fix(harness): correct codex & openai reasoning stream envelopes

[declan-scale]

What

Fixes two reasoning-stream issues found while reviewing the unified-harness release (#424). Both are confined to the streaming envelope (the live StreamTaskMessage* sequence); neither changes the persisted message type, but #1 produces a duplicate persisted message.

1. Codex: duplicate / orphaned reasoning message (🔴 user-visible)

_codex_sync.py emitted Start(ReasoningContent, active) then Full(ReasoningContent, static) at the open index, with no Done. auto_send handles Full by opening its own throwaway streaming context (it ignores event.index), so the original Start context was never closed until end-of-turn teardown — persisting a second, near-empty reasoning message with a later timestamp (duplicate + out-of-order).

Fix: mirror the working _claude_code_sync.py pattern — stream the final reasoning as ReasoningSummaryDelta + ReasoningContentDelta on the open index, then close with Done. The open context accumulates the final ReasoningContent and closes cleanly as one message. The no-started case opens the Start lazily and closes it the same way; the empty-reasoning case still closes with a bare Done.

2. OpenAI: reasoning Start content type regressed ReasoningContent → TextContent (🟠)

convert_openai_to_agentex_events opened reasoning summary/content messages with a TextContent Start. On the migrated auto_send/Temporal path this regressed the pre-migration behavior (which started reasoning with ReasoningContent), so any consumer that branches on the start event's content type to show a "thinking" indicator now sees plain text. The persisted content is rebuilt from the reasoning deltas regardless, so only the live envelope was affected.

Fix: emit ReasoningContent(style="active") for both reasoning Starts. This aligns the converter with the codex/claude_code taps, the (already-corrected) langgraph-sync converter, and what the OpenAI conformance suite already treats as canonical.

Testing

• Rewrote the codex reasoning tests to assert the Start + summary/content deltas + Done sequence and that no Full(ReasoningContent) is emitted; added an empty-block-closes-with-Done case.
• tests/lib/adk/test_codex_sync.py, tests/lib/adk/providers/test_openai_turn.py, tests/lib/adk/test_codex_turn.py, tests/lib/adk/test_claude_code_sync.py, tests/lib/adk/providers/, and the full tests/lib/core/harness/ (incl. conformance) suites pass (203 + 87 green across runs).
• ruff format + ruff check clean on all changed files.

🤖 Generated with Claude Code

Greptile Summary

This PR fixes two streaming envelope bugs in the reasoning message pipeline: a Codex duplicate-message issue caused by mixing Start+Full (which leaves the open context dangling via auto_send's index-ignoring behaviour), and an OpenAI regression where reasoning Start events were typed as TextContent instead of ReasoningContent.

• Codex (_codex_sync.py): item.completed now emits ReasoningSummaryDelta + ReasoningContentDelta + Done on the already-open index, mirroring the _claude_code_sync.py pattern; a lazy Start is opened when no item.started preceded it, and an empty block still closes cleanly with a bare Done.
• OpenAI (sync_provider.py): Both reasoning_summary and reasoning_content Start events now carry ReasoningContent(style="active") so downstream consumers that branch on the start-event type correctly render a thinking indicator.
• Tests (test_codex_sync.py): Existing reasoning tests are rewritten to assert the full Start → deltas → Done sequence and explicitly assert no Full(ReasoningContent) is emitted; a new test covers the empty-block case.

Confidence Score: 4/5

Safe to merge — both fixes are well-scoped to the streaming envelope layer and don't touch persisted message types; the production logic is correct.

The Codex and OpenAI changes are straightforward and well-explained. The lazy-open path in _codex_sync.py correctly opens a Start, emits both a ReasoningSummaryDelta and a ReasoningContentDelta, then closes with Done — but the corresponding test (test_reasoning_no_started_opens_and_closes_one_message) only asserts on the content delta, leaving the summary delta unverified on that path.

tests/lib/adk/test_codex_sync.py — the lazy-open reasoning test is missing the ReasoningSummaryDelta assertion that the normal-path test carries.

Important Files Changed

Filename	Overview
src/agentex/lib/adk/_modules/_codex_sync.py	Replaces Full(ReasoningContent) emission with Start+SummaryDelta+ContentDelta+Done sequence; adds lazy Start for no-started case and imports ReasoningSummaryDelta.
src/agentex/lib/adk/providers/_modules/sync_provider.py	Fixes reasoning Start events to use ReasoningContent instead of TextContent for both reasoning_summary and reasoning_content paths; imports ReasoningContent.
tests/lib/adk/test_codex_sync.py	Rewrites reasoning tests to assert Start+deltas+Done pattern and adds empty-block test; lazy-open path missing ReasoningSummaryDelta assertion.

Sequence Diagram

%%{init: {'theme': 'neutral'}}%%
sequenceDiagram
    participant Codex as Codex API
    participant Proc as _CodexStreamProcessor
    participant Consumer as auto_send / Consumer

    Note over Proc,Consumer: BEFORE (broken)
    Codex->>Proc: item.started (reasoning)
    Proc->>Consumer: "StreamTaskMessageStart(ReasoningContent, active) @ idx"
    Codex->>Proc: "item.completed (reasoning, text=...)"
    Proc->>Consumer: "StreamTaskMessageFull(ReasoningContent, static) @ idx"
    Note over Consumer: auto_send ignores idx on Full — Start @ idx stays dangling until teardown — duplicate message persisted

    Note over Proc,Consumer: AFTER (fixed)
    Codex->>Proc: item.started (reasoning)
    Proc->>Consumer: "StreamTaskMessageStart(ReasoningContent, active) @ idx"
    Codex->>Proc: "item.completed (reasoning, text=...)"
    Proc->>Consumer: "StreamTaskMessageDelta(ReasoningSummaryDelta) @ idx"
    Proc->>Consumer: "StreamTaskMessageDelta(ReasoningContentDelta) @ idx"
    Proc->>Consumer: "StreamTaskMessageDone @ idx"
    Note over Consumer: Open context accumulates deltas and closes cleanly — one message persisted

    Note over Proc,Consumer: OpenAI fix (sync_provider.py)
    Proc->>Consumer: "StreamTaskMessageStart(ReasoningContent, was TextContent) @ idx"
    Note over Consumer: Consumers see correct type for thinking-indicator branch

Loading

%%{init: {'theme': 'base', 'themeVariables': {"darkMode": true, "background": "#0d1117", "primaryColor": "#21262d", "primaryTextColor": "#e6edf3", "primaryBorderColor": "#8b949e", "lineColor": "#8b949e", "textColor": "#e6edf3", "edgeLabelBackground": "#161b22", "actorBkg": "#21262d", "actorBorder": "#8b949e", "actorTextColor": "#e6edf3", "actorLineColor": "#8b949e", "signalColor": "#8b949e", "signalTextColor": "#e6edf3", "noteBkgColor": "#373320", "noteBorderColor": "#d4a72c", "noteTextColor": "#f0e6c0", "labelBoxBkgColor": "#21262d", "labelBoxBorderColor": "#8b949e", "labelTextColor": "#e6edf3", "loopTextColor": "#e6edf3", "activationBkgColor": "#30363d", "activationBorderColor": "#8b949e"}}}%%
sequenceDiagram
    participant Codex as Codex API
    participant Proc as _CodexStreamProcessor
    participant Consumer as auto_send / Consumer

    Note over Proc,Consumer: BEFORE (broken)
    Codex->>Proc: item.started (reasoning)
    Proc->>Consumer: "StreamTaskMessageStart(ReasoningContent, active) @ idx"
    Codex->>Proc: "item.completed (reasoning, text=...)"
    Proc->>Consumer: "StreamTaskMessageFull(ReasoningContent, static) @ idx"
    Note over Consumer: auto_send ignores idx on Full — Start @ idx stays dangling until teardown — duplicate message persisted

    Note over Proc,Consumer: AFTER (fixed)
    Codex->>Proc: item.started (reasoning)
    Proc->>Consumer: "StreamTaskMessageStart(ReasoningContent, active) @ idx"
    Codex->>Proc: "item.completed (reasoning, text=...)"
    Proc->>Consumer: "StreamTaskMessageDelta(ReasoningSummaryDelta) @ idx"
    Proc->>Consumer: "StreamTaskMessageDelta(ReasoningContentDelta) @ idx"
    Proc->>Consumer: "StreamTaskMessageDone @ idx"
    Note over Consumer: Open context accumulates deltas and closes cleanly — one message persisted

    Note over Proc,Consumer: OpenAI fix (sync_provider.py)
    Proc->>Consumer: "StreamTaskMessageStart(ReasoningContent, was TextContent) @ idx"
    Note over Consumer: Consumers see correct type for thinking-indicator branch

Loading

Prompt To Fix All With AI

Fix the following 1 code review issue. Work through them one at a time, proposing concise fixes.

---

### Issue 1 of 1
tests/lib/adk/test_codex_sync.py:441-449
**Lazy-open path missing `ReasoningSummaryDelta` assertion**

The production code for the `idx is None` case always emits a `ReasoningSummaryDelta` before the `ReasoningContentDelta` (lines 407–416 of `_codex_sync.py`). This test only asserts on `content_deltas` and ignores `summary_deltas`, so a future regression that accidentally drops the summary delta on this path would go undetected. Adding the same `summary_deltas` assertions that `test_reasoning_start_deltas_done` carries would close the gap.

_{Reviews (1): Last reviewed commit: "fix(tests): narrow reasoning delta types..." | Re-trigger Greptile}

Greptile also left 1 inline comment on this PR.

[greptile-apps]

Lazy-open path missing ReasoningSummaryDelta assertion

The production code for the idx is None case always emits a ReasoningSummaryDelta before the ReasoningContentDelta (lines 407–416 of _codex_sync.py). This test only asserts on content_deltas and ignores summary_deltas, so a future regression that accidentally drops the summary delta on this path would go undetected. Adding the same summary_deltas assertions that test_reasoning_start_deltas_done carries would close the gap.

Prompt To Fix With AI

This is a comment left during a code review.
Path: tests/lib/adk/test_codex_sync.py
Line: 441-449

Comment:
**Lazy-open path missing `ReasoningSummaryDelta` assertion**

The production code for the `idx is None` case always emits a `ReasoningSummaryDelta` before the `ReasoningContentDelta` (lines 407–416 of `_codex_sync.py`). This test only asserts on `content_deltas` and ignores `summary_deltas`, so a future regression that accidentally drops the summary delta on this path would go undetected. Adding the same `summary_deltas` assertions that `test_reasoning_start_deltas_done` carries would close the gap.

How can I resolve this? If you propose a fix, please make it concise.

Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!