# Session 0017.0 — Transcript

> App: vscode-cowriting-plugin
> Date: 2026-06-10T23-52 → 2026-06-11T06-59 (PST)
> Type: brainstorming
> Goal: Brainstorm Feature #17 (diff-view toggle) into its Solution Design —
> the baseline fork (turn start / last accept / save / pinned) is the
> load-bearing decision; constraints and non-goals in the issue body.
> Outcome: **Spec graduated** — content `specs/coauthoring-diff-view.md`
> (F6, v0.1.0, commit `36bc4e3`). All forks resolved; INV-18/INV-19 added.

## Launch prompt

```
/goal Brainstorm Feature #17 (diff-view toggle) into its Solution Design — the baseline fork (turn start / last accept / save / pinned) is the load-bearing decision; constraints and non-goals are in the issue body
```

## Plan

Brainstorm Feature #17 (diff-view toggle) into its Solution Design.

- Read issue #17 (constraints + non-goals in the body) and the existing
  F2–F5 specs/code as input.
- Work the load-bearing decision first: the **baseline fork** — what the
  diff view diffs *against* (turn start / last accept / save / pinned).
- Explore the remaining design forks, write the Solution Design under the
  content repo's `specs/`, graduate it at finalize (autonomous default).

## Pre-session state

- `main` at `611a180` (PR #16 — proposal threads decide-only fix), in sync
  with origin. Epic #1 fully shipped (F2 #4 · F3 #6 · F4 #12 · F5 #14).
- One dirty file: `sandbox/playground.md` — a 1-line accidental truncation
  ("thre" for "thread") left from the operator's manual play before this
  session. Noted at init; restored at finalize.
- Memory resume pointer (f5-cross-rung-shipped): "Next /goal: Brainstorm
  Feature #17 …" — exactly this session's goal.
- No other in-flight sessions; claim of 0017 was uncontended.

## Turn-by-turn arc

1. **Init** — classified the `/goal` as brainstorming; claimed session 0017
   (`--type brainstorming`, start 2026-06-10T23-52); seeded the Plan block;
   verified the CLAUDE.md org-context import; confirmed the baseline
   (noting the playground stray); resolved app facts (roadmap = this repo;
   content = `vscode-cowriting-plugin-content`; **no BDD corpus** — nothing
   to consume as gherkin input).
2. **Anchor gate** — fetched issue #17 from Gitea (`gitea-api.sh`,
   git.benstull.org): `type/feature`, P2, constraints + firm non-goals in
   the body. Eligible brainstorming anchor (R3: a feature produces its
   design first).
3. **Context exploration** — read the F4 spec (`coauthoring-propose-accept.md`)
   as the form reference, plus `model.ts`, `liveTurn.ts`,
   `proposalController.ts`, `extension.ts`, `store.ts`, and the seam
   (`attributionController.ts:232`). The decisive structural fact:
   **INV-9 + INV-10 mean the machine writes the document only through
   `applyAgentEdit`, only at accept time — so between machine landings,
   every buffer change is operator-authored.**
4. **The baseline fork (load-bearing)** — resolved to a
   **machine-landing–anchored auto-baseline + human pin**: initialize at
   first track, advance on every successful seam application; the diff is
   operator-authored *by construction* (new **INV-18**). The issue's other
   candidates fell away on the merits: turn-start collapses into accept
   under F4's propose-by-default (the document doesn't change at turn
   start); last-save duplicates native compare-with-saved/local-history;
   fixed session-start mixes machine landings in; a multi-baseline selector
   is YAGNI (pin covers the operator-chosen epoch).
5. **Secondary forks** — snapshot home = extension workspace storage via a
   vscode-free `BaselineStore`, never the sidecar (**INV-19**: no
   cross-rung contract impact, no `.threads/` churn); whose-changes =
   whole-document diff, no F3 filtering (state-not-history); view primitive
   = native `vscode.diff` with readonly `cowriting-baseline:` left and the
   **live document** right (editing continues in the diff view); landing
   signal = one additive `AttributionController.onDidApplyAgentEdit` event
   (one ingress → one signal); keybinding `ctrl+alt+d`.
6. **Spec written** — `coauthoring-diff-view.md` (F6), the full Solution
   Design form (handbook §3.3, F4/F5 precedent): BUC/PUC, INV-18/19,
   architecture (BaselineStore + DiffViewController + seam event),
   alternatives table, testing strategy (unit + host E2E, no LLM), failure
   modes (stale-epoch states are visible and pin-recoverable), 4-slice
   delivery plan, traceability to #17's acceptance.
7. **Self-review** — placeholder/consistency/scope/ambiguity pass; one fix
   (named `renderIfOpen` as the `ensureBaseline` hook).
8. **Finalize** — restored the playground stray; submitted the spec
   `--status graduated` (autonomous default) → content commit `36bc4e3`;
   synced the local content clone; memory updated
   (`f6-diff-view-spec-graduated` + index; cleared the stale Next /goal in
   `f5-cross-rung-shipped`); published this transcript.

No wrong turns of substance this session; the one mid-course correction was
realizing publication of the spec belongs to finalize's `submit-spec.sh`
(not a hand commit/PR), and authoring the spec directly in the content
repo's working tree as its canonical home.

## Deferred decisions

_Autonomous-mode low-confidence calls the driver made and would have
liked operator input on. Surfaced at finalize._

- **Absorb-on-advance semantics** (spec §6.7): the baseline advances to a
  whole-document snapshot at each machine landing, so operator edits made
  *before* an accept disappear from the diff ("my changes since the machine
  last landed", not "my whole session"). The alternative — splicing the
  accepted replacement into the baseline so older operator edits stay
  visible — was deferred as an additive refinement (murky edge cases when
  the machine edits operator-authored-since-baseline text). Confident in
  the invariant (INV-18 holds either way); less confident the absorb
  semantics match operator intuition — pin + the epoch-naming diff title
  are the mitigations.
- **Default keybinding `ctrl+alt+d`** for `cowriting.toggleDiffView`
  (spec §6.7): chosen over `cmd+alt+d` (macOS Dock toggle) and `cmd+k d`
  (taken by compare-with-saved); cheap to change.
- **Spec authored directly in the content repo working tree**
  (`vscode-cowriting-plugin-content/specs/coauthoring-diff-view.md`) — its
  canonical home — rather than a scratch `specs/` dir in the plugin repo;
  finalize's `submit-spec.sh` publishes from that path.
- **Restored `sandbox/playground.md`** at finalize (discarded the 1-line
  pre-session truncation, restoring the committed text) — judged stray
  damage from manual play, not work; flagged here since discarding an
  operator working-tree edit is not reversible.

## Cut state (end of session)

| Repo | State |
| --- | --- |
| `vscode-cowriting-plugin` | `main` @ `611a180`, clean, in sync (playground stray restored; no code changes this session) |
| `vscode-cowriting-plugin-content` | `main` @ `36bc4e3` — **specs/coauthoring-diff-view.md added (status: graduated)**; local clone synced; 4 untracked `issues/*.md` capture drafts remain (Author's to publish) |
| session-history (`sessions/0017/`) | this transcript, published at finalize |

- Issue #17: open, unchanged — the spec is its design artifact; the issue
  closes when F6 ships.
- Nothing implemented yet: F6 is spec-only as of this session.

## What lands on the operator's plate

- **Review the graduated spec** (autonomous mode skipped the review gate):
  `/Users/benstull/git/benstull.org/benstull/vscode-cowriting-plugin-content/specs/coauthoring-diff-view.md`
  — especially the two deferred calls above (absorb-on-advance semantics;
  `ctrl+alt+d`).
- The four untracked `issues/*.md` drafts in the content repo are still
  yours to commit/publish or discard (capture-session convention).

## Prompt the operator can paste into the next session

```
/goal Plan and execute Feature #17 (F6 diff-view toggle) from the graduated Solution Design at content specs/coauthoring-diff-view.md — 4 slices per §7.2, host-E2E tier, no LLM in CI
```