feat(web): move the agent plane behind per-session presets

The Web overlay disables base's 32 agent-plane rows and mounts the preset roster instead, so each session composes its own tools and prompt rather than sharing one process-wide set. The TUI keeps base unchanged: it is single-session and composing its agent process-wide is correct there. `roots` is patched in by AppCLIEntry, like `distIndex`: the shipped presets sit beside the composition that names them and the user's live under the Harness home, neither of which a config author chooses. A session's preset is fixed at creation. Naming a different one for an existing identity is `agent-preset-conflict` rather than a switch, because that session's history was produced under the first preset's tools. The guard sits after `await creation`, beside the cwd check, so it covers every path that yields a live agent — freshly created, adopted live, resumed, or recovered by the concurrent-creation catch. A request naming no preset adopts the session as it is, keeping reconnect and retry ordinary. Two bugs the real-composition test caught, both invisible to unit tests: `PresetTree` now refuses to write. The Loader persists a tree whose plugin self-disposed, and tearing an agent down disposes its whole subtree — inherited, that rewrote the shipped composition, truncating a 241-line preset to `[]` the first time a session ended. `dsh-tool-skill` compared against a lookup of its own name in the global layer, so it threw inside any preset: `register()` files into the calling context's scope. It now compares against the definition it registered, which is what the identity check meant all along. The `standard` catalog is asserted exactly, not spot-checked: a row that registers into the wrong layer mounts cleanly and simply contributes nothing, so an omission is this design's quietest failure. It matches the shipped TUI catalog plus `glob`/`grep`, the pair that composition documents as ripgrep-dependent. Re-records `cordis-inspect-jsdoc`, whose rendered `SessionHeader` gains the `agentPreset` field. `fs-glob-sampling` fails identically on pristine master and is untouched here. The browser e2e scaffold gains the roster fact AppCLIEntry supplies. `roots` is resolved and patched in by the CLI entry, like `distIndex` on the webserver row, and this lane boots the shipped tree without that entry — so it has to supply the same fact or the roster resolves nothing and every session in the lane composes an agent with no tools, no persona, and no token meter. Only the shipped root: a developer's own `~/.dsh/.agent-presets` must not decide a golden. The `cordis:group` builtin comes with it, exactly as `boot()` registers it, because a preset resolving package names from its own directory cannot reach `@cordisjs/plugin-group` by name. The lane stays red through this layer and the next four for the reason stated above — the api-proxy injects `subagents`, `workspace`, and `tools`, so `api-gateway` cannot activate and the browser has no `/api` at all. It goes green again in the layer that returns those registries to the host plane; this change is what makes that layer's fix sufficient rather than partial.
2026-08-03 23:44:09 +08:00
parent 91b55b9245
commit 3d68185480
15 changed files with 564 additions and 11 deletions
@@ -12,7 +12,7 @@ At every eligible `agent/pre-step`, the plugin calls `ctx.skills.snapshot()` for

 Every catalog message carries the `skill-catalog` source: a `catalog`-form context whose `entries` record exactly the `name` and `description` pairs it published, plus `update` on a replacement. The digest covers those durable entries, not the rendered prose, so the surrounding `<system-reminder>` framing cannot decide whether a republish is needed and consumers never re-parse the `<available_skills>` block. The plugin scans durable session events backwards without copying them and derives the comparison baseline from the newest visible `skill-catalog` message it can read; unreadable and foreign records are skipped. When the digest changes, the downstream `enter` decision receives a durable user-role message containing the complete replacement catalog; an empty replacement explicitly retires earlier names. If no catalog remains visible but a recognizable historical catalog exists, compaction hid it and the next complete observation re-establishes the current catalog. An incomplete provider snapshot emits nothing and preserves the last-good model view for retry at the next pre-step. If no prior catalog exists and the current view is empty, no tombstone is necessary.

-The catalog is omitted when no model-invocable skills are initially available, and also when that agent's tool view restricts away the shipped `skill` tool or resolves a same-name scoped shadow instead. Visibility changes participate in the digest, keeping prompt guidance, model-visible schema, and executable dispatch aligned.
+The catalog is omitted when no model-invocable skills are initially available, and also when that agent's tool view restricts away the shipped `skill` tool or resolves a same-name scoped shadow instead. Identity is compared against the definition this plugin registered rather than a lookup of its own name, so the plugin works mounted globally or inside one agent's composition, where `register()` files into that agent's layer alone. Visibility changes participate in the digest, keeping prompt guidance, model-visible schema, and executable dispatch aligned.

 `catalogDescriptionMaxLength` controls normalized catalog descriptions; rendering XML-escapes them. Its default is `500` and values must be integers of at least `3`, which reserves room for a truncation ellipsis. The [skill catalog hot-refresh Agent Note](../../../.agents/notes/implemented/feature/2026-07-27-skill-catalog-hot-refresh.md) owns the durable initial catalog and replacement lifecycle.