omp-fabric 1.18.3 → 1.19.0

package/docs/configuration.md

@@ -275,7 +275,9 @@ OMP core calls reject when the native tool reports an error. Successful `bash`,
 
 `fullCodeMode: true` is the default. Fabric takes OMP's native core tools out of the model-facing set and exposes their implementations through `fabric_exec`; `omp.read()` and the other `omp.*` core calls route through the host adapter.
 
-Fabric records which native core tools were active before it takes ownership. Switching to orchestration-only mode or unloading Fabric restores that selection. Fabric applies full-mode ownership only when the session initializes or the mode changes. It never resets an explicitly selected active tool set from input, agent-start, turn-end, or settled lifecycle hooks. The system prompt carries the full-mode execution rule.
+Fabric reaches only the core tools OMP itself has active. A tool disabled in OMP is unavailable through `omp.*` as well: discovery omits it and a call to it fails. `capture.excludeTools` is a different mechanism. It hides a tool from the model and leaves it callable through `omp.*`. Turning a core tool off while full code mode is already active applies to the next session: Fabric has removed that name from the active set, so the change is invisible to it and the tool stays reachable through `omp.*` until the session ends.
+
+Fabric records which native core tools were active before it takes ownership. Switching to orchestration-only mode or unloading Fabric restores that selection. A core tool re-enabled while full code mode is active stays hidden for that session and is not added to the restored selection. Fabric applies full-mode ownership only when the session initializes or the mode changes. It never resets an explicitly selected active tool set from input, agent-start, turn-end, or settled lifecycle hooks. The system prompt carries the full-mode execution rule.
 
 OMP core shows its model-visible skill catalog only while the native `read` tool is active. Full code mode restores the same catalog from OMP's structured skill registry and changes only the loader instruction, so `omp.read` runs inside `fabric_exec`. Native core tools stay hidden. Packaged skills mark cross-document paths with `<skill-dir>`. Fabric replaces that marker inline from OMP's expanded skill `location` or the actual `SKILL.md` read path. It never matches skill names or enumerates directories. Ordinary document reads stay unchanged. When an expanded skill invokes another installed skill, Fabric adds an exact name-to-path resolution hint for that turn, and the delegated `SKILL.md` loads before task work.
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "omp-fabric",
-  "version": "1.18.3",
+  "version": "1.19.0",
   "description": "A programmable tool and agent runtime for Oh My Pi (OMP)",
   "type": "module",
   "main": "./dist/index.js",

package/skills/fabric-advisor/SKILL.md

@@ -17,7 +17,7 @@ Hard pointer: read `<skill-dir>/../fabric-ambient/references/setup.md` completel
 - `strings.instructions`: the prompt below, with the requested focus appended
 
 ```text
-You are an ambient peer advisor for the main coding agent. Review the supplied parent-session event and recent transcript as an outside observer, not a second executor. Focus on correctness, missed user constraints, risky assumptions, edge cases, and cheaper paths to the requested outcome. Inspect the workspace with read-only tools only when evidence is needed.
+You are an ambient peer advisor for the main coding agent. Review the supplied parent-session event and recent transcript as an outside observer, not a second executor. Focus on correctness, missed user constraints, risky assumptions, edge cases, and cheaper paths to the requested outcome. Inspect the workspace with read-only tools only when evidence is needed. Never prescribe a tool to the main agent without transcript evidence that it is callable there. Under full code mode the main agent by default reaches OMP's file tools only as `omp.edit` / `omp.write` inside `fabric_exec`, so an escaping failure in a shell or Node script argues for `omp.edit`, never against `fabric_exec`.
 
 Prefer silence. Return {"action":"silent"} when work is on track. Return {"action":"message","message":"..."} only for one concrete, material observation that could prevent wasted work or a defect while there is still time to act. Cite the evidence and recommendation tersely as advice, not an order. Do not repeat advice visible in the transcript or raise minor style preferences unless the user required them.
 ```

package/skills/fabric-ambient/SKILL.md

@@ -34,7 +34,7 @@ Review the supplied event and recent transcript as an outside observer. Return {
 Use `name=advisor`, `events=["turn_end"]`, and `triggerTurn=false`. Append any requested focus to:
 
 ```text
-You are an ambient peer advisor reviewing the main coding agent. Focus on correctness, missed constraints, risky assumptions, and cheaper paths. Inspect with read-only tools only when needed. Return {"action":"silent"} when work is on track; otherwise return {"action":"message","message":"..."} for one concrete, material observation. Cite evidence and a terse recommendation as advice, not an order. Do not repeat visible advice.
+You are an ambient peer advisor reviewing the main coding agent. Focus on correctness, missed constraints, risky assumptions, and cheaper paths. Inspect with read-only tools only when needed. Never prescribe a tool without transcript evidence it is callable in the main session; under full code mode its file tools are by default `omp.edit`/`omp.write` inside `fabric_exec`. Return {"action":"silent"} when work is on track; otherwise return {"action":"message","message":"..."} for one concrete, material observation. Cite evidence and a terse recommendation as advice, not an order. Do not repeat visible advice.
 ```
 
 The supervisor can wake an idle session; the per-turn advisor must not.