fix(codex): improve local session import compatibility

[Himehane]

fix(codex): improve local session import compatibility

• fix duplicated user messages during transcript import
• fix duplicated assistant messages during transcript import
• normalize imported messages before deduplication
• improve compatibility with recent Codex transcript format
• support session titles from session_index.jsonl
• preserve fallback behavior for older Codex versions
• add regression tests for transcript parsing and title resolution

[github-actions]

Findings

• [Major] Synthetic context filter can drop real imported user prompts — hub/src/web/routes/codexDesktop.ts:296
• [Major] Adjacent duplicate collapse is not limited to mirrored transcript sources — hub/src/web/routes/codexDesktop.ts:823

Summary

• Review mode: initial
• Found two import-parser data-loss risks introduced by the diff. Both affect transcript fidelity during direct Codex import.

Testing

• Not run (automation); attempted bun test hub/src/web/routes/codexDesktop.test.ts, but bun is not installed in this runner.

HAPI Bot

[github-actions]

[MAJOR] This now drops any response_item user message that merely contains <environment_context> anywhere in the prompt. response_item is also where real user prompts are imported, so a user asking about or pasting a transcript/XML block with that tag is silently omitted from the imported HAPI session. Keep the filter to standalone injected blocks instead of arbitrary substrings.

Suggested fix:

function shouldIgnoreInjectedResponseUserMessage(text: string): boolean {
    const normalized = text.trim()
    const lower = normalized.toLowerCase()
    const isAgentInstructions = lower.startsWith('# agents.md instructions')
    const isEnvironmentContext = lower.startsWith('<environment_context>')
        && lower.endsWith('</environment_context>')
    return isAgentInstructions || isEnvironmentContext
}

[github-actions]

[MAJOR] This duplicate check collapses any adjacent imported user/assistant messages with the same normalized text, even when both came from the same transcript source. That can remove legitimate repeated messages; the mirror bug this PR fixes is specifically an event_msg/response_item duplication, so the drop should require different sources.

Suggested fix:

function isMirroredAdjacentDuplicate(
    previous: CodexImportedMessageEntry | undefined,
    next: CodexImportedMessageEntry
): boolean {
    return Boolean(
        previous
        && previous.source !== next.source
        && isAdjacentDuplicateImportedMessage(previous.message, next.message)
    )
}

// Use isMirroredAdjacentDuplicate(previous, entry) in both adjacent-dedupe branches.

[github-actions]

Findings

• No issues found.

Questions

• None.

Summary

• Review mode: follow-up after new commits
• Previous HAPI Bot findings from 0e2c7bee2988a57ff640944ce1e6a583e687c8f8 were rechecked against current head 899c16c223334d3d5f1a5e6b29816b29d757d0eb; no current diff findings above 80% confidence.
• Residual risk: local test automation could not be run in this runner.

Testing

• Not run (automation); bun --version failed with bun: command not found.
• git diff --check passed.

HAPI Bot

[Himehane]

Verified locally.

All tests passed successfully.

• 134 test files passed
• 1131 tests passed

Additionally verified the Codex import behavior manually:

• Duplicate user messages are no longer imported.
• Duplicate assistant messages are no longer imported.
• Session titles are correctly loaded from session_index.jsonl.