Cross-Pollination Brief — May 10, 2026

PM's Saturday was the densest single session of the M2f arc: 9 issues closed, −2,229 LOC removed, 4 worktrees used, and 2 subagents deployed — all in one day. The headline finding: 3 of 5 M2f Group A+B issues had body-vs-reality mismatches where audit-cascade Phase 0 discovered the issue was describing dead or unreachable code. Instead of multi-thousand-LOC migrations, the disposition was deletion. Pattern-067 (Issue-Body Reality Mismatch) filed Emerging. Separately: the rubric recalibration work that followed Friday's #1064 investigation landed clean — Run 7 at 68.9% PASS exceeds the Apr 12 baseline and clears the CEO's M2f gate criterion. Hub CLAUDE.md gained new Multi-Agent Operation discipline; Janus Layer 5 mandate expanded.

Key Insights

1. M2f Group A+B complete: −2,229 LOC via dead-code disposition; Pattern-067 filed

From: PM session log dev/2026/05/09/2026-05-09-0630-lead-code-opus-log.md; commits b62a9080 (#936 deletion), 82bca29c (#935 deletion), 80cbd586 (#933 merge), c9591108 (#932 merge); CURRENT-STATE refresh 15942063; Pattern-067 commit a2bd06d9 Relevant to: Klatch (all agents; audit-cascade discipline; issue-triage methodology; worktree adoption)

Saturday's M2f Group A+B work — covering security (#932 HIBP stub, #933 re-enable key validation) and persistence (#936 UserService, #935 analytics) — produced a striking result: 3 of 5 issues disposed via deletion rather than implementation.

• #933 (re-enable API key validation): bypass had been disabled Oct 30, 2025 to work around format-validator issues. Those issues were fixed in the same Oct 30 commit. The bypass outlasted its reason by 6+ months. Removing it un-broke 5 tests rather than breaking any. Phase 0 investigation caught this; without it, the issue body would have been implemented as written.
• #936 (UserService db persistence): issue body claimed "user data lost on restart." Investigation: UserService wired into AuthMiddleware but create_session() / create_user() have zero production callsites. Real auth is PostgreSQL + AuthService + JWT. UserService always returns None; the data-loss scenario is unreachable. PM disposition: Option A (delete). −435 LOC.
• #935 (BudgetManager/APIUsageTracker persistence): APIUsageTracker has a real INSERT into api_usage_logs (table exists, 0 rows), but both production callers pass without a session parameter — the INSERT gate never fires. Transitive dead code. PM disposition: Option A (delete) + alembic migration to drop the table. −1,378 LOC.

Net: −2,229 LOC across the group. #932 and #933 (the real implementation work) contributed a combined −51 LOC net (fixes + new tests), offsetting the deletions at the margin.

Worktree adoption: Lead Dev created a dedicated worktree for each issue (piper-morgan-product-932, -933, -936, -935) — fully operationalizing the "worktree per substantive session" norm. Cross-agent collision still occurred mid-#932 (another agent's git checkout flipped the shared HEAD); the pre-commit gated branch check caught it with exit 1 and no bad commit landed. Recovery via re-checkout was clean. Subagents for #932 and #933 created their own worktrees independently, confirming the collision-discipline pattern is working at both agent and sub-agent scope.

Pattern-067 (Issue-Body Reality Mismatch) filed Emerging by Lead Dev (self-approval per Pattern-066 precedent): "prophylactic Phase 0 dead-code/unreachable check on issue bodies matching trigger conditions: 'TODO to enable X', 'lost on restart', line-number citations from triage runs N months ago, parenthetical alternatives in acceptance criteria." Promotion to Proven pending one trial-application cycle. Aggregate observed: ~−2,229 LOC removed via deletion instead of migration additions.

Suggested action: Klatch — when any issue body contains the trigger-condition language above, run a Phase 0 dead-code investigation before scoping the migration. The PM M2f experience shows the cost of skipping this is implementing features nobody uses. The "issue body describes a world that no longer exists" shape is especially common for issues originally triaged during active development when code paths were still hypothetical.

2. Canonical retest Run 7: 68.9% PASS — M2f baseline established via rubric recalibration + 3 bug fixes

From: PM session log dev/2026/05/09/2026-05-09-0630-lead-code-opus-log.md (§ 06:35–08:15); commits 49c48c2e (#1065), f623bba2 (#1067), 78be342e (#1066), 14259cdd (Run 7); 15942063 (CURRENT-STATE refresh noting 68.9% exceeds Apr 12 baseline) Relevant to: Klatch (Argus; eval harness methodology; judge calibration; fixture management)

Following Friday's #1064 investigation (fabrication hypothesis refuted), Saturday morning completed the remediation path and ran a clean baseline.

Rubric recalibration (Runs 5–6 before fixes): Run 5 with fixture reset + softened auto-fail rule moved only Q56 (FAIL→PASS) — verdict-rubric gap discovered: the judge prompt's PASS/MARGINAL criteria still required "no zeros" while FAIL only fired on 2+ zeros; single-zero responses fell into a scoring limbo and defaulted to FAIL. Run 6 closed the gap (PASS criteria: total >= 7; MARGINAL: total in {5,6}). Remaining 8 FAILs in Run 6 were all legitimate.

3 bug fixes shipped (30–45 min each):

• #1065: 3 hardcoded "setup wizard" references in intent_service.py:4436/4555/4647 replaced with natural-language guidance. Fixed Q8/Q31/Q31 FAILs.
• #1067: 4th subsumption rule in _apply_subsumption_filter — when QUERY is update_document_query AND PORTFOLIO also matched, drop PORTFOLIO. Fixed Q40 (double-intent false positive). Regression-tested "archive project foo" still routes correctly.
• #1066: #N slot-fill regex fallback in _handle_update_issue matching the pattern already present in 3 sibling handlers. Fixed Q58.

Run 7 result:

68.9% exceeds Apr 12 baseline (65.6%). Apr 16's 72.1% peak is pre-recalibration and not directly comparable. CEO benchmark criterion ("preceding work meets the most recent benchmarks") met. M2f audit-cascade unblocked.

Discovered work filed: #1068 (2 pre-existing test failures in milestone routing, unrelated to #1067), #1069 (templated attention_query handler, P:low cosmetic), #1070 (multi-turn evaluation harness, P:low methodology). Rubric-recalibration memo distributed to CXO+PPM before proceeding — PM-authorized "proceed without sign-off, review can land after."

Suggested action: Klatch (Argus) — if the eval harness has a similar verdict-rubric gap (PASS/MARGINAL criteria that are stricter than FAIL criteria on the same scale), single-dimension failures can accumulate as FAILs even when the overall response is marginal-quality. Worth auditing the rubric symmetry: does the judge's FAIL condition have a lower bar than PASS+MARGINAL? If so, mid-quality responses fall into a scoring limbo. The fixture-reset bake-in pattern (idempotent cleanup script that runs before each retest) is also worth adopting directly.

3. Hub CLAUDE.md: Multi-Agent Operation discipline codified; Janus Layer 5 mandate expanded

From: Hub commits 89cae65 (Multi-Agent Operation subsection), f20dfbc (concurrent-operation discipline), dc7797b (Janus Layer 5 mandate expansion); all xian-approved May 9 Relevant to: Klatch (Daedalus, Mnemosyne, and any other multi-agent coordination; Layer 5 self-stewardship)

Two structural changes to the hub's shared CLAUDE.md (Layer 2), both approved Saturday morning:

Multi-Agent Operation subsection establishes three rules for Janus+Themis on the same working tree:

1. Mail-on-main: Anything under docs/mail/ lands on main directly, never on feature branches. Mail's purpose is to be read by the other agent, so it bypasses branch isolation. Feature branches are for code, entity-file updates, CLAUDE.md edits.
2. Branch hygiene for non-mail work: Fetch main before starting if a branch is more than a few days old. Rebase with --force-with-lease before merge.
3. Lightweight concurrent-operation: Commit and push your own work promptly; before structural git ops (rebase, force-push), fetch and check recent activity (git log origin/main --oneline -5). If you see uncommitted work that isn't yours, surface it to xian rather than touching it.

The note in the PR: "Heavier coordination (worktrees, branch registries, merge-keeper sweeps) is the PM PA synthesis template if friction surfaces, but DinP's two-agent setup doesn't need it yet."

Janus Layer 5 mandate expansion (36 LOC added, 5 changed) — from a single "scope" paragraph to four named working areas: Public-facing surfaces, Cross-pollination flow, Federated mail, Operational hygiene. Key additions: Operating Principles (6 explicit principles, including "Kind, not nice"), Active Stewardship of Own Context section ("Janus is a partner in defining who Janus is, not a passive recipient of an externally-authored definition"), and an expanded Boundary with Themis that routes Themis's strategic demands on the website via mail.

Suggested action: Klatch — the mail-on-main rule and the Active Stewardship principle are both transferable. For multi-agent git coordination: establishing a single file class (mail, or a mail-equivalent) that always lands on main prevents the "is this memo visible to the other agent?" problem. For Layer 5 maintenance: the Active Stewardship framing (agent proposes refinements to own scope at session tidy-up for xian approval) is cleaner than waiting for xian to drive the update. If Daedalus or Mnemosyne notices drift between their Layer 5 and current practice, the right move is proposing the fix, not flagging it as an observation.

Sources Read

• piper-morgan-product/dev/2026/05/09/2026-05-09-0630-lead-code-opus-log.md — full read; M2f Group A+B end-to-end session log
• piper-morgan-product commit a2bd06d9 — Pattern-067 (Issue-Body Reality Mismatch) filed Emerging; 115 LOC
• piper-morgan-product commit 15942063 — BRIEFING-CURRENT-STATE refresh; M2f-A+B complete, M2f-C deferred, Run 7 baseline
• piper-morgan-product commits 9832bb5a, 5d06a44e — #1072 regex→pattern fix (deprecated API, not cross-relevant)
• designinproduct commits 89cae65, f20dfbc — CLAUDE.md Multi-Agent Operation subsection + concurrent-operation discipline
• designinproduct commit dc7797b — Janus Layer 5 mandate expansion; diff read in full
• klatch — 48h log: brief-delivery commits only; no new session activity
• weather — 48h log: brief-delivery commits only; no new narrated insights
• atlas, globe, cuneo, one-job, optilisten, nyt-crossword — 48h logs empty; skipped

Not re-reported (covered in prior briefs): #1064 fabrication hypothesis refuted (May 9); PreCompact hook (#86) ships (May 9); Patterns-063/064/065 promoted (May 9); #1063 stale tests closed (May 9); &&-chain gating gap (May 8); worktree-required-for-subagents (May 8); #1053 subagent arc (May 8).

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.