refactor(cli)!: one shared base config with per-surface overlays

`dsh` shipped two config trees that were 43 rows the same: apps/cli/cordis.yml
composed web as 74 flat rows, while the TUI booted examples/tui-agent/cordis.yml
whose single `@deepseek-ai/dsh-tui-demo` row mounted twelve plugins behind a
twenty-key pass-through Config. Neither file was what its location claimed —
apps/cli hardcoded the "example" as the product default and the "demo" bundle
was the application — and every capability change had to be made twice.

- apps/cli/base.cordis.yml holds the 43 shared rows; tui.cordis.yml and
  web.cordis.yml are patch lists stating only what differs per surface
- overlays apply as SIBLING patch lists at one include level, because include
  patches never cross an include boundary. Precedence: base < surface <
  (--config | personal ~/.dsh/config.yaml) < launcher flag/profile patches
- `--config` now applies an overlay INSTEAD OF the personal one, so a demo or
  test tree never inherits the user's route; new `--config-replace` boots a file
  as the entire tree (the old `--config` behaviour). Both survive /resume
- vendor/include: index each `insert`ed row as it is added so a later patch can
  configure or disable it. Upstream built the id index once before the patch
  loop, leaving every surface-only row — the whole TUI front door — silently
  unpatchable from user config. Logged as local modification 8
- session identity moves to dsh-agent-loop's CONFIGURED_AGENT_IDENTITIES_KEY;
  dsh-tui's MAIN_SESSION_ID_KEY is deleted (only the bundle read it)
- delete examples/tui-agent, examples/cordis-agent, packages/examples/tui-demo;
  TUI tests → apps/cli/tests, cordis e2e → packages/cordis/tool-cordis/tests,
  examples/code-mode survives as an overlay leaf
- `dsh web` gains --config, threaded into AppCLIEntry as an extra overlay

Three latent defects surfaced and are fixed here: the TUI captured the optional
sessionQuery service once at construction and could permanently disable /resume
when it won the mount race; the session-store root silently reverted to a
project-local ./.sessions; --config-replace was dropped by the resume handoff.

Verified by booting each tree through the real Loader (TUI 55 entries, web 75,
zero unsettled) rather than reading YAML. All eight terminal snapshots replay
byte-identically; 14/14 PTY smoke, 112/112 snapshots, 25/25 doc-sync, hygiene
and lint clean.
This commit is contained in:
Turtle
2026-07-29 13:58:21 +08:00
parent e7c0a5b794
commit f290a8b851
182 changed files with 1659 additions and 2400 deletions
@@ -0,0 +1,53 @@
# Agent Note: One shared base config with per-surface overlays
Status: implemented
English | [中文](2026-07-29-shared-base-config-overlays.zh.md)
## Problem
`dsh` shipped two full config trees that were 43 rows the same. `apps/cli/cordis.yml` composed the web surface as 74 flat rows, while the TUI booted `examples/tui-agent/cordis.yml`, whose single `@deepseek-ai/dsh-tui-demo` row mounted twelve plugins and re-declared their configuration as its own twenty-key pass-through `Config`.
Neither file was what its location claimed. `examples/tui-agent` was not an example: `apps/cli/src/tui.ts` hardcoded it as the product's default config, and it owned the TUI PTY smoke, the eight terminal snapshot scenarios, and the PTY harness the `cordis-agent` leaf imported. `dsh-tui-demo` was not a demo either — it was the application, mounted by the shipped binary from `packages/examples/`.
The duplication was the load-bearing problem. Of the 43 shared rows, 38 were byte-identical and 5 differed for a defensible per-surface reason, so every capability change had to be made twice and could silently drift. The bundle also inverted a default: `composeTuiApp` read `config.goals ?? {}`, so the shipped TUI mounted goals, `tool-goal`, `goal-session`, and `/goal` although no config key requested them.
## Decision
One shared base, one overlay per surface, composed as sibling patch lists.
`apps/cli/base.cordis.yml` holds the 43 rows both surfaces mount. `apps/cli/tui.cordis.yml` and `apps/cli/web.cordis.yml` are **patch lists**, not trees: each states the handful of rows whose value is surface-specific and inserts its own rows. The launcher includes the base once and applies every overlay as a sibling patch list at **one** include level, because include patches never cross an include boundary — stacking overlays as nested includes would silently stop reaching base rows.
Precedence is list order, last write winning per row: base, then the surface overlay, then either a `--config` overlay or the personal `~/.dsh/config.yaml`, then the launcher's own flag and profile patches.
`--config <path>` now applies an overlay **instead of** the personal overlay, so a demo or test tree never inherits the user's provider and model. `--config-replace <path>` boots a file as the entire tree, bypassing base, surface overlay, and personal overlay alike; that is what the old `--config` did, so trees like `examples/web-cordis` moved to the new flag. Both flags survive the `/resume` execve handoff, or resuming would silently change the agent.
A patch replaces its target row's whole `config` rather than merging, which shapes the split: a row whose value differs per surface lives in the overlays, never in the base, so no row is patched by three layers at once. Session identity therefore cannot ride a config key at all — it moved to `dsh-agent-loop`'s `CONFIGURED_AGENT_IDENTITIES_KEY`, as [the launcher-owned identity note](../architecture/2026-07-28-launcher-owned-resume-identity.md) now records.
`examples/tui-agent`, `examples/cordis-agent`, and `packages/examples/tui-demo` are deleted. The TUI tests move to `apps/cli/tests/`, the cordis-toolset e2e to `packages/cordis/tool-cordis/tests/`, and `examples/code-mode` survives as a genuine example: an overlay that patches `tools.mode` and inserts the code runtime.
## Alternatives considered
**Leave both trees flat and duplicated.** Rejected: 43 rows maintained twice is the defect, and a gate asserting they stay identical would freeze the duplication rather than remove it.
**Nest the overlays as includes (`code-mode` → `tui` → `base`).** Rejected after testing the Loader: patches do not cross an include boundary, so the outer file's patches are dropped with only a warning. A three-level chain left `tools` unpatchable, and a base behind one include made every personal patch a silent no-op.
**Put the union of all rows in the base and have each overlay disable what it does not want.** Rejected: the base stops meaning "shared", and each surface carries rows it exists only to switch off.
**Keep the per-surface rows in the base and let overlays patch them.** Adopted only for the five rows that must exist in both trees, because a patch cannot create a row. Their base entries carry the plugin name and the config both surfaces share; each overlay states the rest.
## Consequences
An overlay or `--config` tree that named `@deepseek-ai/dsh-tui-demo`, or patched the `tui-agent` row, no longer resolves. Overlays now patch the row that owns each key: the model route on `agent-loop`, the persona on `system-prompt`, presentation on `tui`.
A patch whose `id` matches no row stays a Loader warning rather than an error. That is deliberate: one personal overlay is shared across surfaces, and `insert` rows match nothing by design, so a row that exists only under `web` must not fail the TUI's boot.
`dsh web` gains `--config`, threaded into `AppCLIEntry` as an extra overlay. `AppCLIEntry` reads both the base and its surface overlay when recovering row defaults for its own patch merge, since a flag override must preserve the overlay's other fields on the same row.
## Verification
Composition is checked by booting each tree through the real Loader and inspecting settled entries, not by reading YAML: the TUI settles 55 entries and web 75, both with zero unloaded or unsettled rows, and web's `httpServer` up. The three-layer case (`base` + `tui` + `code-mode`) confirms `tools.mode` reaching `code` over the TUI overlay's `native`.
All eight terminal snapshot scenarios replay byte-identically after moving, and the 14-case PTY smoke passes, including two cases that assert a personal overlay reaches an **inserted** row — the behavior the vendored `plugin-include` fix enables ([`vendor/README.md`](../../../../vendor/README.md) local modification 8, covered by `packages/ui/app-boot/tests/config-reload.spec.ts`).
Flattening surfaced three latent defects, each fixed here: the TUI captured the optional `sessionQuery` service once at construction and so could permanently disable `/resume` when it won the mount race; the shipped session-store root silently reverted to a project-local `./.sessions`; and `--config-replace` was dropped by the resume handoff.