Session Retro — S265 (subo-workflow sub-orchestrator)

Second retro written under the S264-ratified habit (dossier §4.G). Doubles as a format proof-of-concept generalising beyond the main-track S264 record — the same 6 categories, a different track/role. Will migrate into the retro ledger as a record under {48.3}.

Session: S265 · 2026-05-25 · subo-workflow sub-orchestrator (Task ID-48 {48.1}+{48.2}, co-investigating ID-23) · worker branch cmux-worker-subo-workflow-e472c743 → af9d4e03

1. Bugs discovered

Reference-doc drift: the empirical-grounding axis exists in task-checker.md (L337) but is missing from checker-output-schema.md’s axis enum (~L52). Captured as Subtask {48.7}.
session_refs format inconsistency: the new ID-48 ref was written bare (S265) vs the kh-{track}-S{N} convention (kh-main-S264). Fixed in-session → kh-subo-workflow-S265.
Self-inflicted (workflow gotcha): a large inline node -e '...' script with parens / pipes / ${} choked zsh (parse error near )). Fix = write to a temp .cjs and run it.

2. Failed assumptions

“The validate-reality-before-spec guardrail must be built.” FALSE — it already exists comprehensively (task-planner.md §OQ-3 + task-checker.md §empirical-grounding). This was the meta-loop catch: the dossier framed the existing mandate as net-new work. A survey agent surfaced it; I verified it by direct read before letting it reshape scope.
“The two confirmed seeds are the same failure mode.” FALSE — ID-32 B4 (internal contract / corpus shape) and ID-28 bind_target (external API) straddle OQ-3’s explicit internal/external scope boundary. That boundary IS the insight.
Dossier note F: “/code-review is a rename of /simplify.” FALSE — verified by reading both installed surfaces; they do different jobs.

3. Architecture / design decisions

Root-cause taxonomy split into Sub-type A (external-API, already covered) vs Sub-type B (internal-contract/corpus-shape, net-new) — drives the whole PLAN.
Retro surface = 4th JSON ledger AND keep the Mempalace diary (complementary jobs).
workflow-evaluator = agent + companion skill combo (task-executor↔implement-subtask precedent).
Hook-1 (block agent/skill edits) = sentinel-file mechanism (PreToolUse sees tool input only, not conversation history).
Sibling-only deps preserved; the cross-Task ID-23 sequencing was escalated as OQ-C, not bent — the forcing function worked as designed.

4. Rejected approaches

Re-authoring the empirical-verification mandate — rejected (exists).
Mass-renaming 27 spec dirs to ID-prefix — rejected (churn; forward-only convention).
Renaming docs/specs/workflow-evaluation/ mid-investigation — rejected (breaks the dossier/retro/task-list/PLAN cross-refs); deferred to a post-close subtask.
Editing the shared SDLC role/skill files in parallel with ID-23 — rejected (file collision); sequence-after / fold-in.
Implementing {48.3}+ this session — rejected (OQ-gated + ID-23-gated).

5. Workflow improvements

The ID-48 programme itself (root-caused, decomposed, OQs isolated).
Lean-context fan-out: three parallel surveys returned conclusions-only, keeping the orchestrator context small — the efficiency discipline ID-48 catalogues, applied to its own run.
Load-bearing verification: verified the single most scope-reshaping agent conclusion (OQ-3 already exists) by direct read before acting — the exact meta-loop discipline this Task institutes.
New gotcha for capture: inline node -e with special chars → temp .cjs file.

6. Unresolved questions

The 5 in OQ-pending.md (OQ-A retro-ledger build; OQ-B ID-prefix + rename timing; OQ-C ID-23 cross-Task sequencing; OQ-D evaluator efficiency-eval cadence; OQ-E PRODUCT-skip).
Does the parent O-of-O want a final_report.yaml (id32-style) from this terminal, or is the worker-branch commit + this retro + OQ-pending sufficient for teardown integration?