@@ -0,0 +1,60 @@
+# Conversation fidelity
+
+What Modex shows about a coding turn, where each fact comes from, and what it does not claim. It is
+the user-facing summary of tracker [#273](https://github.com/TypeSafeAI/modex/issues/273). The
+contracts behind it are [conversation events](conversation-events.md),
+[evidence](conversation-evidence.md), [task plans and records](task-context.md) and the
+[event journal](event-journal.md). The acceptance matrix and its open gates are in
+[the acceptance receipt](reviews/2026-10-18-conversation-fidelity-acceptance.md).
+
+Coding still runs only through `codex app-server` and `claude -p`. Nothing described here executes,
+replays or re-sends anything by being shown, reloaded or reconnected.
+
+## Supported
+
+| Feature | Codex (app-server, codex-cli 0.163.0) | Claude (`claude -p`, Claude Code 2.1.302) |
+| --- | --- | --- |
+| Message identity and phase | Native item ids; `commentary` and `final_answer` phases | Message id and block index; no phase (unspecified) |
+| Final answer | The latest attempt's `final` message | The latest attempt's last reply |
+| Reasoning | Summary parts; reasoning text kept apart; empty reasoning is "not shared" | Thinking text when `thinking_display` is `summarized` (summary) or other (provider text); signature-only or redacted is "not shared" |
+| Tool outcome | From native status: succeeded, failed, denied (declined), unknown | Succeeded, failed, or denied when Modex refused it |
+| Exit code | Reported | Not reported |
+| Citations | Memory citations | None; links in the reply are labelled as the model's |
+| Plan | `turn/plan/updated` | `TodoWrite` |
+| Save to Notes | Yes | Yes |
+| Stop | Native interrupt | Process stop (CF-2, [#275](https://github.com/TypeSafeAI/modex/issues/275), refines this) |
+
+## Fallbacks
+
+- **History from before these changes** reads unchanged. It keeps the old last-reply heuristic and
+  gains no invented provenance.
+- **A value a CLI does not report** stays unspecified, unknown or not shared. Modex never guesses
+  one.
+- **Work a dead process left running** becomes unknown everywhere it is shown, not running and not
+  failed, and is not run again.
+- **A phone that predates these fields** shows a tool's outcome in its title, and a plan as a line of
+  text.
+- **Journal off,** the default, keeps the earlier reload behaviour. Journal on replays the same items
+  as the transcript.
+
+## Privacy limits
+
+- **Opaque reasoning** (signatures, redacted thinking) is never read, stored or shown.
+- **Task records** leave out reasoning, command output and paths outside the project.
+- **Streamer Mode** masks what the window shows. It does not sanitize the clipboard, accessibility
+  text, images or arbitrary secrets in transcripts. Copy and Save to Notes write what the transcript
+  holds.
+
+## Migration and rollback
+
+Every new item field is optional and additive. An older build reading a newer transcript ignores the
+fields it does not know and shows the item as before. A plan item shows as text on older phones; an
+older desktop ignores it. Nothing migrates state files in place. The journal stores events
+unchanged, so rolling back to an older build leaves its replay working.
+
+## Reference captures
+
+The CF-4 typed-turn spec captures a typed turn at 900×600 in Graphite, Jev and OpenCoven
+(`MODEX_SCREENSHOTS=<dir>` keeps them). They are test captures from the mock engine and injected
+items, not provider screenshots. No reference images are committed yet; the receipt lists that as
+open.
@@ -0,0 +1,69 @@
+# Conversation fidelity acceptance, 2026-10-18 (preparation)
+
+The acceptance receipt for tracker [#273](https://github.com/TypeSafeAI/modex/issues/273) and its
+gate [#281](https://github.com/TypeSafeAI/modex/issues/281) (CF-8). **The gate is not passed.** This
+receipt records what is verified, how, and every gate still open. A green mock or fixture test is not
+recorded as live-provider or human acceptance.
+
+## Candidate
+
+No integrated candidate is pinned yet: CF-2 (#275, PR #286) and CF-3 (#276) have not landed, and
+several children have parts still open. Each row names the build it was checked on.
+
+## Children
+
+| Child | Landed | State |
+| --- | --- | --- |
+| CF-1 #274: identity, phases, outcomes | #282 → `7780093` | Done; closed |
+| CF-2 #275: Claude Stop and cleanup | PR #286, green, waiting to update over CF-4 | Owned by another session |
+| CF-3 #276: human requests | — | Not started by its owner |
+| CF-4 #277: typed turns, Markdown | #284 → `8e26534` (typed turns); Markdown slice is PR #283 | Partial |
+| CF-5 #278: evidence | #287 → `d3711c6` (first part) | Partial |
+| CF-6 #279: journal and clients | #285 → `fb76aac` (CF-1 half) | Partial; request half waits for CF-3 |
+| CF-7 #280: plans and Notes | PR #289 (first part) | Partial; question cards wait for CF-3 |
+
+## Integration matrix
+
+| Requirement | Evidence | State |
+| --- | --- | --- |
+| Golden path: request → plan → approval → real edit → check → answer → Save to Notes → Open → restart restores, nothing re-runs | `e2e/conversation-fidelity.spec.ts` "golden path", against the real app, main process and `claude -p` adapter with a fake Claude CLI. Asserts file bytes, execution counts `{edits: 1, checksStarted: 1, checksFinished: 1}` before and after restart, plan statuses, outcomes, note revision 1 → 2 on the same page, and identical `thread:items` after restart | Verified with a fixture CLI, not a live provider |
+| A plan or clarifying question answered during the path | — | Open: needs CF-3 |
+| Paired-client reconnect in the path | CF-6 unit tests (journal replay, the phone's full and delta snapshots); `journal-resume.test.ts` | Verified in unit tests; not yet inside the golden fixture |
+| Denial: file unchanged, nothing runs | "a denial leaves the file as it was and runs nothing": bytes unchanged, counts `{}` | Verified (fixture) |
+| Command failure | "a failing check is a failed command": the tool is failed, and the plan step stays in progress | Verified (fixture) |
+| Unknown outcome | CF-1 and CF-6 unit tests: a dropped stream; work left running at a restart | Verified (unit) |
+| Stop while executing | "Stop while a check runs": stopped, the check never reports, nothing retried | Verified (fixture); CF-2 refines Stop |
+| Request withdrawal; simultaneous desktop and phone answers; reconnect after a decision send | — | Open: needs CF-3 |
+| Old snapshot migration | `test/conversation-migration.test.ts` | Verified (unit) |
+| Journal off | The golden fixture runs with the journal off, the default | Verified (fixture) |
+| Expired cursor | Existing `event-journal.test.ts` resume and reload cases | Verified (unit, existing) |
+| Stale Note revision | `task-context.test.ts`: a note edited since is kept and the new record becomes a new note | Verified (unit) |
+| Missing artifact | — | Open: CF-5's owned-artifact actions |
+| Outcomes, reasoning and provenance match recorded events | CF-1 fixture tests on Claude Code 2.1.302 and codex-cli 0.163.0 captures; CF-5 evidence tests | Verified on captures, not live |
+| Themes at 900×600 | `e2e/typed-turns.spec.ts`: Graphite, Jev and OpenCoven, no sideways scroll | Partial: no wider reference viewport, no committed references |
+| Keyboard and reduced motion | Existing `app.spec.ts` accessibility test; the new elements add no animation | Partial; human VoiceOver open |
+| Long transcript and streaming measurement | CF-5: a 3 MiB Codex stream from 584 ms to 76 ms in a unit test | Open: no named-environment measurement |
+| Live Codex and Claude acceptance | Captures only (2026-10-18, Haiku 5.5 and gpt-6-astra, benign tasks) | Open: needs authorized live runs through the app |
+| Human VoiceOver | — | Open: human |
+| Fresh `pnpm build`, `typecheck`, `test`, `test:e2e` on a pinned candidate | Per-PR results in #282, #284, #285, #287, #289 | Open until a candidate is pinned |
+
+## Commands and results for this preparation
+
+On `cf8-fidelity`, stacked on CF-7, macOS (Apple Silicon), Node v24.18.1, pnpm 10.34.5:
+
+- `pnpm run build`: passed.
+- `npx playwright test e2e/conversation-fidelity.spec.ts`: 4 of 4 passed (45 s).
+
+These are acceptance tests of behaviour CF-1 through CF-7 already built, so they had no red phase.
+
+## Next
+
+1. Land CF-2 (#286) and CF-3. Then add the question, withdrawal, simultaneous-answer and
+   reconnect-after-decision variants to the golden fixture.
+2. Pin a candidate SHA and run fresh full checks on it.
+3. Run authorized live Codex and Claude smoke tasks through the app, and record versions, models and
+   limits.
+4. Get human keyboard and VoiceOver acceptance, commit reference captures, and measure a long
+   transcript on a named machine.
+EOF
+cd /Users/buns/Documents/GitHub/TypeSafeAI/modex/.worktrees/cf8-fidelity && git status --short