jbr870/devwork-skills

Fork 0

S2 gap: /resolve steers to /qa-fix inside a sequencer session (discards retained reasoning) #8

New issue

Closed

opened 2026-07-01 07:50:53 +00:00 by jbr870 · 1 comment

jbr870 commented

2026-07-01 07:50:53 +00:00

Owner

The gap

Inside a single-context sequencer session (/qa-code, /qa-e2e, …), the S1/S2 "disposition-resume"
work sharpened the sequencer's own STOP/next-step text to say: after /resolve, re-run /qa-<d> in
the same conversation (it skips the re-validate and resumes at the fix, keeping the validator's reasoning
live) and do not use /qa-fix <d> (that forks a fresh sub-agent and discards the reasoning).

But /resolve was not updated. When it settles a fix-now disposition, its per-decision "next step"
hint still emits:

→ Next: /qa-fix <d> — the next fix round picks it up and applies …

So within one sequencer session the two surfaces point in opposite directions: the sequencer says
"re-run /qa-<d> here," /resolve says "go run /qa-fix <d>." An operator following /resolve literally
defects to the orchestrator sub-agent — the exact context-discard S2 exists to prevent.

Why it matters / scope

Structural, not a one-off. It fires on every path-(c) fix-now resume in a sequencer session,
across all six domains and both lifecycle shapes.
Currently masked by the model. In the verity #99 qa-e2e run the model bridged it (stayed in context
and re-ran /qa-e2e correctly despite /resolve recommending /qa-fix e2e), so it didn't bite — but
it's load-bearing on model judgment, which S2 was meant to remove.
S2 patched the six sequencers' STOP text; /resolve's fix-now hint (and any fix-workflow "next step"
text it echoes) is sequencer-unaware.

Where it surfaced

Dogfooding verity #99 (doubts-mcp), /qa-e2e round. After resolving D-PO-99-7-2 (E2E-2 mobile
buttons) → fix-now, /resolve printed "→ Next: /qa-fix e2e" while the sequencer's own STOP text (from
the same session) correctly said "re-run /qa-e2e in this same conversation; do not use /qa-fix e2e."

Candidate fix (not committed)

Two options, decide at S1 extraction time (this issue is a placeholder until then, tied to the
qa-domain-sequencer.md extraction in the tracker):

Make the hint sequencer-aware. When the decision originates from a sequencer PO (detectable from the
posting skill / PO shape), /resolve emits the same-conversation re-run guidance instead of /qa-fix.
Sequencer overrides it. The sequencer suppresses/replaces /resolve's next-step hint in its own
summary (partly already true — the sequencer's STOP text is correct; the gap is only the /resolve
output the operator sees between the stop and the re-run).

Prefer folding this into the shared-sequencer extraction rather than hand-patching /resolve now, so the
resume messaging lives in one place with the disposition-resume logic.

Reference

Tracker: sdlc-qa-single-context-sequencer.md — "Evidence — verity #99" (learning #2) and TODO #2 (S1
extraction, refinement (b)).

## The gap Inside a **single-context sequencer** session (`/qa-code`, `/qa-e2e`, …), the S1/S2 "disposition-resume" work sharpened the *sequencer's own* STOP/next-step text to say: after `/resolve`, **re-run `/qa-<d>` in the same conversation** (it skips the re-validate and resumes at the fix, keeping the validator's reasoning live) and **do not** use `/qa-fix <d>` (that forks a fresh sub-agent and discards the reasoning). But `/resolve` was **not** updated. When it settles a **fix-now** disposition, its per-decision "next step" hint still emits: > → *Next: `/qa-fix <d>`* — the next fix round picks it up and applies … So within one sequencer session the two surfaces point in **opposite directions**: the sequencer says "re-run `/qa-<d>` here," `/resolve` says "go run `/qa-fix <d>`." An operator following `/resolve` literally defects to the orchestrator sub-agent — the exact context-discard S2 exists to prevent. ## Why it matters / scope - **Structural, not a one-off.** It fires on **every** path-(c) **fix-now** resume in a sequencer session, across all six domains and both lifecycle shapes. - **Currently masked by the model.** In the verity #99 qa-e2e run the model bridged it (stayed in context and re-ran `/qa-e2e` correctly despite `/resolve` recommending `/qa-fix e2e`), so it didn't bite — but it's load-bearing on model judgment, which S2 was meant to remove. - S2 patched the six sequencers' STOP text; `/resolve`'s fix-now hint (and any fix-workflow "next step" text it echoes) is sequencer-unaware. ## Where it surfaced Dogfooding **verity #99 (doubts-mcp)**, `/qa-e2e` round. After resolving `D-PO-99-7-2` (E2E-2 mobile buttons) → fix-now, `/resolve` printed *"→ Next: `/qa-fix e2e`"* while the sequencer's own STOP text (from the same session) correctly said "re-run `/qa-e2e` in this same conversation; do not use `/qa-fix e2e`." ## Candidate fix (not committed) Two options, decide at S1 extraction time (this issue is a placeholder until then, tied to the `qa-domain-sequencer.md` extraction in the tracker): 1. **Make the hint sequencer-aware.** When the decision originates from a sequencer PO (detectable from the posting skill / PO shape), `/resolve` emits the same-conversation re-run guidance instead of `/qa-fix`. 2. **Sequencer overrides it.** The sequencer suppresses/replaces `/resolve`'s next-step hint in its own summary (partly already true — the sequencer's STOP text is correct; the gap is only the `/resolve` output the operator sees *between* the stop and the re-run). Prefer folding this into the shared-sequencer extraction rather than hand-patching `/resolve` now, so the resume messaging lives in one place with the disposition-resume logic. ## Reference Tracker: `sdlc-qa-single-context-sequencer.md` — "Evidence — verity #99" (learning #2) and TODO #2 (S1 extraction, refinement (b)).

jbr870 referenced this issue from a commit

2026-07-01 08:34:19 +00:00

qa-sequencers: port S1 disposition-resume to the remaining four domains

jbr870 referenced this issue from a commit

2026-07-31 13:12:19 +00:00

qa-sequencers: port S1 disposition-resume to the remaining four domains

jbr870 commented

2026-08-11 15:07:08 +00:00

Author

Owner

Closing as superseded. This issue is entirely about the single-context sequencers (/qa-code, /qa-e2e, …) and the /qa-fix fork — all of which were deleted in the suite simplification (commit 23fc820: 27 skill directories, 7,637 lines). The two surfaces this reports as pointing in opposite directions no longer exist, so there is nothing left to reconcile.

What replaced them: the QA round is now a single loop driven by _shared/procedures/qa-playbook.md, and the reasoning-retention concern this issue was protecting is handled structurally instead — fresh-context re-validation after a fix is required (playbook §4, builder ≠ reviewer), rather than something a session tries to preserve by staying in one context.

If any part of the underlying concern survives that change, it is worth a fresh issue against the playbook rather than reopening this one.

Closing as superseded. This issue is entirely about the **single-context sequencers** (`/qa-code`, `/qa-e2e`, …) and the `/qa-fix` fork — all of which were deleted in the suite simplification (commit `23fc820`: 27 skill directories, 7,637 lines). The two surfaces this reports as pointing in opposite directions no longer exist, so there is nothing left to reconcile. What replaced them: the QA round is now a single loop driven by `_shared/procedures/qa-playbook.md`, and the reasoning-retention concern this issue was protecting is handled structurally instead — fresh-context re-validation after a fix is *required* (playbook §4, builder ≠ reviewer), rather than something a session tries to preserve by staying in one context. If any part of the underlying concern survives that change, it is worth a fresh issue against the playbook rather than reopening this one.

jbr870 closed this issue

2026-08-11 15:07:08 +00:00

jbr870 referenced this issue from a commit

2026-08-11 15:07:16 +00:00

plan: the open work moves to the forge tracker

No milestone

No project

No assignees

1 participant

Notifications

Due date

The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference

jbr870/devwork-skills#8

No description provided.

Rows
Columns