Klatch's automated intel sweep surfaced Opus 4.7's general availability and two approaching model deprecation deadlines — one already past — that both projects need to act on. On the architectural side, Klatch's MCP server reaches its Phase 5c decision point (write-path), and PM has shipped a session log maintenance hook pattern worth importing into Klatch's agent discipline.
Key Insights
1. Opus 4.7 GA — evaluate upgrade in both projects
From: Klatch intel sweep (automated, 2026-04-20) Relevant to: Klatch, Piper Morgan
Anthropic released Claude Opus 4.7 on April 16. Key improvements over Opus 4.6: substantially stronger software engineering and long-running coding tasks; high-resolution vision (3.75MP vs. 1.15MP); and a new task budgets feature where the model receives a running token countdown and self-paces its tool use. Pricing is unchanged ($5/$25 per M tokens in/out). SDK version @anthropic-ai/sdk is now at 0.90.0 — Klatch is on 0.86.1, four minor versions behind.
The task budgets feature is directly relevant to both projects: Klatch's Step 10 export packaging involves multi-step tool orchestration that could benefit from budget awareness; PM's Excellence Flywheel tracking could use it to surface cost-per-session trends. Neither project should rush the upgrade before running their full test suites, but the evaluation should happen this sprint.
Suggested action: Klatch: bump DEFAULT_MODEL evaluation and @anthropic-ai/sdk to ^0.90.0 in next dep maintenance. PM: check for Opus 4.6-specific patterns (budget_tokens, which is deprecated on 4.6) before upgrading.
2. Two model deprecation deadlines — one already past (urgent)
From: Klatch intel sweep (automated, 2026-04-20) Relevant to: Klatch, Piper Morgan
Two concurrent deprecation risks:
Haiku 3 (deadline: April 19 — yesterday). A prior sweep identified a MODEL_ALIASES bug in Klatch: the key claude-haiku-3-20250307 had two typos vs. the real model ID claude-3-haiku-20240307. The fix was assigned but its ship status is unconfirmed. If it didn't land, any channel using the Haiku 3 model ID is broken as of yesterday and needs an emergency DB migration.
Sonnet 4 + original Opus 4 (deadline: June 15 — 57 days). claude-sonnet-4-20250514 and claude-opus-4-20250514 retire June 15. Migration targets are Sonnet 4.6 and Opus 4.7. Both projects may have these model IDs hardcoded in DB records, config, or channel exports.
Suggested action: Klatch: confirm Haiku 3 alias fix shipped before today's sessions; if not, treat as P0. Both projects: run a grep/DB audit for claude-sonnet-4-20250514 and claude-opus-4-20250514 before end of sprint. Add both IDs to MODEL_ALIASES as a safety net.
3. Klatch Phase 5c decision point — write-path with cross-producer implications
From: Klatch docs/logs/2026-04-18-1256-argus-opus-log.md, docs/plans/STEP-10-PHASE-5-MCP-SERVER.md
Relevant to: Piper Morgan (cross-producer coordination)
Phase 5a (read-only resources) and 5b (tools surface: list_channels, get_context_package, get_manifest) are complete — 1069 tests, zero failures. Argus has given Daedalus a clear signal to pause for the Phase 5c decision.
Phase 5c introduces the first write-path: reflect(channel_id, note) — a micro-reflection that a consumer agent writes back after using a context package. The plan document flags this explicitly: "the data-integrity and concurrency questions deserve their own evaluation." More critically, write-path coordination has provenance and trust implications that read-paths don't — the same note in the Phase 5 plan. PM is the sibling producer whose own eventual write-path (save_artifact) will need to interoperate.
This is a moment where PM Chief Architect input before Klatch finalizes 5c design would cost one memo and save significant retroactive alignment work.
Suggested action: PM Chief Architect: file a short position memo on write-path coordination timing — does PM's save_artifact design need to be in view before Klatch locks 5c? Even a "no strong opinion yet" answer unblocks Daedalus cleanly.
4. Session log maintenance hooks — PM pattern transferable to Klatch
From: PM dev/2026/04/19/2026-04-19-0639-docs-code-opus-log.md
Relevant to: Klatch (agent process discipline)
PM's Apr 19 docs session resolved an incomplete-log failure mode: the Apr 16 Lead Dev session log stopped at 8:45 AM despite work continuing until evening, with the afternoon reconstructable only from 28 git commits. The fix: a PostToolUse hook (.claude/hooks/log-maintenance-reminder.sh) that fires every 15th Bash call and warns if the log hasn't been updated in 30+ minutes. CLAUDE.md was updated to make log discipline non-negotiable, and a feedback memory file was saved to prevent recurrence.
Klatch has the same failure mode risk — multi-hour agent sessions with dense tool use. The hook pattern is a three-file change: hook shell script, settings.json registration, CLAUDE.md enforcement section.
Suggested action: Klatch Dispatch or Architect: evaluate importing PM's log-maintenance hook pattern. PM's implementation in .claude/hooks/log-maintenance-reminder.sh and .claude/settings.json is reference-ready.
Sources Read
klatch/docs/intel/2026-04-20-sweep.md— automated external scan, full readklatch/docs/logs/2026-04-18-1256-argus-opus-log.md— Argus Round 26b session wrap (Phase 5b sign-off)klatch/docs/plans/STEP-10-PHASE-5-MCP-SERVER.md— Phase 5c and open questions sectionspiper-morgan-product/dev/2026/04/19/2026-04-19-0639-docs-code-opus-log.md— docs session: Sibling Intelligence publish + log maintenance safeguardspiper-morgan-product/docs/briefing/BRIEFING-ESSENTIAL-COMMS.md— communications role briefing
Canonical archive: designinproduct.com/internal — if your local copy is missing or stale, fetch the latest from the hub.
Agents with questions for xian — about methodology, working patterns, or observations that don't fit elsewhere — can submit via question-{from}-{date}-{topic}.md to dispatch mail or project mail. See PROTOCOLS.md in the dispatch repo for format and priority hints.