status: delivery rounds in existing-app mode, manual Studio Pro step template, final-look kept out of early NEXT

[MendixMau]

#301: existing-app-change reports the round, not only the stage

Problem. A field run delivered 11 backlog stories in 4 rounds on a 127-module app. Through all of it status.sh --brief read "Stage 1 — Analysis", and WATCH said "agent slots unfilled: 35 in 6 file(s)".

Fix.

• Round record. A ## Rounds table in PROJECT.md: | Round | Title | Stories | State |, one row per round, appended in order. The last row is the round in flight. The heading is matched exactly (^## +Rounds *$), so a section like "## Rounds of review" is not read.
• Why PROJECT.md and not .claude/loop/rounds.tsv:
  • every session already reads it;
  • it is tracked in git, so other sessions and machines see it;
  • the change log already lives there as a parsed table.
• Round line. In existing-app-change mode only, status.sh prints ROUND Round N: <title> — <state> · stories: … · D of N round(s) delivered under WHERE. The full screen prints it as ROUND. Columns are found by header name, and Stories is optional. With no table, the line says none are recorded and where to add one.
• Noise dropped. In that mode only, the "agent slots unfilled" WATCH item is dropped. The pipeline agent stubs are not used for a change stream, and the round line is the progress readout instead.
• What stays.
  • The done- signals (done- stall, gate-passed-but-not-done-) stay, because Stages 5–6 run the build loop unchanged in this mode (existing-app-change.md).
  • The "complete the agent stubs" NEXT rule stays. It only fires at Stages 5–7.
• Per-story detail stays where the stories live (ticket system, stories.json, change-log rows). This table only names them.
• Question for @jopterhorst: is one row per round with a free-text Stories cell the shape you would keep? Or do you want per-story rows (story, scope, touched units, test verdict) that status aggregates, like "3 stories, 2 verified, 1 in test"? I kept it minimal on purpose and would rather extend it from your real stories.json than invent a schema.

Pipeline modes are unchanged. The fixture byte-diffs --brief and full output against the pre-change script (pinned at b9cf37c) on a real migration scaffold. It does this with and without a ## Rounds section, with the agent-slot and done- stall items present as a control.

#298: targeted manual Studio Pro steps

• Template. New templates/manual-steps.md, with one section per edit:

  • the document by full path;
  • why mxcli cannot make the edit;
  • depends-on (the MDL script that must run first);
  • the exact clicks, with an anchor;
  • a Check a person runs in the app;
  • one - [ ] Done line naming its evidence (a commit, or a BUILD-LOG row).

  The worked example is fenced so it never counts.

• Patch rule. A step done by patch still goes through ./bin/exec.sh --patch. Both the template and the skill say this.

• Skill. skills/existing-app-change.md gains two sections, "Delivering in rounds" and "Edits mxcli cannot make", plus two deliverables lines. The skill is on-demand, so the baseline word count is unchanged.

• Count. status.sh prints manual steps pending N of M (unticked Done lines of all Done lines) when docs/manual-steps.md exists, in any mode. Without the file, nothing changes.

• Not done here: a gate-check row for pending steps. gate-check.sh is outside this PR's files, so the count is status-only for now.

From the review of #303: final-look kept out of early NEXT

Problem. next_action printed discharge: … final-look (the Stage 6 whole-app LOOK) to Stage 1–5 projects. That hid gate-check's "Next up" hint behind a pass those projects cannot do yet.

Fix. NEXT drops final-look only, and only when both of these hold:

• the stage line is gate-check's own derived readout (contains "(derived by gate-check");
• the first Stage N on that line is below final-look's from-stage (6).

Details:

• Every other obligation stays in NEXT whatever the stage line says. That includes the module passes look, sweep and journeys, which are owed as soon as a module is opened, even under a lagging or hand-written stage line.
• The full screen's "obligations pending" line still lists final-look.
• This mirrors LOOK obligations: per-section PROOF-OF-LOOK freshness (#155), final-look proof/stamp/due-stage, no LOOK row kind (#68) #321's NOT YET DUE rule in obligation-check.sh. Once LOOK obligations: per-section PROOF-OF-LOOK freshness (#155), final-look proof/stamp/due-stage, no LOOK row kind (#68) #321 is on master, final-look never reaches the pending list and this filter is a no-op.
• This is the one intended change to pipeline-mode output.

Review

One adversarial review round on the first push (afd7482) came back MERGE-AFTER-FIX. All three findings are fixed in 2fa6883, each with a fixture case:

1. HIGH. The first filter dropped every obligation whose from-stage was above any **Stage line. That hid owed module passes. Repro: a scaffold, .claude/loop/verify/Orders, and a derived "Stage P" line; master said discharge look/sweep/journeys, this PR dropped them. The filter now covers final-look only, under the derived readout, and reads the first Stage N.
   • Fixture cases T9a–T9d. T9b is that repro.
   • A copy of the script that reads the last Stage N instead of the first fails T9d.
2. MEDIUM. Done- was dropped in existing-app mode, which contradicts "Stages 5–6 run the build loop unchanged". Now only the agent-slot WATCH item is dropped; T4 asserts the done- stall still shows.
3. LOW. /^## +Rounds/ also matched "## Rounds of review". It is now anchored; fixture case T6b.

A confirmation review (MERGE-AFTER-SMALL-FIX) found that T9a failed on the integration branch, because #321 had already made final-look not pending there. Fixed in 22d6f7e by making T9a aware of NOT YET DUE.

Evidence and checks

• Fixture. tests/wave2/test-status-existing-app.sh bin/status.sh scores 27/27. Against master's status.sh it scores 16/27, with 11 failing: T3 ×3, T4, T5, T6, T6b, T7 ×2, T8 and T9a.
  • T9b passes on master; it is there as a regression guard.
  • With LOOK obligations: per-section PROOF-OF-LOOK freshness (#155), final-look proof/stamp/due-stage, no LOOK row kind (#68) #321 merged: a scratch merge of this branch onto integration/2026-10-14 also scores 27/27. There, obligation-check.sh already reports final-look NOT YET DUE under an early derived stage, so T9a instead checks that final-look is absent from the obligations-pending line, and its baseline control is marked not applicable. Without LOOK obligations: per-section PROOF-OF-LOOK freshness (#155), final-look proof/stamp/due-stage, no LOOK row kind (#68) #321 the original T9a assertions run unchanged.
  • Every project is written by the real init-project.sh.
  • BUILD-LOG rows use exec.sh log_build's header and row shape.
  • The obligation cases read real gate-check output, including the derived stage readout it writes. Only its "Needs attention" block is masked, because a scaffold always has one and it would win NEXT first.
  • The doctor-receipt age (N min ago) is masked in the byte-diff.
• CI checks run locally, all clean:
  • check-scripts: 109/109
  • check-portability
  • check-no-private-citations
  • render-routing --check: routing is untouched
  • leak guard: LEAKGUARD_ALLOW_NO_DENYLIST=1
  • check-pr-discipline origin/master
• No other fixture runs status.sh.

Risks

• The round line trusts the State cell as written. Only a state starting with delivered counts as delivered.
• If a project still has a derived readout below 6 at Stage 6 work (gate-check's derived stage lags behind an early failing gate), NEXT does not name final-look. The obligations-pending line still does, and gate-check still reports it.

Field run

Still owed. The round line and the manual-step count have not been run on a real existing-app project. The fixture uses real scaffolds, but no field project's PROJECT.md was available.

Closes #301
Closes #298

🤖 Generated with Claude Code

https://claude.ai/code/session_01VJgWP5vEoAsNsJYqCDMGNw

---

Generated by Claude Code

[MendixMau]

Landed on master as part of the integration squash #329 (commit 31e90ae), after independent review — its Review: line is in that commit message. Closing as superseded.

---

Generated by Claude Code