Drop temperature for models that removed sampling params (400 otherwise).
Preserve signed thinking blocks across tool rounds — Anthropic rejects a
tool_use whose preceding thinking block was dropped. Raise default
max_tokens to 32k for thinking-by-default models and request summarized
display so the reasoning pane is fed.
Revises the prior fix: the model's reply is already complete in
final_content before the paced stream loop, so truncating to
needlessly dropped the unstreamed tail. On SSE receiver-drop, stop
streaming but commit the FULL reply (interrupt flag skips the pointless
N+1 and marks the turn); the surface recovers the whole reply on resume.
turn.rs streaming loop returned early on tx.send failure (receiver gone)
before the session commit, orphaning the turn: the assistant's partial
reply never landed, so the next turn's history snapshot read the user's
follow-up with no preceding reply -> a user->user gap. The model then
answered as if from nowhere (the 'whatever it was...' non-sequitur),
model-independent and intermittent. Commit the partial like an interrupt
instead, the way the cancel branch already does, so the thread links.
ToolCard styled the status glyph with a hardcoded byte offset 0..3, but
✓/✗/⟳ are 3-byte chars sitting at bytes 2..5 of " {glyph}". The split
landed mid-char and panicked at render in tuie's style slicing. Compute
the glyph region as 2 + glyph.len_utf8() instead.
Then took cargo clippy -- -D warnings from 312 failures to clean:
- scoped #![allow(dead_code)] on WIP scaffolding (federation, sensorium,
gitea_memory, model_router, session, subagent…); gate stays live on
active code so new orphans still fail
- scoped #![allow(deprecated)] on the legacy ratatui render path, marked
pending removal at tuie parity — no migration on code we're deleting
- declare the gui feature (forwards to tuie/gui) — the cfg was real intent
- real fixes: duplicate SaveAndGoBack arm + dead Err arm, base64::encode,
4 unused imports, dead assignment, private-type leak, dedup'd if/else
branches, manual clamp/strip, &PathBuf→&Path, collapsible matches
agents: wire the rest of the detail widgets into the section children so
PRIMARY badge etc actually render; drop the HARDCODED debug text.
settings: test needs a provider in the map before navigating to Providers
cat.
clippy: pin CI to rust 1.94 per Cargo.toml rust-version; also apply what
cargo clippy --fix could auto-fix for 1.96 compat.
Vanguard gets his own inference pipe. z.ai GLM Coding Plan, /paas/v4.
BifrostClient learned to leave versioned URLs alone. ProviderRegistry
dispatches per agent; every engine asks who it's talking to before
dialing. Annie and Hal none the wiser.
Introduce an `LlmProvider` trait (the engine<->LLM seam, sibling to the
`Backend` harness<->engine trait) so inference can route to providers
beyond the Bifrost gateway. Two impls behind it:
- `BifrostClient` - existing OpenAI-compatible gateway (default).
- `OpenAiOAuthProvider` - rides the Codex CLI's ChatGPT login
(`~/.codex/auth.json`) and drives `chatgpt.com/backend-api/codex/responses`
(Responses API) with no API key. Self-refreshes the token (single-flight,
write-back, CLI re-read fallback) and translates the engine's OpenAI-chat
request to/from the Responses API + SSE accumulation.
Selected via `[bifrost] provider` ("bifrost" | "openai-oauth"). The engine
keeps speaking the existing ChatCompletionRequest/CompletionResult/
InferenceStrain currency, so all six inference call-sites are unchanged -
only the field type flips to `Arc<dyn LlmProvider>`.
Model ids are translated at the provider boundary (oauth/catalog.rs::resolve):
Bifrost-namespaced ids (`openai/...`, `-precision`) map onto served ChatGPT
models; `-fast` -> priority service tier.
Verified live to the wire level: builds+links, server boots in oauth mode
(reads the Codex token), and chatgpt.com accepts the request (auth, endpoint,
headers, payload all valid). The SSE->CompletionResult accumulation is NOT yet
verified against a successful completion (blocked by a subscription usage limit
at test time) - needs one live turn to confirm end-to-end.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- Itinerary tool (set/advance/describe/clear) with YAML persistence
and todo-linked enrichment, rendered as TUI header strip
- Mid-turn subconscious checkpoint: every N tool rounds, Aster assesses
progress; HALT verdict breaks the loop with full correction pass
- Nickname tool: agent learns human's name via set/get/clear, stored
in system/human.md frontmatter, TUI uses it in chat bubble labels
- Removed hardcoded human.md write from memory::init (wizard owns it)
- Prompt nudge for itinerary when live todos exist without a route
