Fixes OpenRouter rejecting requests where max_tokens equals the served
context window. Prefer top_provider.context_length on discovery and clamp
max_tokens to ContextWindow minus a proportional reserve (window/8 capped
at 4096). Reworked from the original PR so the reserve derives from the
window, not MaxOutput: models whose output already fits are untouched and
small-window models are not over-penalized.
Co-authored-by: Neil-urk12 <neil-urk12@users.noreply.github.com>
OpenRouter enforces input + max_output <= served context_length and
rejects requests where max_tokens equals the whole window, which happens
for models whose catalog MaxOutput is set equal to ContextWindow (e.g.
nemotron-3-super-120B). Two parts:
- discover.go (from #24): prefer top_provider.context_length when it is
smaller than the inflated model-level context_length, so ContextWindow
reflects the limit OpenRouter actually serves.
- openai.go: clamp max_tokens to ContextWindow minus a reserve. The
reserve is derived from the window (window/8, capped at 4096), never
from MaxOutput, so models whose output already fits the window are
untouched and small-window models (gpt-4) are not over-penalized.
Adds buildRequest clamp tests (fits-window no-op, large-window cap,
small-window proportional reserve, floor, explicit-request passthrough)
and an httptest-based DiscoverOpenRouter test for the served-context
preference.
Co-authored-by: Neil-urk12 <neil-urk12@users.noreply.github.com>
Thread the resolved model's catalog MaxOutput through to
provider.Request.MaxTokens so each turn requests the model's full output
capacity. Fixes Bedrock silently truncating long writes/edits at its
4096 default (stopReason=length). Other providers already defaulted to
MaxOutput on a zero request, so this is a no-op for them. Also surfaces
StopLength explicitly in the TUI instead of ending silently.
Co-authored-by: Raymond Gasper <raymondgasper@fastmail.com>
Turns omitted MaxTokens on the provider request, so Bedrock applied its
conservative 4096 default and silently truncated long writes/edits with
stopReason=length. In the TUI this read like the interaction timed out.
Thread the resolved model's catalog MaxOutput through to the request:
catalog Model.MaxOutput -> Resolved.MaxOutput -> Agent.MaxTokens
-> provider.Request.MaxTokens
Zero still falls back to each provider's own default, so models without a
catalog MaxOutput are unaffected. The SDK path inherits this via NewAgent.
Also surface StopLength explicitly in the TUI ('response hit the output
limit -- ask it to continue') instead of ending silently.
Tests: TestAgentPropagatesMaxTokens (Agent.MaxTokens reaches the wire) and
TestBedrockBuildRequestMaxTokens (non-zero flows through; zero -> 4096).
- it currently rejects requests where input + max_output exceeds the
serving provider's context lmit (which may be tighter than the
model-level value)
- use the smaller of ContextWindow and MaxOutput as the cap, with a
4096-token input reserve
Run repairToolUseResultPairs on outbound messages in oneTurn so an
in-process aborted turn (cancel, connection drop, ECONNREFUSED) no
longer leaves a dangling tool_use that gets rejected by Anthropic/OpenAI
on the next request. Pure and idempotent, no-op on valid transcripts.
A turn aborted mid-flight (cancel, connection drop, dev-server
ECONNREFUSED) can leave an assistant tool_use block with no matching
tool_result in the live transcript. repairToolUseResultPairs already
fixes this, but only ran in OpenSession (load time), so an in-process
abort left the transcript broken until restart. The next request was
then rejected by Anthropic/OpenAI with 'tool_use ids were found without
tool_result blocks'.
Run the same repair on the outbound messages in oneTurn. It is pure and
a no-op on valid transcripts, so there is no hot-path cost beyond a
single linear scan and no behavior change for healthy sessions.
Allow extensions to open panels outside slash-command responses, enabling
human-in-the-loop tool gates and secret/input collection patterns. Removes
the review-only spontaneous panel plan before merge.
Co-authored-by: Raymond Gasper <raymondgasper@fastmail.com>
- Footer now shows '● this panel has focus' so users know keypresses
are going to the panel, not the editor
- Prompt line uses '› approve this action? [y/n]' cursor glyph
- Unhandled keys re-render with '› unrecognised key — press y or n'
instead of silently swallowing the input
- docs/extensions.md: add open_panel spontaneous frame section with
blocking tool pattern explanation, concurrent-panel note, and
references to new examples; add approve/secret to See also list;
add roadmap entry
- examples/extensions/approve/: approve_action tool — opens a y/n
panel from inside the tool handler, blocks until user responds
- examples/extensions/secret/: fetch_with_password tool — masked
password input panel, secret never leaves the extension process
Allow extensions to emit an open_panel frame at any time, not just as
the action of a command_response. This makes it possible to build
approval gates, secret collection, and freeform user-input prompts
directly inside tool handlers.
Changes:
- extproto: add OpenPanelFromExt wire type
- extensions/manager: route spontaneous open_panel frames to hooks.OpenPanel
- ext/ext.go: add Extension.OpenPanel() SDK method
- tests: TestSpontaneousOpenPanel (manager), TestOpenPanelEmitsCorrectFrame,
TestBlockingToolWaitsForPanelKey, TestBlockingToolDenied (SDK)
- docs/plans: add spontaneous-panel.md design doc
The blocking tool pattern (open panel → block on channel → key event →
tool_result) requires no additional wire changes; it falls out of
standard Go concurrency on the extension side.
Part 3 (intercept timeout for built-in tool gating) is out of scope
and tracked separately.
Inject a stub toolConfig when the message history contains toolUse or
toolResult blocks but req.Tools is empty (e.g. the /btw side-chat sends
the frozen main transcript). Bedrock's Converse API otherwise rejects
the request with HTTP 400. Bedrock-only; other providers unaffected.
Co-authored-by: Raymond Gasper <raymondgasper@fastmail.com>
Bedrock's Converse API returns HTTP 400 with "toolConfig field must be
defined when using toolUse and toolResult content blocks" whenever the
message history contains toolUse or toolResult blocks but toolConfig is
absent from the request.
The /btw side-chat sends the frozen main transcript as context with no
tools defined. If the main conversation included tool calls the serialised
messages will contain toolUse/toolResult blocks, triggering the 400.
Fix: add bedrockMessagesHaveToolBlocks() to detect this case and, when
req.Tools is empty but tool blocks are present in the history, inject a
minimal stub toolConfig with an inert placeholder tool. Bedrock accepts
the request and the stub can never be invoked since no tool_use stop
reason can fire when the advertised tool list is effectively empty.
Adds Converse API cachePoint blocks at the system prompt boundary and on
the last user message for Bedrock Claude models (PriceCacheWrite > 0),
mirroring the Anthropic provider's caching strategy. Nova models are
excluded (automatic caching).
Co-authored-by: Raymond Gasper <raymondgasper@fastmail.com>
Place Bedrock Converse API cachePoint blocks at the system prompt
boundary and after the last user message on every request to Claude
models (those with PriceCacheWrite > 0 in the catalog).
This mirrors the existing Anthropic provider strategy (cache_control:
ephemeral on system, tools, and last user message) using Bedrock's
equivalent syntax: a {"cachePoint":{"type":"default"}} content block
appended to the relevant arrays.
Changes:
- bedrockRequest.System widened from []map[string]string to
[]map[string]interface{} to accommodate mixed text/cachePoint blocks
- bedrockCachePoint: shared sentinel content block var
- bedrockModelSupportsCaching: gates on PriceCacheWrite > 0; strips
geo prefixes before catalog lookup; falls back to anthropic.claude-
prefix check for unknown models (cachePoint is silently ignored by
the API if unsupported)
- buildRequest: resolves model ID before caching check; injects
cachePoint into system array and calls bedrockTagLastUserCache
- bedrockTagLastUserCache: appends cachePoint to last user message
Nova models (PriceCacheWrite == 0) are excluded — they use Bedrock's
automatic caching and don't need explicit markers.
Tests: 8 new cases covering model detection, Claude vs Nova presence/
absence, multi-turn last-message targeting, no-system safety,
nil/empty panic safety, and JSON wire shape.
A long provider error (e.g. a Bedrock 400 with an embedded JSON body)
was appended as one line and clipped at the terminal edge. Wrap v.Err to
the build width via wrapLine (which hard-breaks unbreakable blobs), keep
the marker on the first row, and indent continuation rows so the whole
message is readable.
Newer Bedrock models (Anthropic Claude 4.x, DeepSeek) reject invocation
by their bare foundation-model ID with on-demand throughput, demanding a
cross-region inference-profile ID instead (HTTP 400). Rewrite such IDs
at request time by prepending the region-matched geo prefix
(us/eu/apac/us-gov), so selecting anthropic.claude-sonnet-4-5-... in a
us-east-1 setup invokes us.anthropic.claude-sonnet-4-5-...
Already-prefixed IDs, ARNs, and families that don't need a profile are
left untouched, preserving explicit choices and custom application
inference profiles.
On VS Code's xterm.js the transcript is taller than the viewport, so an
in-place clear (home + erase-to-end) only wipes the visible rows and the
scrolled-away part lingers in retained scrollback, stacking a duplicate
on the next full repaint.
- Clear() (Ctrl+L) now emits \x1b[3J under keepScrollback to actually
drop that scrollback, then homes and repaints. Accepts VS Code's
viewport-snap since the user explicitly asked for a clean screen.
- Overlay close (esc on a dialog, slash/file popup dismissal) now runs
the same Clear() so closing a picker purges the stale overlay rows
instead of leaving them in scrollback.
- Resize() does the same purge under keepScrollback; previously it
skipped the wipe and left a half-repainted old-width frame until the
user pressed Ctrl+L.
Other terminals keep their no-snap clear path.
- Shell escape: typing "!cmd" runs it via the bash tool's shell in the
session cwd, honoring the /jail sandbox. Output is parked below the
transcript as a styled terminal-log block until the next prompt or
/clear, so it never enters the model conversation. Shares busy state
with the agent: esc cancels it and no turn or other escape can start
while one is in flight.
- VS Code terminal: full repaints used \x1b[2J, which xterm.js scrolls
into scrollback and duplicates the frame. Clear in place via cursor
home + erase-to-end under keepScrollback; Clear()/Resize() no longer
eagerly wipe. Force a viewport-safe Invalidate on slash/file popup
open and close transitions there.
- Restore the live tool-call overlay behavior (keep in-flight boxes
visible until the tool_result reaches the transcript) and drop the
forced repaint at turn start.
- Document the shell escape in the README.
Map short/alternate provider names (bedrock -> amazon-bedrock, vertex,
gemini, azure, copilot, codex, ...) to their canonical ids in Resolve so
an alias is never treated as unknown and silently downgraded to
anthropic. Add a region-aware hint to Bedrock 403 responses on the
bearer route.
- User themes from $ZOT_HOME/themes/*.json with partial overrides
(colors, syntax, spinner) and dark/light fallback.
- /settings color-theme picker; selection persisted in config.json.
- Theme-only extensions: extension.json plus theme.json (or
themes/theme.json) load without spawning a subprocess.
- write-zot-themes built-in skill and docs/themes.md.
- README, extensions docs, and embedded docs index updated.
On the openai-codex (Responses API) route a tool result serialized to a
string-only function_call_output, dropping ImageBlock content, and the
agent loop's tool-image mirror only fired for provider "openai". So
images returned by read reached the TUI but never the model, which then
correctly reported it received no image content.
Extend the mirror to fire for "openai-codex" too (its client already
serializes user-message images as input_image, so the bytes arrive),
and have the codex tool-result serializer emit a short placeholder for
an image-only result instead of an empty output the API may reject.
Adds a test covering both behaviors.
Opus 4.7+ only support adaptive thinking: explicit thinking budgets
(thinking:{type:enabled,budget_tokens:N}) and non-default sampling
params return 400. The Anthropic client now sends thinking:{type:
adaptive} plus output_config.effort and omits temperature for these
models, while older models keep the budget-based path. Adaptive models
are detected via a new Model.AdaptiveThinking flag with an id-substring
fallback so the same family reached through an Anthropic-Messages proxy
is handled too.
For adaptive Anthropic models served over the OpenAI-compatible chat-
completions wire (openrouter, opencode, ...), reasoning_effort now maps
maximum -> xhigh instead of clamping to high, preserving the model's
full reasoning ceiling. Adds AnthropicAdaptiveEffort and
OpenAICompatAnthropicEffort with tests.
