refactor: delete dead codex-core subsystems and inert picker pagination

.github/workflows/rust-ci.yml

@@ -127,7 +127,9 @@ jobs:
           MOCK_ACP_AGENT_BIN: ${{ github.workspace }}/nori-rs/target/mock-acp-out/mock_acp_agent
         run: |
           cargo test --profile ci-test --target ${{ matrix.target }} --package tui-pty-e2e
-          cargo test --profile ci-test --target ${{ matrix.target }} --package nori-acp
+          cargo test --profile ci-test --target ${{ matrix.target }} --package nori-harness
+          cargo test --profile ci-test --target ${{ matrix.target }} --package nori-acp-host
+          cargo test --profile ci-test --target ${{ matrix.target }} --package nori-config
           cargo test --profile ci-test --target ${{ matrix.target }} --package nori-tui
           cargo test --profile ci-test --target ${{ matrix.target }} --package nori-cli
           cargo test --profile ci-test --target ${{ matrix.target }} --package codex-protocol

docs.md

@@ -31,7 +31,7 @@ The project was originally forked from OpenAI Codex CLI and has been adapted to
 │                  nori-tui                       │
 │        Interactive Terminal Interface           │
 ├────────────────────────┬────────────────────────┤
-│     nori-acp (acp/)   │   codex-core (core/)   │
+│ nori-harness (harness/)│  codex-core (core/)    │
 │  ACP Agent Connection  │  Config, Auth, Tools   │
 │  Subprocess Spawning   │  Sandbox, Utilities    │
 ├────────────────────────┴────────────────────────┤
@@ -95,9 +95,8 @@ Nori acts as an MCP client:
 
 ### Things to Know
 
-- Nori-authored crates use a `nori-` prefix (`nori-cli`, `nori-tui`, `nori-acp`, `nori-protocol`, `nori-installed`); inherited crates keep the legacy `codex-` prefix from the OpenAI Codex fork and are progressively adopted or removed per `docs/specs/crate-layering.md`
-- The `nori-config` feature flag enables Nori-specific configuration paths (`~/.nori/cli/`) instead of the legacy Codex paths (`~/.codex/`)
-- The `unstable` feature flag gates experimental ACP features like model switching
+- Nori-authored crates use a `nori-` prefix (`nori-cli`, `nori-tui`, `nori-acp`, `nori-acp-host`, `nori-config`, `nori-protocol`, `nori-installed`); inherited crates keep the legacy `codex-` prefix from the OpenAI Codex fork and are progressively adopted or removed per `docs/specs/crate-layering.md`
+- Configuration always uses the Nori paths (`~/.nori/cli/`); the old `nori-config` and `unstable` cargo feature flags were removed during the crate-layering cleanup (no cargo feature may change which crate owns a responsibility)
 - Cross-platform sandboxing is implemented using Landlock (Linux), Seatbelt (macOS), and restricted tokens (Windows)
 - Snapshot testing with `insta` is used extensively for TUI regression testing
 - The project has two justfiles: a root `@/justfile` implementing the Shared Local Runner Layer spec (standardized `help`, `dev`, `test`, `doctor` targets) and `@/nori-rs/justfile` for Rust-specific workflows. The root justfile wraps `nori-rs` by running `cd nori-rs && cargo ...` for each target. Both coexist -- run `just` from the repo root for the standard targets, or `cd nori-rs && just` for the Rust-native recipes

docs/reference/transcript-format.md

@@ -0,0 +1,189 @@
+# Nori transcript format
+
+Reference for Nori's on-disk session transcript format, for tools that want to
+read or write Nori transcripts without depending on the `nori-harness` crate.
+Transcripts capture the client-side view of a conversation: what the user
+typed, what the assistant responded, and which tools ran.
+
+Canonical implementation: `nori-rs/harness/src/transcript/` (crate
+`nori-harness`) — `types.rs` (schema), `recorder.rs` (writer), `loader.rs`
+(reader), `project.rs` (project id derivation).
+
+## File locations
+
+Transcripts live under the Nori home directory:
+
+- `$NORI_HOME` if set, otherwise `~/.nori/cli` (see `nori-config`).
+
+Layout:
+
+```text
+$NORI_HOME/transcripts/by-project/{project-id}/
+  ├── project.json            # Project metadata
+  └── sessions/
+      └── {session-id}.jsonl  # One file per session
+```
+
+- `{session-id}` is a UUIDv4 generated at session start.
+- Session files are created with mode `0600` on Unix.
+- The file is created fresh (truncated) when the session starts, then
+  appended to line by line. Each line is flushed after writing, but the file
+  is never fsynced, so the tail of a crashed session may be missing.
+
+### Project ids and `project.json`
+
+The project id is 16 lowercase hex characters derived by hashing, in priority
+order:
+
+1. the normalized git remote URL of `origin` (e.g.
+   `git@github.com:user/repo.git` and `https://github.com/user/repo.git` both
+   normalize to `github.com/user/repo`), else
+2. the git root absolute path, else
+3. the canonicalized working directory path.
+
+The hash is Rust's `DefaultHasher`, which is **not guaranteed stable across
+Rust versions**. Treat project ids as opaque directory names: discover
+projects by listing `by-project/` and reading each `project.json`, and do not
+try to recompute ids yourself.
+
+`project.json` (pretty-printed JSON, rewritten on every new session in the
+project):
+
+```json
+{
+  "id": "a1b2c3d4e5f60718",
+  "name": "my-repo",
+  "git_remote": "git@github.com:user/my-repo.git",
+  "git_root": "/home/user/src/my-repo",
+  "cwd": "/home/user/src/my-repo",
+  "created_at": "2026-07-03T12:30:45.123Z",
+  "updated_at": "2026-07-03T12:30:45.123Z"
+}
+```
+
+`git_remote` and `git_root` are `null` outside a git repo. Because the file is
+rewritten each session, `created_at` reflects the most recent session, not the
+first.
+
+## Line format
+
+Session files are JSONL: one self-contained JSON object per line, `\n`
+terminated. Every line shares an envelope, with the entry's fields flattened
+into the same object:
+
+- `ts` — ISO 8601 timestamp with millisecond precision, UTC (`Z` suffix),
+  e.g. `"2026-07-03T12:30:45.123Z"`.
+- `v` — schema version, currently `2`.
+- `type` — the entry kind (snake_case), plus the entry's own fields.
+
+The **first line must be a `session_meta` entry**; readers reject files whose
+first line is not valid session metadata.
+
+### `session_meta` (first line)
+
+```json
+{"ts":"2026-07-03T12:30:45.123Z","v":2,"type":"session_meta","session_id":"7f9c2f6a-1c1e-4a9b-9a3e-2f0d8b7c6d5e","project_id":"a1b2c3d4e5f60718","started_at":"2026-07-03T12:30:45.123Z","cwd":"/home/user/src/my-repo","agent":"claude-code","cli_version":"0.9.0","git":{"branch":"main","commit_hash":"1975265abc..."},"acp_session_id":"acp-sess-abc123"}
+```
+
+Optional fields, omitted when absent: `agent` (ACP agent slug, e.g.
+`claude-code`, `codex`, `gemini`), `git` (and within it `branch`,
+`commit_hash`), and `acp_session_id` (the ACP agent's own session id, used to
+resume via `session/load`).
+
+### `user`
+
+```json
+{"ts":"2026-07-03T12:31:02.001Z","v":2,"type":"user","id":"msg-001","content":"What files are in src?","attachments":[{"type":"file_path","path":"/tmp/screenshot.png"}]}
+```
+
+`attachments` is omitted when empty. Each attachment is tagged by `type`:
+`file_path` (`path`) or `base64` (`data`, `mime_type`).
+
+### `assistant`
+
+```json
+{"ts":"2026-07-03T12:31:10.500Z","v":2,"type":"assistant","id":"msg-002","content":[{"type":"thinking","thinking":"Let me check src."},{"type":"text","text":"src contains main.rs and lib.rs."}],"agent":"claude-code"}
+```
+
+`content` is a list of blocks tagged by `type`: `text` (`text`) or `thinking`
+(`thinking`). `agent` is optional.
+
+### `tool_call` / `tool_result`
+
+```json
+{"ts":"2026-07-03T12:31:05.000Z","v":2,"type":"tool_call","call_id":"call-001","name":"shell","input":{"command":"ls src/"}}
+{"ts":"2026-07-03T12:31:05.250Z","v":2,"type":"tool_result","call_id":"call-001","output":"main.rs\nlib.rs","exit_code":0}
+```
+
+`input` is arbitrary JSON. On results, `truncated` (bool) is omitted when
+false and `exit_code` is omitted when not applicable. `call_id` correlates
+call and result.
+
+### `patch_apply`
+
+```json
+{"ts":"2026-07-03T12:31:20.000Z","v":2,"type":"patch_apply","call_id":"call-002","operation":"edit","path":"/home/user/src/my-repo/src/main.rs","success":true}
+```
+
+`operation` is `edit`, `write`, or `delete`. `error` (string) is present only
+on failure.
+
+### `client_event`
+
+Wraps a normalized ACP-native event from the `nori-protocol` crate. The
+payload lives under `event` and is internally tagged by `event_type`
+(snake_case), with the variant's fields flattened alongside the tag:
+
+```json
+{"ts":"2026-07-03T12:31:06.000Z","v":2,"type":"client_event","event":{"event_type":"tool_snapshot","call_id":"call-001","title":"Edit src/main.rs","kind":"edit","phase":"completed","locations":[],"invocation":null,"artifacts":[],"raw_input":null,"raw_output":null}}
+```
+
+Variants (see `nori-rs/nori-protocol/src/lib.rs` for payload shapes):
+`tool_snapshot`, `approval_request`, `message_delta`, `plan_snapshot`,
+`session_phase_changed`, `prompt_completed`, `load_completed`,
+`queue_changed`, `context_compacted`, `replay_entry`,
+`agent_commands_update`, `session_capabilities_changed`,
+`session_update_info`, `session_config_update`, `session_mode_changed`,
+`thread_goal_updated`, `thread_goal_cleared`, `warning`. This set tracks the
+protocol crate and changes more often than the core entry types; readers
+should be prepared to skip unknown `event_type`s.
+
+## Token usage
+
+Nori transcript entries do not carry token counts. The only usage data that
+can appear is the optional `usage` field inside `session_update_info` client
+events. The token statistics shown in the TUI footer are parsed from the
+underlying agent's *own* transcript files (`~/.claude/projects/`,
+`~/.codex/sessions/`, `~/.gemini/tmp/`) by
+`nori-rs/harness/src/transcript_discovery.rs`; those are third-party formats
+and out of scope here.
+
+## Versioning and compatibility
+
+- `v` is currently `2`. The loader does not branch on it; it exists for
+  forward compatibility.
+- **Unknown fields are ignored** (no `deny_unknown_fields` anywhere in the
+  schema).
+- **Unknown or unparseable lines are silently skipped** by the full-transcript
+  loader — except the first line, which must parse as `session_meta` or the
+  load fails.
+- Blank lines are skipped.
+- Lightweight scanners (session pickers, first-user-message preview, user-turn
+  counts) inspect only the `type` field per line, so extra entry kinds never
+  break listing.
+
+Writer expectations for third-party producers:
+
+- One JSON object per line, no pretty-printing, `\n` terminated.
+- First line is `session_meta` with at minimum `session_id`, `project_id`,
+  `started_at`, `cwd`, `cli_version` (plus the `ts`/`v`/`type` envelope).
+- Name the file `{session_id}.jsonl` and keep `session_id` consistent between
+  the filename and the metadata — loaders derive paths from the id.
+
+## Reading transcripts programmatically
+
+If you can take a Rust dependency, use the `nori-harness` crate
+(`nori-rs/harness`): `TranscriptLine` / `TranscriptEntry` are the canonical
+serde types, `TranscriptLoader` handles discovery, listing, and
+tolerant parsing, and `TranscriptRecorder` is the canonical writer. Anything
+this document says is a simplification of those types.