Oh My Pi v17.2.13: Bedrock role_arn credential resolution, Cursor double-execute fix, GLM-5.2 support

@oh-my-pi/pi-agent-core

• Cursor re-execution: settled tools no longer re-execute when an owned dialect projector rebuilds toolCall blocks; snapshotAssistantContentBlock now copies kCursorExecResolved.

@oh-my-pi/pi-ai

• User-Agent: standardized on omp/<version> via shared utility.
• Bedrock/Cursor StreamOptions.headers: fixed — caller-supplied tracing/attribution headers were silently dropped.
• Antigravity Flash: hang fixed — after successful response headers with no SSE events, provider now cancels stalled body and fails over after 60 seconds (longer allowance for Pro reasoning starts).
• Cursor exec-bridge: ArkType validation fixed — undefined optional fields (cwd, case, skip, timeout) are now dropped instead of written as value || undefined.
• Cursor tools.format dialect: wrapInbandToolStream now copies kCursorExecResolved so agent-loop no longer double-executes settled bash/grep/todo calls.
• Codex Responses Lite: fixed — requests for opaque codenames (e.g., Daybreak) now include reasoning.context: "all_turns" and no longer fail with HTTP 400.
• Cursor usage reporting: adapted for current Pro/Pro+/Ultra /api/usage-summary payloads using individualUsage.plan instead of the older individualUsage.overall bucket.
• Google passive callers: empty or thinking-only STOP responses now accepted as successful silence instead of exhausting retry budget.
• AWS credentials: role_arn profiles now resolve via STS AssumeRole/AssumeRoleWithWebIdentity, honouring role_session_name/duration_seconds/external_id; Bedrock detected on EKS/IRSA and multi-account setups.
• Bedrock EC2 detection: Nitro DMI markers (board_asset_tag instance ids, Amazon EC2 vendor fields) now recognized in addition to Xen ec2 UUID prefix.
• DeepSeek Responses thinking: after prewalk hand-off plus mid-run compaction, replayed assistant turns now include a reasoning_text item when the target requires reasoning replay in thinking mode.

@oh-my-pi/pi-catalog

• User-Agent: standardized on omp/<version>.
• Muse Spark 1.2: meta/muse-spark-1.2 and muse-spark-1.2-contributor now correctly listed as image-capable (input: ["text", "image"]) with same Responses reasoning metadata as 1.1; contributor uses its discounted pricing.
• GLM-5.2 thinking: levels fixed across Baseten, CoreWeave, HuggingFace, and other uppercase-ID resellers — no longer using generic xhigh effort ladder. Added Baseten zai-org/GLM-5.2-Fast and Fireworks glm-5.2-fast as reasoning models.

@oh-my-pi/pi-coding-agent

• SearXNG safesearch: new searxng.safesearch setting added.
• omp update: now honours omp.dist field in npm manifest. Major-version bumps without one are treated as binary-only; bun/npm-managed installs migrate to standalone GitHub release binary. Windows shim installs (omp.cmd/omp.ps1) taken over by omp.exe.
• Cloudflare AI Gateway: routing support added for Gemini search.
• Exa MCP: search provider added.
• TinyFish search: domain inclusion/exclusion filtering and URL deduplication added.
• /vibe mode: pre-vibe toolset now recorded on mode_change entry and restored properly on exit — bash, edit, write, grep, glob, task, hub no longer silently unavailable.
• Composer drafts: extension-filtered pasted image payloads and source links preserved when /goal, /plan, or /vibe submits.
• Agent Hub: lineage registration timestamps now display in user's local timezone (fix for UTC-only display, raw note truncated at source).