Standing: read the dev-workflow feedback log and look for systemic issues nf-5t1 ← Beads

Twinned from https://github.com/cfsh/ninefold/issues/771 by tools/beads/import_github.py's reconcile pass.

Standing task, filed with the design_docs/006 round so the log does not become a channel nobody reads.

Read ~/nf/data/dev-feedback.jsonl periodically and look for the systemic shape several entries share.

Why this issue exists

dev.py feedback (design 006 S16, implementation #768) gives dev sessions a place to log friction *without* filing a patch-shaped issue. @aedanpope's diagnosis, 2026-08-09:

> "we've seen dev instances spawn a proliferation of minor dev workflow bugs and nits, often these bugs are proposing patch solutions ontop of patch solutions. Dev workflows have a lot of context on the problem they are solving, and the friction they are hitting - but not necessarily the best context on how to fix their friction."

⚠️ Feedback nobody reads is worse than no feedback, because it launders the obligation: the dev discharges the report, and nothing happens. This issue is the guard against that, and design 006 §5.10 says plainly that if this issue goes stale, S16 should be closed rather than left absorbing reports.

What the pass is

Not triage of individual entries. Group by seam, not by report — the value is in what several devs hit at the same place, which is exactly the vantage no individual session has.

• Read the log since the last pass. Entries carry severity (blocked / friction / idea), the session, the worktree and head sha, the last command, and a pointer to that session's transcript JSONL.
• ⚠️ Follow the transcript pointers. Design 006 §2.12 is the existence proof that transcripts answer questions no summary does — three sweeps over reaped transcripts found vectors (a silently clamped timeout, --body corrupting posted text, gh resolving its repo from cwd) that no filed issue had.
• Scan severity: blocked first.
• Output is a small number of systemic findings — or a design round if the shape is big enough — not one issue per entry. Turning the log back into a 1:1 issue stream would reproduce exactly what #381 measured: nine residual issues from nine PRs, ≈1:1 with PR count, with no natural termination.

Current scale of the problem this is meant to reduce

Measured 2026-08-09: 45 of 197 open issues name a workflow tool (poll, merge_pr, review_now, wrap_up, presubmit, session_state, probe_gate, dispatch). Design 006 §2.7 counts 26 against the poll loop alone. If S16 works, that arrival rate falls and this pass is where the replacement signal shows up.

Cadence

Unset deliberately — nobody has watched the log fill yet, and picking an interval now would be the anticipated-need pre-building B4 warns against. A sensible first pass is once the file has meaningful volume; whoever takes it can propose the cadence from what they find.

Blocked on S16 landing (#768) — there is no file until then. Filed now so it is not forgotten, per @aedanpope's ask.

Unclaimed and unlabelled — backlog, for a triage round.