-summarization`, `agent-session-branching`,
`branch-summary-extensions` and `compaction-summary-reasoning`: all
pass.
- `npm run check` passes.
- Not tested against a real provider endpoint.
Fixes#11444
Requested by @davidbrai
Co-authored-by: triagent <triagent@localhost>
Retry Bedrock strict tool limit rejections once with strict: "prefer" tools non-strict, as the Anthropic provider already does. Share the limit error messages and the prefer check between both providers.
closes#11416
GPT-5.5 and older namespace tools loaded through additional_tools or tool
search and reject replayed calls without the namespace, so switching to them
after another model's tool calls failed with 400. Drop dynamic tool loading
for these models on OpenAI and OpenCode; keep mid-conversation system messages.
closes#11415
closes#11132
pi sends every opencode and opencode-go request to the shared base
(`https://opencode.ai/inference/v1`, `/inference/go/v1`), Jev included.
this switches them to the documented protocol-specific endpoints, which
is also what models.dev lists per model:
| api | before | after |
|---|---|---|
| openai-completions / openai-responses | `/inference/v1` |
`/inference/openai/v1` |
| anthropic-messages | `/inference` | `/inference/anthropic` (SDK
appends `/v1/messages`) |
| google-generative-ai | `/inference/v1` | `/inference/google/v1beta` |
| jev-1.13 / jev-1.13-free | `/inference/v1` | `/inference/systemone/v1`
|
go uses the same paths under `/inference/go/...`.
changes: base URLs in `generate-models.ts`, the jev URL expectations in
`test/classifier-models.test.ts`, and a stale comment in
`providers/opencode.ts`. `models.generated.ts` has no diff since the
catalog data is gitignored.
verified with the regenerated catalog:
- all 111 opencode/opencode-go models hit their new URL, none 404
- real requests with an OpenCode key returned replies on
anthropic-messages, google-generative-ai, openai-responses,
openai-completions and jev
- go returned 403 "subscription required" from the new paths (my key has
no Go sub), so routing works there too
`npm run check` passes. `./test.sh` passes for ai and coding-agent; the
only remaining failures were timeouts that pass in isolation and pi-env
tests that need cargo locally.
---
_Drafted with AI assistance, reviewed by me._
The Node driver takes the count from node:sqlite; the Durable Object driver reads
SELECT changes() after the statement, since the cursor's rowsWritten also counts
index writes (verified in workerd). run is unchanged and costs nothing more.
Inline `ctx.ui.custom()` components never report pi as blocked through
OSC 7501 program status. This is on purpose, because the same API hosts
loaders. As a result, a custom dialog that waits for the user (an
approval prompt, the `question.ts` example) shows up as `working`, and
the terminal cannot tell that pi needs input.
The extension dialog stack already stores a blocked status for each
dialog and reports the visible one. This change adds an optional
`blocked` option to `ctx.ui.custom()`:
```ts
ctx.ui.custom(factory, { blocked: { kind: "question", message: "Pick a color" } });
```
The inline branch of `showExtensionCustom()` passes it to
`pushExtensionDialog()`. Components without the option behave as before.
Overlays ignore the option because they are not part of the dialog
stack. `docs/tui.md` documents the option.
Testing:
- Added a test to `test/interactive-mode-extension-dialogs.test.ts`:
- A custom component that opts in reports `blocked`.
- A plain custom component stacked on top of it clears the blocked
status.
- Closing the plain component restores the blocked status.
- Closing the opted-in component clears it.
- The test fails without the change to `interactive-mode.ts` and passes
with it.
- `npm run check`: biome and lint pass, and there are no type errors in
coding-agent. The only errors are existing type errors in
`packages/ai/test` from locally generated model data. They are not
related to this change.
Fixes#11428
Requested by @badlogic
Co-authored-by: triagent <triagent@localhost>
When thinking was hidden and the hidden-thinking label was set to `""`
(for example with `ctx.ui.setHiddenThinkingLabel("")`), every assistant
message that contained thinking still drew blank rows:
- thinking followed by a tool call drew 2 blank rows;
- thinking followed by text drew 3 blank rows before the text.
`AssistantMessageComponent.updateContent()` had three causes:
- It counted hidden thinking as visible content when deciding on the
leading and trailing spacers.
- It always added the label component. With an empty label the theme's
escape codes are still there, so the component rendered one blank row.
**What changed**
- A hidden thinking run with an empty label is now skipped entirely. Its
run index is still advanced, so click overrides on other runs keep their
indices.
- Spacers are now added lazily, just before the next visible block. The
result is the same as before: one spacer before the first visible block,
and one after a visible thinking run that is followed by more visible
content. But spacing now follows what actually renders.
- Default behaviour (the "Thinking..." label, or expanded thinking) is
unchanged.
**Testing**
- Added two tests to `test/assistant-message.test.ts`:
- a hidden thinking run with an empty label renders nothing, both before
a tool call (`[]`) and before text (`["", "answer"]`);
- with the default label, rendering still gives `["", "Thinking...", "",
"answer"]`.
- The empty-label test fails without the fix and passes with it.
- `npm run check` passes. The `test/assistant-message.test.ts` and
`test/interactive-mode*` tests pass.
- Rendering the component directly, with hidden thinking and an empty
label:
- thinking + tool call: before, 2 blank rows; after, `[]`;
- thinking + text: before, 3 blank rows then the text; after, `["", "
hi"]`.
Fixes#11402
Requested by @badlogic
Co-authored-by: triagent <triagent@localhost>
`pi install`, `pi remove` and `pi update --extensions` run npm without
`--no-audit`, so npm prints its "N vulnerabilities … run `npm audit
fix`" advice after every install. pi gives users no way to clear that
warning, and following it causes harm. In `~/.pi/agent/npm`, `npm audit
fix` runs without `--legacy-peer-deps`, so npm installs the
host-provided `@earendil-works/pi-*` peers there: a second copy of pi.
The comment above `getNpmInstallArgs` warns against exactly that.
The change adds `--no-audit --no-fund` wherever pi already passes
`--legacy-peer-deps` to npm:
- `getNpmInstallArgs` (npm package installs and updates)
- `getGitDependencyInstallArgs` (dependency installs for git packages)
- npm uninstall
pi's managed self-update (`runManagedNpmCi`) already passes both flags.
bun and pnpm arguments are unchanged.
I did not edit `CHANGELOG.md`, per CONTRIBUTING.md.
**Testing**
- Updated the expected npm argument arrays in
`test/package-manager.test.ts`. With the source change reverted, 11
tests fail; with it, all 132 pass.
- `npm run check` passes.
- Ran the built CLI by hand with `PI_CODING_AGENT_DIR=/tmp/pa pi install
npm:lodash@4.17.20`:
- Before the change, the output ended with npm's "To address all issues,
run: npm audit fix" block.
- After, it shows only "added 1 package" and "Installed
npm:lodash@4.17.20".
Fixes#11409
Requested by @badlogic
Co-authored-by: triagent <triagent@localhost>
With the kitty keyboard protocol active, these function keys were not
recognized:
- F1, F2 and F4 arrive as `CSI P`/`Q`/`S`, or as `CSI 1;<mod> P/Q/S`.
- Any F-key pressed with a modifier, e.g. `CSI 15;5~` for Ctrl+F5.
`matchesKey` returned false whenever a modifier was set and only
compared against the legacy sequences. `parseKittySequence` did not know
any F-key forms. So shortcuts bound to keys like `f1` or `ctrl+f5` never
fired.
Separately, F13–F35 (and other kitty functional keys) arrive as CSI-u
codepoints in the Private Use Area. `decodeKittyPrintable` returned them
as text, so the editor inserted invisible U+E020+ characters into the
prompt.
Changes in `packages/tui/src/keys.ts`:
- Add F1–F12 internal codepoints. Map `~` numbers 11–15, 17–21, 23 and
24 to them.
- Parse the `CSI [1;mod[:event]] P/Q/S` form. F3 is left out because
`CSI 1;<n>R` collides with cursor position reports, so kitty sends F3 as
`CSI 13~`.
- Let the `f1`…`f12` cases in `matchesKey` fall back to
`matchesKittySequence` so modifiers work. Name the keys in `parseKey`
output.
- Make `isKeyRelease` and `isKeyRepeat` recognize `:3`/`:2` events on
P/Q/S, so F1/F2/F4 releases don't trigger shortcuts a second time.
- `decodeKittyPrintable` now rejects the kitty functional-key PUA range
(57344–63743). Keypad keys are normalized to ASCII first, so they still
decode.
Testing:
- Added tests in `packages/tui/test/keys.test.ts`. They cover kitty
F1–F12 with and without modifiers, F3 not being confused with a cursor
position report, release and repeat detection, and F13+ not decoding as
text. Two of the new tests fail without the change; all pass with it.
- Ran `npm test` in `packages/tui` and `npm run check`; both pass.
- Before the change: `parseKey("\x1b[P")` and `parseKey("\x1b[1;5P")`
returned `undefined`, `matchesKey("\x1b[15;5~", "ctrl+f5")` returned
false, and `decodePrintableKey("\x1b[57376u")` returned `"\ue020"`.
- After the change: these give `f1`, `ctrl+f1`, true and `undefined`.
Fixes#11419Fixes#11345
Requested by @badlogic
Co-authored-by: triagent <triagent@localhost>
The `openai` provider still defaulted to `gpt-5.5`. The same default
applies to API-key logins and to "Sign in with ChatGPT", so logging in
through ChatGPT selected `openai/gpt-5.5`. The reporter says that model
no longer works with a ChatGPT subscription. The `openai-codex` and
`github-copilot` providers already default to `gpt-6.1-sol`.
This change sets `defaultModelPerProvider.openai` to `gpt-6.1-sol` in
`packages/coding-agent/src/core/model-resolver.ts`. The new value
applies in three places: the model picked after `/login` when no model
was set, the fallback model at startup, and the default model for
`--provider openai`. API-key users get the new default too.
I did not test against a real ChatGPT account, so the claim that the
subscription rejects `gpt-5.5` is not checked here.
Testing:
- Updated the assertion in `test/model-resolver.test.ts` ("openai
defaults track current models"). It now expects `gpt-6.1-sol`; the
model-resolver tests pass.
- `npm run check` passes.
- `pi --list-models gpt-6.1` still lists `openai/gpt-6.1-sol`, so the
new default is in the catalog.
Fixes#11442
Requested by @badlogic
Co-authored-by: triagent <triagent@localhost>
`ctx.ui.setEditorComponent()` always cleared the editor slot and put the
new editor there with focus. If an extension dialog
(`confirm`/`select`/`input`/`editor`, or an inline `custom()` component)
was open at that moment, the editor replaced it on screen. The dialog
was still on the dialog stack and its promise still pending, but nobody
could see or answer it. It hung until its timeout, or forever without
one.
Now `setCustomEditorComponent` leaves the editor slot and focus alone
while a dialog is open. The new editor is still created and wired up,
and `showTopExtensionDialog` shows it when the last dialog closes.
Without an open dialog nothing changes.
Testing:
- Added a regression test to
`test/interactive-mode-extension-dialogs.test.ts`. It opens a confirm,
replaces the editor, then checks that the confirm stays visible and
answerable and that the new editor is focused afterwards. It fails
without the fix and passes with it.
- `npm run check` passes.
- Ran the reproduction from the issue in tmux (`pi -e ./a.ts -e ./b.ts`,
where a.ts opens a non-blocking `confirm` in `session_start` and b.ts
calls `setEditorComponent`).
- Before: only the editor shows, and the confirm times out.
- After: the confirm shows, and Enter gives `answer: true`.
Fixes#11431
Requested by @badlogic
Co-authored-by: triagent <triagent@localhost>
When `context` is set, the grep tool reads the file again to print the
lines around each match. It turned CRLF into LF and also turned every
lone CR into LF before splitting the text into lines. ripgrep, which
supplies the match line numbers, counts lines only at LF. A lone CR
therefore added an extra line on pi's side, so every later line was off
by one.
For example, in a file containing `before\rprogress\nneedle\nafter`, a
search for `needle` with `context: 1` showed `progress` as the line-2
match and left out `after`.
The fix keeps the CRLF normalization and splits only on LF, so pi
numbers lines the same way ripgrep does. Displayed lines already have
`\r` removed, so the line before the match now shows as
`beforeprogress`.
**Testing**
- Added a test in `packages/coding-agent/test/tools.test.ts` for a file
with a lone CR searched with `context: 1`. It fails without the fix and
passes with it.
- `test/tools.test.ts`: all 87 tests pass.
- `npm run check` passes.
- Reproduction from the issue. Before the fix:
```
f.txt-1- before
f.txt:2: progress
f.txt-3- needle
```
After the fix:
```
f.txt-1- beforeprogress
f.txt:2: needle
f.txt-3- after
```
Fixes#11398
Requested by @badlogic
Co-authored-by: triagent <triagent@localhost>
The root and ./testing exports of pi-telemetry had no source
condition, so Durable's and pi-env's tests, which resolve with source,
failed to import it without a build: the Environments workflow has
failed on every run since 2026-10-18.
Durable publishes a stored commit inside Storage.commit's adopt: the
telemetry events, its own observers (document watches, conversation
views, task graph, scheduler, waits), and a dispatch of host
subscribeCommits listeners one microtask later, before every wait and
abort the commit triggers. A commit the storage rejects after adoption is
published before the Session fails; a throwing telemetry adapter is
reported as host_error "telemetry" instead of rejecting a durable commit.
Chord's watch listeners take one change object, { value, ops, retired }
({ value } for native watches); retirement is a flag, not an end.
Durable's document watches follow their address, and a task document's
watch ends after its retirement call. New synchronous observers for
adapters: observeDoc, observeConversation, observeTaskGraph; the watches
are built on them. The env file watcher's change type is renamed
FileChange.
Add gpt-6.1-terra to the OpenAI, Azure OpenAI Responses, and OpenAI
Codex providers with official pricing ($1 input, $0.05 cached input,
$1.25 cache writes, $5 output, long-context tier above 272k), tool
search, additional tools, and mid-conversation effort changes, as for
GPT-6.1 Sol. Unlike Sol it accepts reasoning.effort "none", so off is
supported; both reject "minimal" (verified against Codex). Codex serves
it with a 500K context window.
- ai: OAuth logins only accept http(s) URLs from servers (shared
trustedHttpUrl, now also checked by Radius); pi-messages rejects
contentIndex values that are not the next or an existing block; the
CLI auth file is written with mode 0600
- ai: OpenCode DeepSeek V4.1 Flash (renamed from V4 Flash upstream) no
longer sends prompt_cache_retention
- tui: setTitle() replaces control characters; export
replaceControlCharacters()
- coding-agent: the login dialog only opens web URLs and never prints
control characters from sign-in, device-code, or info URLs
- durable: partial commits no longer merge a new __proto__ member into
Object.prototype
Split build-check-test into build-check plus a test matrix (coding-agent,
durable, other). Install only fd and ripgrep, and only in jobs whose tests
use them; the cairo/pango/jpeg/gif/rsvg dev libraries were for compiling
canvas, which npm ci --ignore-scripts never does.
A query that equals a tool's full name or the part after a __ separator
(thread_read for mcp__app__thread_read) now ranks that tool first in
tool_search and codemode searchTools(). BM25 length normalization could
otherwise rank a shorter description sharing the name's words above it.
closes#11331
`scripts/check-entry-graphs.mjs` built the list of files each export
reaches with `path.relative()`, which returns paths with `\` separators
on Windows. The allow and forbid patterns are written with `/`, which
caused two problems on Windows:
- The allow entry `packages/telemetry/src/index.ts` never matched, so
the bare `index.ts` forbid pattern flagged it and the check failed even
though nothing was wrong.
- Forbid patterns that contain a `/` (`providers/`, `runtime/host.ts`,
...) never matched, so real violations went unreported.
The fix converts the relative paths to `/` before any comparison. On
Linux and macOS nothing changes.
Testing:
- The scripts have no unit tests, so no test was added.
- `node scripts/check-entry-graphs.mjs` and `npm run check` pass on
Linux.
- I simulated Windows path handling with `path.win32`, without running
Windows itself:
- Before the fix: `relative(...)` gives
`packages\telemetry\src\index.ts`, and
`endsWith("packages/telemetry/src/index.ts")` is false.
- After the fix: the path becomes `packages/telemetry/src/index.ts`, and
the suffix check is true.
Fixes#11333
Requested by @badlogic
Co-authored-by: triagent <triagent@localhost>
`--model "<provider>/<id>"` returned "not found" when the provider id
itself contains a `/`. URL-derived ids such as
`llama-server=http://host:10800` are one example.
`splitProviderPrefix()` in `model-resolver.ts` split the reference at
the first `/`, which turned the prefix into `llama-server=http:`. No
provider has that id, so resolution fell through to not-found.
`splitProviderPrefix()` now compares the text against the provider ids
in the catalog. A provider matches when the text starts with its id
followed by `/`, and the longest matching id wins. Ordinary provider ids
resolve exactly as before.
Testing:
- I added two cases to the `resolveModelReference` table in
`test/model-resolver.test.ts`, using a provider
`llama=http://host:10800`: one plain, one with a `:high` thinking
suffix. Both fail without the change and pass with it. The test runs
every case against the catalog in both orders.
- `node ../../node_modules/vitest/dist/cli.js --run
test/model-resolver.test.ts`: 64 passed.
- `npm run check`: passes.
- Before the change, a standalone check of `resolveModelReference` with
the reporter's exact provider and model
(`llama-server=http://host:10800/unsloth/Qwen3.8-27B-GGUF:BF16`)
returned `not-found`. It now returns `resolved`.
- One test, `test/model-runtime-classifiers.test.ts` ("applies provider
compat from models.json to classifier models"), fails in my environment
both with and without this change. It is unrelated to this fix.
Fixes#11394
Requested by @badlogic
Co-authored-by: triagent <triagent@localhost>
When the agent directory is reached through a symlink (for example
`~/.pi` linked into a dotfiles repo), pi passed the symlinked path as
`--prefix` to `npm install` and `npm uninstall`. npm then writes lock
entries such as `../../../dotfiles/pi/agent/npm/node_modules/foo`
instead of `node_modules/foo`. Each later install adds one more nesting
level, so `package-lock.json` keeps growing.
This change resolves the install root with `realpathSync` before passing
it to the package manager:
- in `getNpmInstallArgs()`, which covers install and update for npm,
pnpm and bun;
- in `uninstallNpm()`.
The paths pi loads extensions from are unchanged.
**Testing**
- Added a regression test in `test/package-manager.test.ts`. It uses a
symlinked agent dir and checks that install and uninstall get the real
path as `--prefix`. The test fails without the fix and passes with it.
- Wrapped the existing `--prefix`/`--cwd` expectations in
`realpathSync(...)`. They now match on systems where the temp dir is
itself a symlink, such as macOS.
- Ran `npm run check` (clean) and `test/package-manager.test.ts` plus
`test/package-manager-ssh.test.ts` (all pass).
- Manual reproduction with `~/.pi` as a symlink, running `pi install
npm:ms` several times:
- before: lock keys grew to
`../../../dotfiles/pi/agent/npm/node_modules/ms`,
`../../../dotfiles/dotfiles/pi/...`, and so on;
- after: `[ '', 'node_modules/ms' ]`, and `pi remove npm:ms` leaves `[
'' ]`.
Fixes#11358
Requested by @badlogic
Co-authored-by: triagent <triagent@localhost>
`./pi-test.sh --no-env` is supposed to run pi without any API keys, but
its hard-coded `unset` list had fallen behind
`packages/ai/src/env-api-keys.ts`. Credentials such as
`DEEPSEEK_API_KEY`, `MOONSHOT_API_KEY` and `ANTHROPIC_AUTH_TOKEN` were
still picked up from the environment.
This change adds the 26 missing variables to the list:
- the Anthropic auth token and federation variables
- the API keys for Ant Ling, Baseten, Cloudflare, DeepSeek, Fireworks,
Google Cloud, Kimi, Meta, Moonshot, Nvidia, Qwen token plan, Radius,
Together, Typesafe, Xiaomi (all plans) and ZAI coding CN
The list now covers every variable name in `env-api-keys.ts`.
Testing:
- No automated test covers this shell script, so none was added.
- A loop over every variable name in `env-api-keys.ts` confirmed each
now has an `unset` in `pi-test.sh`.
- Before the change, `PI_CODING_AGENT_DIR=$(mktemp -d)
DEEPSEEK_API_KEY=x ./pi-test.sh --no-env --offline -ne --list-models
deepseek` listed DeepSeek models.
- After the change, the same command with both `DEEPSEEK_API_KEY` and
`MOONSHOT_API_KEY` set prints "No models available."
- `bash -n pi-test.sh` and `npm run check` pass.
Fixes#11359
Requested by @badlogic
Co-authored-by: triagent <triagent@localhost>
A commit may be marked deferrable ({ persistence: { kind: "deferred", withinMs? } }) on Session, Conversation,
TaskRuntime and a tool's api.commit. When every write changes an existing conversation document with history
"latest" and fork "initial", it is adopted in memory and shown to watches at once, and stored with the next commit
that reads or writes the document, before the document leaves memory, at close, or within withinMs (default
settings.progress.persistWithinMs, 1000 ms). The generation partial throttle and the tool output throttle use it.
subscribeCommits() gets these commits only once stored, with exactly what was stored; internal observers get
memory-only publications. Spec §1/§13 changed, new §4.1. 32 streams over SQLite: storage transactions 562/s -> 99/s,
disk 14.3 -> 4.9 MB/s, CPU 29% -> 22% of one core. A crash loses up to about 1 s of streamed content.
Tracker revisions, applyImmutable() and applyImmutableBatches() copy each container on a changed path with
spread (Object.assign() onto a null-prototype object) instead of one Object.defineProperty() per key: same own
data properties, prototype and own __proto__ member, about 25 times faster per container. The copy-count test
counts key listings through a proxy instead of spying on Object.keys.
Adds a tracker GC regression test: 10,000 adopted changes retain only the document, not a context registration
per change. Follow-up to #9265.
Kimi Coding, Moonshot, and OpenCode Go generated K3.1 Pro and K3.1 Pro Highspeed with Kimi K2 settings because K3 detection matched exact IDs. Detect the K3 family by ID pattern, and Kimi Coding K3-style models by their effort options. Add implied subscription costs and clamp gateway output limits like K3.
- chord/delta: PendingOps (push, size, take -> PendingTake): raw pending ops, r drops everything
before it, a leaf index merges appends and primitive sets per path (cleared by any other op),
past 10,000 ops only "replace with the current value"; one batch is kept by reference and
received ops are never mutated
- chord: Watch queues frames in PendingOps while its listener is busy or not started and delivers
one call with the newest value, the folded ops (or [["r", newest]] after an overflow) and the
newest frame's context; replaces the 100-frame collapse. NativeWatch is unchanged and takes an
optional replace()
- durable: watchEvents keeps every event batch in order (NativeWatch with a snapshot replace);
document, view and task-graph watches fold; spec, README, example 04 and tests updated
- tests: fold fuzzers (tracker, appendText, splices/reorders, root arrays, diffRevisions) with
codec round trips, adversarial cases, cap, ownership, cost and GC guards
The suite's evicting run (PI_DURABLE_DOCUMENT_RETENTION_MS=0) evicts the unwatched pi.live after every
commit, so each partial commit replayed the document's delta chain and the per-commit cost grew with the
stream. Pin the document retention in that test so it measures appends, as the default retention does.
Follow-up to #9265.
- chord: Delta appendText(draft, key, text) / StringKeys<T> record an append op without the
O(length) prefix compare that a plain `draft[key] += text` costs in prepare()
- durable: partial commits append text and thinking deltas with appendText (assignJson plan,
checked by length), and apply tool-call argument changes from their own StreamingJsonParser
of the raw toolcall_delta text instead of reassigning the arguments snapshot
- durable "." entry-graph budget 73 -> 76 for the parser it now imports
- ai: StreamingJsonParser / ToolCallArgumentsStream read each argument delta once (tolerant,
immutable snapshots, linear cost); every provider, pi-messages and the faux provider use it
and keep stream state (partialJson, partialArgs, index, customInput, streamIndex, Bedrock
reasoning bytes) off the content blocks; grammar tool input appends per delta
- agent: streamProxy uses the same parser
- chord: Delta append ops own their text (ownedSlice), so ops and replicas no longer keep the
whole growing string alive
- durable, ai, mcp: bounded error excerpts and summaries own their characters
- coding-agent: tool calls render streamed arguments once per frame, not once per delta
Placeholder rows below a Kitty image line can contain box padding,
borders, or side-by-side text. Stopping the backward search at the
first non-blank row dropped those images once their first row scrolled
above the viewport. Search back at most the tallest registered image
height instead; the existing hiddenRows check keeps results correct.
Image placeholder rows inside padded or colored boxes contain spaces and
SGR codes, which stopped the backward search for the image line, so the
whole image disappeared once its first row scrolled out of view.
With a sessionId, the faux provider reported newly cached prompt tokens
both as input and as cacheWrite, so input + cacheRead + cacheWrite
exceeded the estimated prompt. Every prompt token now goes either to
cacheRead or to cacheWrite, and input is 0 while caching is simulated.
