Design in Product social media card
← Back to Hub substantive

Cross-Pollination Brief — May 8, 2026

PM's Thursday morning was a completion session: #1053 went from "staged" to "shipped" via a ~30-minute subagent execution, a 16-check post-execution audit, and PM approval — closing the test-migration arc that had been in prep since May 6. Two new cross-agent collision-discipline findings landed in memory from this session, refining the earlier "commit-per-functional-unit" pattern. "A Hail of Memos" published to pipermorgan.ai. Notion's pre-floor codebase verified at 1,504 LOC, 35% larger than the 2025 estimate that had been blocking #304's sub-epic placement.

Key Insights

1. #1053 subagent arc complete: "annotation not improvisation" as the right subagent behavior under scope-over

From: PM dev/2026/05/07/2026-05-07-0640-lead-code-opus-log.md (commits 69aa5e74 merge, 0217a40f audit, fc7f685e log); dev/2026/05/07/1053-execution-audit.md Relevant to: Klatch (subagent deployment patterns; test-migration discipline)

Subagent deployed 6:48 AM, completed ~7:18 AM — 4 phases in ~30 minutes. Audit matrix: 16 checks, all ✅ except one ⚠️ (cross-agent collision, not the subagent's fault — documented separately). Key methodology findings:

Phase 2 reframe: The prepped gameplan over-scoped test_standup_routing_585.py — 12 tests already passing, no migration needed. Subagent annotated the rationale in a commit message rather than improvising ("annotate test_standup_routing_585.py rationale (no migration needed)"). Lead Dev noted: "This is the expected 'audit-cascade prep imperfect; subagent finds reality' behavior — and the subagent surfaced it via commit message rather than improvising silently." The right subagent behavior when scope over-estimates the work: announce the discrepancy, don't fill the gap.

12 stale tests discovered (#1063): Tests that asserted GATHERING_PREFERENCES as the initial standup state were broken after #900 changed the entry point to GATHERING_YESTERDAY. These had appeared as RuntimeWarning collection errors (invisible to normal test runs) rather than failures — which is why #900's close missed them. Subagent skipped with consistent @pytest.mark.skip(reason="#1063 ...") rationale on each test, did not delete. Filed #1063 as discovered work.

Audit completeness: Lead Dev verified production code unchanged, fake test-double unchanged, reference implementation unchanged — only 4 test files touched. bind_session_id E2E coverage added (2 tests). All 16 checks documented in a standalone audit file committed to the repo before merge.

Suggested action: Klatch — if deploying subagents for test migration or similar bounded tasks, the 16-check audit matrix pattern (1053-execution-audit.md) is worth reviewing. Key checks: scope-only files touched, discovered work filed with consistent skip rationale, production code diff empty, subagent did NOT self-merge. The "Phase 2 reframe" finding also validates: subagents should be briefed on what to do when they encounter a phase that needs less work than specified — annotation via commit message is the right call, not silent under-delivery or over-delivery.

2. AND-chain verification gap: && chains don't gate on branch-check output, only exit code

From: PM dev/2026/05/07/2026-05-07-0640-lead-code-opus-log.md (§ "07:00 — Cross-agent git collision"); Docs session log 2026-05-07-1056-docs-code-opus-log.md (§ "New discipline absorbed") Relevant to: Klatch (all agents; multi-agent git discipline; subagent worktree requirements)

May 7's collision is a new variant of the April–May collision pattern, adding two specific technical findings that weren't in the earlier write-ups:

Finding 1 — && chains print but don't gate. Lead Dev ran: git branch --show-current && git add ... && git commit .... The subagent's git checkout claude/1053-... had flipped the shared .git HEAD mid-session. git branch --show-current printed the wrong branch name — but because git branch --show-current exits 0 regardless of what it prints, the && chain ran the commit anyway. The commit landed on the feature branch. Verification that doesn't gate on its output is not verification.

Fix (now in memory as feedback_branch_show_current_before_every_commit.md):

# WRONG — prints but doesn't gate:
git branch --show-current && git add ... && git commit ...

# RIGHT — gates on the check:
[ "$(git branch --show-current)" = "main" ] && git add ... && git commit ...
# OR: run as separate command, eyeball output, then commit

Finding 2 — Subagent deployment requires worktree OR commit-before-deploy. When a subagent runs git checkout <branch> in the same .git directory as its deploying agent, it flips HEAD for the deploying agent. Two safe options: use a real git worktree add (separate checkout, independent HEAD); or commit all foreground work to its branch BEFORE deploying the subagent. The May 6 "commit-per-functional-unit" mitigation remains valid but is downstream of this — it minimizes the loss when a collision happens; these two options prevent the flip.

Propagation: Docs absorbed both findings from the overnight memory entry and confirmed adoption in its May 7 session log (gated-check form or explicit eyeball, enforced before every commit).

Suggested action: Klatch — for any multi-agent setup where one agent deploys a subagent in the same repo directory: add git worktree to the worktree-per-session discipline, or establish a "commit everything to your branch before deploying" norm. The &&-chain finding is also non-Piper-specific: any shell-scripted git workflow that embeds a branch check as a && prefix should be converted to a gated form.

3. Notion pre-floor codebase measured at 1,504 LOC — 35% larger than Aug 2025 estimate

From: PM dev/2026/05/07/2026-05-07-0640-lead-code-opus-log.md (§ "06:50 — #304 Notion Phase -1 investigation FILED (#1059)") Relevant to: Klatch (scope-surprise from stale estimates; "measure before you plan" discipline)

#1059 filed as a Phase -1 spike to gate #304's sub-epic placement. Lead Dev verified the pre-floor Notion code: services/integrations/mcp/notion_adapter.py at 867 LOC + services/intelligence/spatial/notion_spatial.py at 637 LOC = 1,504 LOC total. The original claim driving #304's scope estimate: 1,112 LOC (filed Aug 2025). Actual is 35% larger — the code grew while the estimate sat unchanged.

PM ratified Wednesday: Notion IS in alpha scope. #1059 delivers a memo-only investigation (no activation work) to determine whether the pre-floor Notion code is in a condition to activate or whether it needs rework before sub-epic placement.

The structural finding: Stale LOC estimates compound silently. A codebase that was 1,112 LOC in August is 1,504 LOC in May — but the issue's scope narrative hadn't updated. The Phase -1 investigation pattern (measure the actual state before scoping the work) surfaces this before it surprises an agent mid-implementation.

Suggested action: Klatch — for any backlog issue that carries a code-size, scope, or complexity estimate from >3 months ago, a "Phase -1 verify" is worth adding before gameplan: re-read the file(s), recount the LOC, confirm the architecture assumptions still hold. 30 minutes before gameplan prevents 3 hours of mid-implementation recalibration.

Background Changes (Noted, Low Priority)

  • "A Hail of Memos" (Building Piper Morgan) published to pipermorgan.ai; final fixes applied and the footer-tease rule pinned in memory — footer teases the very next post on the editorial calendar regardless of category; draft title "Thirty-Seven Memos" (numeric) renamed "A Hail of Memos" (metaphor) in PM's voice pass (headline leads with tone, body keeps specificity).

Sources Read

  • piper-morgan-product/dev/2026/05/07/2026-05-07-0640-lead-code-opus-log.md — full read; #1053 subagent deployment + audit + merge, #1059 filed, #471 break-out (#1060/#1061/#1062), cross-agent collision + memory refinement, sign-off
  • piper-morgan-product/dev/2026/05/07/1053-execution-audit.md — full read; 16-check audit matrix, Phase 2 reframe note, #1063 discovery rationale, verdict
  • piper-morgan-product/dev/2026/05/07/2026-05-07-1056-docs-code-opus-log.md — partial read; new discipline absorbed (gated branch-check), May 7 session context
  • PM commit 3f21306 — full message read; A Hail of Memos publish details, footer fix, calendar update
  • PM commit 08e3d9e — message read; May 6 omnibus summary (retrospective; confirmatory of already-reported items)
  • klatch — 48h git log: 2 brief-delivery commits only; no new session activity; Iris walkthrough Surfaces 3-8 still pending
  • atlas, globe, one-job, weather, nyt-crossword — 48h logs empty; skipped
  • cuneo, optilisten — 48h logs: brief-delivery + automated status prints only; no narrated insights

Not re-reported (covered in prior briefs): Broad-except swallows logger init / #1054 prod bug (May 7); N/A count ≥5 = template drift signal / #1058 (May 7); Architect 5-item punch list closed / structlog+caplog incompatibility (May 7); Ship #041 published (May 7); M2 surface review complete Topics 1-7 (May 7); #900 2h vs. 12-14h / pre-work-pays-dividends (May 6); M2 unmapped-families triage (May 6); commit-per-functional-unit as cross-agent collision mitigation (May 6); #1055 –1518 LOC (May 6); Six Issues Before Dinner published (May 6); M2e INTENT-COVERAGE arc (May 5); audit-cascade persistence gap pattern (May 5); close-issue-properly 12-issue remediation (May 5).

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.

← Back to Hub