Trebell Code 1.3.6

Trebell Code 1.3.6

Version 1.3.6 is a focused Native reliability and efficiency follow-up to v1.3.5. It keeps the v1.3.5 architecture intact while removing several avoidable recovery turns caused by conventional model path/tool-call mistakes.

Native recovery without guessing

• Relative workspace/... paths now receive the same bounded workspace-root fallback as /workspace/... only when the literal path does not exist. A real workspace directory still wins, so Trebell does not silently reinterpret valid user paths.
• When trebell_repo/read_source hits a known "file is not indexed" boundary and trebell_workspace/read_file is already exposed, Native can perform that bounded read internally rather than spending another model turn asking for the same file through a second tool. The model still sees the original repository tool lifecycle plus explicit fallback provenance.
• Provider-flattened Trebell tool aliases are repaired only when they identify exactly one exposed tool. Ambiguous names are left untouched instead of being fuzzy-guessed.
• If a later same-turn trebell_terminal/run call loses only its executable, Native may restore the command only when exactly one earlier successful terminal call in that same turn has identical args and cwd. Failed executions, timed-out commands, signal-killed runs, empty-args calls, and ambiguous command matches are never used as repair templates.
• Recipe tool allowlists now also filter the schemas sent to the model. Disallowed tools remain blocked by the execution gateway even if a provider hallucinates a hidden call, so this reduces prompt cost without weakening the policy boundary.
• Large virtualized tool output keeps its persistent retrieval handle and signal-aware failure lines while using a smaller hot preview, reducing repeated noisy context without discarding the full stored output.

Measured behavior

• The earlier /workspace/... normalization had already reduced one failure-repair regression from 31,920 input tokens / 10 model turns / 17 tool calls to 8,826 input / 4 turns / 5 calls while independently passing verification.
• On the current v1.3.6 candidate, the same failure-repair benchmark again completed successfully in 4 model turns / 6 tool calls / 9,191 provider input tokens, with 0 failed tool calls and independent verification passing. Model/provider behavior remains stochastic; this is a measured run, not a fixed guarantee.
• A repository-tool-description compression experiment was rejected: although it reduced the provider-visible schema from 5,117 to 4,793 JSON characters, a reproduced three-scenario run still used 44,968 input tokens / 17 model turns / 21 tool calls, worse than the cleaner 35,397 input / 14 turns / 17 calls baseline. The fuller descriptions remain in place.
• The live Native benchmark now exposes per-turn input/output/cache usage and request-cost evidence so future optimizations can be judged from the expensive turn instead of only aggregate totals.
• On the same 92 KB noisy-output repair task, the earlier full-schema / 6,000-character-preview run used 18,745 provider input tokens / 6 model turns / 4 tool calls. The current combined allowlist-filtering + bounded-preview build used 15,345 input / 6 turns / 4 calls with independent verification passing, about 18% less provider input in that measured pair. On the first recipe-restricted request specifically, the visible tool surface fell from 11 functions / 5,117 schema characters / ~1,280 estimated schema tokens to 1 function / 772 characters / ~193 estimated schema tokens. Both runs reported zero cached input tokens; provider/model behavior is stochastic, so these are measurements rather than a fixed guarantee.

Validation status before release

• Focused Native loop/session/path coverage passed, including negative cases that prove recovery does not cross permission/tool exposure boundaries or reuse ambiguous/failed terminal evidence.
• Latest full deterministic suite on the candidate: 838 passed, 0 failed.
• v1.3.5 remains untouched; v1.3.6 is a new release and will receive its own installer, updater metadata, blockmap, release metadata, and Git tag.

The full deterministic suite was completed before the final release-hardening commits; the corrected packaging run then reran package-specific Codex, desktop/browser, real-model Native, NSIS, and updater gates on the published commit.

Final release hardening

• Explicit required-tool recovery now narrows the provider-visible schema to the one required tool for the bounded retry, then restores the normal tool surface afterward.
• Packaged Windows PowerShell failures now decode CLIXML envelopes into readable errors, and clean release bootstrap installs dependencies before checking for Electron.
• The final corrected packaging run used -SkipTests because the deterministic suite had already completed successfully. Focused Native loop/session coverage passed 42/42 and release-critical PowerShell/updater coverage passed 9/9; packaged Codex, desktop/browser/SnapShot, real-model deepseek-v4.1, NSIS, and updater metadata/blockmap gates were run again on the exact published commit.