fix(prompt): identity guard against foreign agent-platform workspaces#555
Merged
Conversation
…#550) When the sandbox root contained another product's data dir (Hermes Agent's SOUL.md, skills/, memories/, etc.), the model would browse those files and infer it was a sub-profile or runtime layered on top of the host product. The system prompt defines the assistant as a standalone coding assistant; the workspace contents shouldn't override that. Two layers: - Top-of-prompt identity section names the failure mode explicitly: workspace files describe the user's project, never what Reasonix is. Calls out SOUL.md / AGENT.md / PERSONA.md / foreign skills+memories trees / a foreign REASONIX.md as not-the-spec, and tells the model to answer "who are you?" from the prompt instead of running ls. - detectForeignAgentPlatform(rootDir) flags the workspace at launch if one of those markers sits at the root, and the code-mode banner adds a one-line warning suggesting --dir <real-project>. Conservative signal: lone skills/ doesn't trigger (common in coding repos); skills/ + memories/ pair does. Closes #550
This was referenced May 10, 2026
Closed
Merged
ChasLui
pushed a commit
to ChasLui/DeepSeek-Reasonix
that referenced
this pull request
May 23, 2026
…esengine#550) (esengine#555) When the sandbox root contained another product's data dir (Hermes Agent's SOUL.md, skills/, memories/, etc.), the model would browse those files and infer it was a sub-profile or runtime layered on top of the host product. The system prompt defines the assistant as a standalone coding assistant; the workspace contents shouldn't override that. Two layers: - Top-of-prompt identity section names the failure mode explicitly: workspace files describe the user's project, never what Reasonix is. Calls out SOUL.md / AGENT.md / PERSONA.md / foreign skills+memories trees / a foreign REASONIX.md as not-the-spec, and tells the model to answer "who are you?" from the prompt instead of running ls. - detectForeignAgentPlatform(rootDir) flags the workspace at launch if one of those markers sits at the root, and the code-mode banner adds a one-line warning suggesting --dir <real-project>. Conservative signal: lone skills/ doesn't trigger (common in coding repos); skills/ + memories/ pair does. Closes esengine#550
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
A user reported the model claiming a false architectural relationship
to whatever AI platform's data dir happened to sit at the workspace
root — e.g. "the underlying runtime is Hermes Agent" when launched
against a directory containing `SOUL.md`, `skills/`, `memories/`,
or a foreign `REASONIX.md`. The assistant's identity should come from
the prompt, not from `ls`.
Root cause
Two contamination vectors:
with "You are Reasonix Code, a coding assistant" but never says
the workspace is the user's project, not a spec for what
Reasonix is. When the model gets asked "who are you?" and runs
`directory_tree`, it pattern-matches whatever it finds ("oh,
there's a SOUL.md and a skills/ — I must be a sub-profile of this
platform") and confidently invents a relationship.
itself at any directory the user passes, including data dirs that
were clearly never meant to be code projects.
Fix
Two layers — (1) is the load-bearing one, (2) is a UX nudge:
the failure mode explicitly. Lists the marker files that describe
somebody else's runtime, not yours: `config.yaml` with agent /
persona keys, `SOUL.md`, `AGENT.md`, `PERSONA.md`, foreign
`skills/` / `memories/` trees, a foreign `REASONIX.md`. Tells
the model to answer identity questions from the prompt — don't run
`ls` to figure out what it is.
`detectForeignAgentPlatform(rootDir)` returns the marker(s) that
flag a foreign workspace. `code.tsx`'s startup banner adds a one-
line warning suggesting `--dir ` when any fire.
Conservative signal: lone `skills/` doesn't trigger (it's common
in coding repos); a `skills/` + `memories/` pair does.
Test plan
`code-prompt.test.ts`, 6 in `project-memory.test.ts`)
confirm the warning fires, then ask "who are you?" — expect a
Reasonix Code answer, no Hermes / SOUL references
Closes #550