Trebell Code 1.3.5

Trebell Code 1.3.5

Version 1.3.5 separates Trebell Native cleanly from external coding harnesses, expands direct provider support, and adds measured context/tool-loop optimizations without trading away coding quality or verification.

Native vs external harness architecture

• Trebell Native now owns its own agent loop, prompt, context, tools, persistence and direct provider transport instead of treating Codex as an inference bridge.
• Codex uses the real Codex app-server, native account/config/model catalog and native continuation identity. Selecting a Trebell Native provider no longer changes Codex inference.
• Claude Code, OpenCode and ACP-backed runtimes keep their native prompt/session semantics and receive only a small Trebell application-context addition.
• Cross-runtime thread ownership stays truthful: saved external-harness threads resume through the harness that owns them rather than being silently converted to another runtime.
• AgentRouter still receives its required Codex-compatible client fingerprint at the provider-transport layer; this is compatibility metadata, not Codex harness substitution.

Direct provider support

• Trebell Native has first-class direct adapters for official OpenAI, Anthropic and Gemini APIs in addition to AgentRouter, JustWorker, HCNSec and VyceAI.
• Official OpenAI uses the Responses API and preserves Trebell namespaced function-tool identity.
• Official OpenAI Responses requests now carry a deterministic prompt_cache_key derived from the stable model/instructions/tool/first-task prefix, so later turns in the same task keep the same cache-routing/accounting key while materially different task prefixes separate cleanly. OpenAI still performs prompt caching automatically where supported; Trebell does not claim a cache hit unless the provider reports one.
• Anthropic uses the Messages API with Anthropic-native authentication semantics.
• Gemini currently uses Google's OpenAI-compatible endpoint; native cached-content behavior remains intentionally unclaimed until measured.
• Provider telemetry records request/response bytes, latency, normalized usage, cached-input fields when exposed, endpoint and wire API without persisting secrets.

Measured Native efficiency improvements

• ACP runtime context is injected once per external ACP session instead of being repeated on every prompt.
• The obsolete provider compatibility bridge and Responses-to-Chat bridge path were removed from normal architecture.
• Conventional /workspace/... model paths are normalized to the active Trebell workspace, eliminating repeated failed tool calls on Windows. A failure-repair benchmark dropped from 31,920 input tokens / 10 model turns / 17 tool calls during the regression to 8,826 input / 4 turns / 5 calls after the path fix while still independently passing verification.
• Repeated virtualized tool output is cooled across later user turns to a compact receipt plus searchable handle and high-signal failure lines. On the 92 KB noisy-output benchmark this measured 23,063 input tokens / 7 turns / 6 calls, versus 29,696 input immediately before the change and 37,526 input in the preserved v1.3.3 baseline.
• Native now performs one bounded recovery when the user explicitly names an exact exposed tool and the model tries to answer without calling it. Negated, vague and stale-context tool mentions do not trigger this behavior.
• A tool-order experiment showed the requested tool was called with both normal and reordered schemas, so Trebell keeps schema ordering stable rather than shuffling tools and breaking stable schema hashes.
• A proposed reduction of the baseline repository manifest to only search_code was rejected after it caused three focused regressions. search_symbols, search_files and read_source remain directly available until a lower-cost migration is proven without behavior loss.

Current prompt/context telemetry

One real Vyce Native coding smoke on the release candidate measured:

• 4 model turns and 8,082 provider input tokens total;
• 582 estimated system-prompt tokens;
• 11 baseline tool functions and 1,280 estimated tool-schema tokens;
• one stable prefix hash and one stable tool-schema hash across all four requests;
• 0 provider-reported cached input tokens;
• real workspace edit plus terminal verification;
• independent verification passed;
• no Codex app-server in the Native request path.

These are measurements, not fixed guarantees; model behavior and provider routing remain stochastic.

Live/runtime validation

• Real Codex live smoke passed with the Codex-owned model catalog and gpt-6-luna, while Trebell Native remained configured for a different provider.
• OpenCode 1.18.32 passed a real tool/coding smoke with opencode/muse-spark-1.3-contributor-free.
• Antigravity ACP 1.2.1 passed a real tool/coding smoke with gemini-3.8-flash-high.
• Grok reached its real runtime but was blocked by an upstream Rate limited response during the latest rerun.
• Claude Code remained installed but unauthenticated; Cursor's launcher remained unavailable on this machine, so neither is claimed as a live pass.
• Real Trebell Native provider smoke passed for JustWorker, HCNSec and VyceAI. AgentRouter reached the real route but remained blocked by upstream HTTP 402 budget-pool exhaustion.
• Official OpenAI, Anthropic and Gemini adapters are contract-tested; no live success is claimed because those API keys were not configured in the test environment.

Packaged Windows validation

• A real unpacked Electron build passed the installed Codex relay and command-execution smoke.
• Packaged Codex reported its bundled runtime as authenticated through ChatGPT and kept a Codex-owned model catalog even while the Trebell Native provider was switched to VyceAI.
• The packaged desktop/browser smoke passed browser navigation, typing/clicking, cookies, screenshots, video recording, desktop capture, SnapShot persistence, zoom controls, notifications and background/startup toggling.
• A real packaged Trebell Native deepseek-v4.1 run called packaged computer screenshot plus browser open/snapshot/type/click capabilities and created/verified a filesystem proof file.
• Windows packaging now builds unpacked and NSIS artifacts in unique isolated temp directories, uses the already-installed Electron distribution to avoid archive-extraction races, retries once in a fresh isolated directory for observed Windows materialization races, and only copies completed installer/update artifacts into desktop-dist.
• The second packaged smoke now waits for the first Electron process tree to fully terminate and uses fresh GUI/app-server/CDP/fixture ports before relaunching, avoiding a Windows single-instance/socket teardown race between the desktop smoke and the real-model Native smoke.

Automated regression baseline

• npm test: 829 passed, 0 failed on the consolidated pre-release tree.
• Latest full Playwright UI suite in headless Google Chrome: 193 passed, 0 failed.
• Production UI build succeeds.

The release pipeline still performs its own deterministic tests and packaged smoke before publishing.