Workspace instructions (AGENTS.md) from parent/shared repo are never injected in the web app; home patch cannot configure the per-session plugin

[manucuervodj]

Environment

DSH version: 0.1.0-rc.6 (both @deepseek-ai/dsh and @deepseek-ai/dsh-web-app)
Mode: dsh web (web profile, bundles @deepseek-ai/dsh-base + @deepseek-ai/dsh-web-app)
OS: Windows
Agent preset: code (default)
Plugin: @deepseek-ai/dsh-agent-instructions

Summary

In the web app, the workspace-instruction baseline only ever includes AGENTS.md files from the innermost project root down to the session cwd. When repos are nested (e.g. GitHub/ is a repo containing GitHub/Harness/ as another repo), the shared AGENTS.md at the parent repo root is never discovered, because project-root discovery stops at the first directory containing a .git marker walking up from the cwd.

Additionally, the documented configuration seam for this plugin — the home-level patch $DSH_HOME/cordis.patch.yml — is inert in the web app: the host-plane agent-instructions row is disabled by the web-app bundle, and the instance that actually composes session baselines is mounted per-session from the agent preset, which the home patch cannot reach.
Expected behavior

A session whose cwd is C:\Users\manue\Documents\GitHub\Harness should receive a baseline containing both GitHub/AGENTS.md (shared rules) and Harness/AGENTS.md (repo-local rules), since both are ancestors of the cwd.
Actual behavior

The baseline contains only Harness/AGENTS.md. The session log shows a single Instructions from: AGENTS.md section with no budget-omission notice, confirming the parent file was never discovered (not merely truncated).
Root cause

Per-session plugin instance lives in the agent preset, not the host.
    @deepseek-ai/dsh-web-app/cordis.patch.yml disables the host row: - id: agent-instructions / disabled: true (same as other agent-plane rows like tool-subagent, tool-workflow).
    Each agent preset (config/agent-presets/*/agent.cordis.yml) mounts its own agent-instructions row with only maxBytes: 65536, so projectRootMarkers falls back to the default ['.git'].
    $DSH_HOME/cordis.patch.yml patches only the host composition, so configuring the preset's row from the home patch is impossible. The plugin inventory confirms two rows: include:agent-instructions (host, Disabled) and include:agent-presets:agent-instructions (preset, Mounted/Enabled).

Project-root discovery stops at the innermost marker.
    findProjectRoot(cwd, markers) walks up from the cwd and returns the first directory containing a marker (default .git). With cwd = GitHub/Harness, the walk returns GitHub/Harness (it has .git), so ancestorChain(projectRoot, cwd) never includes GitHub/, and GitHub/AGENTS.md is never loaded.

Budget silently drops the broader file (secondary).
    Even with discovery fixed, the default maxBytes: 65536 is smaller than the shared file (68,104 bytes in our case; both files render at 72,035 bytes), and the renderer drops whole broader files before truncating the most-specific one. The budget notice is easy to miss.

Reproduction

Create a parent repo GitHub/ with AGENTS.md and .git, and a nested repo GitHub/Harness/ with its own AGENTS.md and .git.
Start a web session with cwd GitHub/Harness using the default code preset.
Inspect the first request: the baseline contains only Harness/AGENTS.md.

Workaround (verified)

Create a user preset that copies the shipped code preset and overrides the agent-instructions row:
yaml

• id: agent-instructions
  name: '@deepseek-ai/dsh-agent-instructions'
  config:
  maxBytes: 131072
  projectRootMarkers: ['.dsh-root']

Place a .dsh-root marker file at the shared root (GitHub/), and set agent-presets.default: in $DSH_HOME/settings.yaml (hot-reloaded, no restart needed). Sessions then receive both AGENTS.md files.
Suggested fixes

Let the home/user layer configure preset rows. Either allow $DSH_HOME/cordis.patch.yml (or settings) to patch rows inside preset compositions, or have presets inherit a user-configurable overlay for shared rows (e.g. agent-instructions).
Support nested project roots in instruction discovery. Consider loading candidates from every ancestor directory that contains a root marker (root→cwd chain across nested repos), or add an opt-in like includeAncestorRoots: true / a sharedRootMarkers option, so a parent repo's AGENTS.md is not shadowed by the innermost .git.
Reconsider the default budget. maxBytes: 65536 is smaller than many real-world AGENTS.md files; either raise the default, make the omission notice more prominent, or make maxBytes configurable from settings per preset.

Additional notes

The session header records agentPreset at creation; blank-session drafts fix the preset at creation time, so config changes only affect sessions created afterwards (documented behavior, but easy to trip over when testing).
agentPreset.list correctly reports the new default immediately after editing settings.yaml (hot-reload works); only the preset composition itself is read at session mount.

[etzhe]

Related — cross-linking a discussion I just opened: #5333 (#5333) proposes a native auto-load of a user's ~/.dsh/memory/ (+ its MEMORY.md index) alongside the AGENTS.md/CLAUDE.md chain. It's the same "instruction/context injection" concern as here — in #519 it's AGENTS.md (from a parent/shared repo) never being injected in the web app; in #5333 it's that a user's markdown memory dir is never auto-loaded at all. Cross-linking in case they're related.

[manucuervodj]

Update: Workaround working in production

We solved this by creating a custom agent preset that explicitly loads both AGENTS.md files. Here's what we did:

1. Created a custom preset based on the shipped code preset, overriding the agent-instructions row to expand discovery:

- id: agent-instructions
  name: '@deepseek-ai/dsh-agent-instructions'
  config:
    maxBytes: 131072
    paths:
      - 'AGENTS.md'
      - '../AGENTS.md'

2. Set it as default in settings.yaml:

agent-presets:
  default: your-preset-name

3. Verified by starting a fresh session — the baseline now includes both AGENTS.md files (repo-local and parent/shared).

Key notes:

• Sessions created before the settings change keep the old preset (expected behavior, as noted in the original post).
• maxBytes was raised to 128KB since both files combined may exceed the default 64KB.
• No .dsh-root marker needed — using relative paths in the paths array was simpler.

Not a full fix for the discovery logic, but it's been stable for us. Posting in case it helps others hitting the same nested-repo issue.

[etzhe]

Thanks for posting the working workaround — overriding the per-session agent-instructions row with paths: ['AGENTS.md', '../AGENTS.md'] (and raising maxBytes to fit both files) is a clean, practical route for the nested-repo case. It's also a useful data point for the cross-linked #5333: it shows the baseline-composition seam is already configurable per-session, which is the same hook a native auto-load of ~/.dsh/memory/ (+ its MEMORY.md index) would want to use. I'll reference your preset approach there. Agreed it's a layout workaround rather than a fix to the project-root discovery logic itself.