v0.1.0-alpha.23 — Agentic tool-use rescue paths + per-attempt timeout (additive)

[baabakk]

Released 2026-06-24. Install: pnpm add @llm-ports/core@alpha @llm-ports/adapter-openai@alpha

Four additions targeting the agentic tool-use failure modes ADW's 2026-06-19 multi-team build-loop diagnostic surfaced. All additive, no breaking changes.

1. Harmony tool-call extraction (adapter-openai)

When tool_calls is empty AND reasoning_content contains a parseable harmony tool call, the call is hoisted into the executable path. Zero extra LLM calls. Closes the DeepInfra gpt-oss harmony tool-use gap that alpha.22 left open.

2. Zero-tool-call corrective rescue (adapter-openai)

When the model returns prose with empty tool_calls and the request had a tools array, the adapter retries once with a corrective system message: "Your previous response did not include a tool call. Tools are available — call them via the standard tool_calls array. Do not describe what you would do; do it."

Single-shot retry. Five discriminators prevent over-firing (no tools, populated tool_calls, empty content, populated reasoning_content, tool-result message in conversation history). Closes the mimo-parasail prose case where the model emitted 69 tokens of "I would do this..." and ADW promoted empty stubs to main.

3. Telemetry tags (core)

Two new RetryReason values: "harmony-tool-call-extracted" and "zero-tool-call-prose-retry". Filter the existing adapter onRetry hook on these to distinguish "rescued" from "clean zero-output (failover candidate)."

4. Per-attempt timeout (core)

const registry = createRegistryFromEnv({
  // ...
  perAttemptTimeoutMs: 30000,
});

Each provider attempt gets a fresh timer; on timeout, the abort propagates → adapter throws → Registry walks to next provider. Per-attempt, not chain-wide. Composes with user-supplied signal. Closes the ADW production wedge from 2026-06-19T15:40 UTC where mimo grinded silently for 3+ minutes with no escape.

What this does NOT fix

The Case B "under-production" pattern (model makes some tool calls, then stops with the planned manifest incomplete) is NOT addressed at the adapter layer. The adapter sees a clean multi-call completion; only the orchestration knows the manifest is incomplete. ADW (and similar agentic orchestrators) should add a "planned ≠ written" guard at the workflow layer — explicitly out of scope per the design discussion.

Backwards compatibility

All four additions are additive. Two existing test files updated to use mockResolvedValue instead of mockResolvedValueOnce so ASK 2's rescue retry has a target. Test intent preserved.

Tests

• 13 harmony extraction tests
• 8 zero-tool-call rescue tests
• 8 per-attempt timeout tests
• 792 total (was 763 in alpha.22; +29, 0 regressions)

Coming next (alpha.24 plan)

• Catalog architectural redesign — behavioral fingerprinting opt-in + catalog freeze policy. The static catalog gets formally demoted to a perf optimization (not load-bearing for correctness). Addresses the architectural pushback that motivated alpha.22.
• Comprehensive reasoning-model coverage across Parasail, DeepInfra, SambaNova (research artifact + catalog additions)

Full release notes | Empirical source: ADW Dev Log b1eeee2