Trebell Code 1.3.3

Trebell Code 1.3.3

Version 1.3.3 hardens Trebell Native as a real first-party coding harness and adds live-provider proof that Native is independent from the Codex app-server path.

Trebell Native real-model validation

• Added npm run test:vyce:native, a real-provider end-to-end Native harness gate.
• The live test wires attachAgentRelay directly to ProviderManager.turn; it never starts the Codex app-server.
• The real model must receive the Trebell Native system prompt and Native tool schemas, inspect source, edit a file, run terminal verification, produce a final answer, and leave output that an independent Node process verifies afterward.
• The v1.3.3 validation passed with VyceAi + deepseek-v4.1.
• Live instrumentation records model turns, provider request size, schema size, tool calls, arguments, and token usage without exposing credentials.

Native coding-agent prompt

• Added a first-party Trebell Native system prompt instead of relying on tool schemas and optional thread developer instructions alone.
• The prompt teaches evidence-driven coding, permission boundaries, verification discipline, untrusted tool-data handling, minimal coherent edits, and honest failure reporting.
• It tells the model to use exact paths/commands supplied by the user or tools instead of rediscovering them.
• It encourages batching independent read-only tool calls to reduce unnecessary model round trips.
• The current release candidate measures the system prompt at roughly 532 estimated tokens; prompt size is not the dominant Native cost.

Progressive repository tools

• Reduced the baseline Native repository tool surface to high-value search/read primitives plus stable trebell_repo/discover + trebell_repo/invoke.
• Advanced capabilities such as semantic rename/code actions, call hierarchy, Git history/blame, deterministic verification control, and durable repository knowledge are exposed only when specifically requested.
• Generic discovery queries such as “repository structure and project layout” no longer inflate the advanced schema set.
• Discovery returns capability metadata as data instead of mutating the provider-visible schema manifest.
• Native MCP follows the same stable-manifest idea with generic discovery/call/resource tools while policy still resolves the real target tool.
• On the measured coding path, baseline functions dropped from 31 to 11 and first-request tool-schema JSON dropped from about 14.3 KB to about 4.9 KB.

Native context efficiency

• Native UI turns now keep scoped repository instructions but replace large preloaded source excerpts with a compact untrusted repository seed map.
• On the Trebell repository audit task, the repository evidence preamble dropped from about 2,778 estimated tokens to about 340 tokens while preserving selected paths and orientation.
• Exact source is fetched on demand through Native repository/workspace tools instead of being replayed in every model/tool round trip.
• Byte-identical repeated source reads are replaced with a compact unchanged-observation marker.
• Large redacted tool output is stored outside hot model context behind bounded trebell_output/search and trebell_output/read handles, with a signal-aware preview that retains likely failure/assertion lines.

Tool-loop reliability

• Added one bounded recovery attempt for empty terminal model responses; a second empty completion now fails visibly.
• Normalized common safe argv mistakes such as args: "verify.mjs" and command: "node verify.mjs" without implicitly interpreting shell syntax.
• Normalized model-style workspace-root paths such as /src/app.js before both policy evaluation and execution while preserving traversal/absolute-path boundaries.
• Added schema validation for Trebell-owned tool arguments before policy evaluation/execution.
• Invalid or unexpected tool arguments fail closed instead of reaching the executor.
• Long-running server/watcher control lives behind lazy trebell_process tools instead of bloating the always-present one-shot terminal schema.
• Project-list background refreshes no longer erase unrelated user-action errors, so failed cross-environment activation remains visibly reported after Trebell safely rolls the environment back.
• Agent Browser recording now prefers WebM codecs, flushes a final MediaRecorder chunk before stop, and fails visibly on empty output. The packaged hidden-window smoke temporarily exposes the capture source to the compositor and now verifies non-empty encoded video.

Native inference telemetry and provider capabilities

• Every Native model step now records a stable inference id, prompt-section estimates, tool-schema/stable-prefix hashes, normalized provider usage, cache-read/write tokens, request/response bytes, latency, provider endpoint/wire API, and context-window utilization where known.
• Trebell preserves internal provenance so actual user text can be measured separately from Trebell application/untrusted working context without changing provider-visible serialization.
• Provider capability metadata distinguishes protocol compatibility from verified prompt caching, explicit cache control, stateful continuation, persistent connections, native compaction, and cache telemetry.
• A controlled Vyce repeated-prefix experiment observed 0 cached input tokens, so Trebell does not pretend the current Vyce Chat Completions route has a cache/stateful continuation feature it has not demonstrated.

Live efficiency result

The first measured real Native coding run consumed roughly 40.4k input tokens for the small validation fixture. The latest equivalent run completed successfully in 4 model turns / 11,997 input tokens / 373 output tokens, with no failed tool calls, one stable prefix, one stable tool-schema hash, and independent post-turn verification.

A new bench:vyce:native live benchmark also covers multi-file refactoring, failing-test repair, and large noisy command output in disposable repositories with independent verification and per-scenario token/latency/tool metrics. The final v1.3.3 benchmark passed all three scenarios:

• multi-file refactor: 5 turns / 16,349 input tokens;
• failing-test repair: 4 turns / 13,148 input tokens;
• 92 KB noisy-output repair: 7 turns / 37,526 input tokens with virtualized command output.

The three tasks totaled 16 model turns / 67,023 input tokens. Large-output work remains an optimization area: virtualization materially bounds retained context and preserves useful failure signals, but model/tool-loop decisions can still dominate total cost.

Release validation

• Explicit TREBELL_HOME values now isolate Electron userData too, preventing single-instance/prefs collisions in isolated release and desktop-browser validation.
• Unknown /api/* routes now return an explicit 404 instead of falling through to frontend assets; the retired mobile-device API is verified absent regardless of whether ui/dist already exists.
• 808 / 808 deterministic tests passed.
• 138 / 138 full visual/screenshot tests passed in offline provider-safe mode.
• 2 / 2 targeted Native UI E2E tests passed (compact context seed + thread-owned background process/runtime UI).
• The real Vyce Native coding gate passed with deepseek-v4.1, including independent post-turn verification.