Per-repo config.yaml not read by fullsend run; workspace layout diverges from ADR 0033

[ifireball]

What happens

Per-repo installation writes configuration to .fullsend/config.yaml (via fullsend github setup), but agent workflows invoke fullsend run with --fullsend-dir defaulting to $GITHUB_WORKSPACE (repo root). The CLI then looks for config at {fullsend-dir}/config.yaml — i.e. config.yaml at the repo root, not .fullsend/config.yaml.

As a result, per-repo config fields are not applied during agent runs:

• runtime — on main, fullsend run always uses the default Claude backend and does not consult per-repo config. Setting runtime: dummy (or any other runtime) in .fullsend/config.yaml has no effect.
• allowed_remote_resources — tryLoadOrgConfig reads {fullsend-dir}/config.yaml and parses org-config shape only; per-repo config at .fullsend/config.yaml is never loaded for allowlist validation or config-driven agent resolution.
• create_issues, status notification defaults, and other per-repo config fields — same path mismatch; not reachable when fullsend-dir is the workspace root.

Separately, the Prepare workspace step in reusable agent workflows (reusable-triage.yml, etc.) always copies upstream harness/agents/skills content to the workspace root (harness/, agents/, …), regardless of install_mode. Only the customized overlay source path differs for per-repo (.fullsend/customized/ vs customized/).

This diverges from ADR 0033, which specifies that per-repo mode should:

• Prepare workspace with root .fullsend/ (populating .fullsend/harness/, etc.)
• Pass --fullsend-dir=.fullsend to fullsend run

The implementation never completed that layout. Per-repo installs also do not receive harness wrapper files under .fullsend/harness/ (unlike per-org config repos, which get them via HarnessWrappersLayer). Instead, harness files are materialized at workflow runtime by copying from vendored or sparse-checked upstream defaults into the workspace root.

What should happen

When a repository is installed in per-repo mode:

• Config written to .fullsend/config.yaml should be read consistently by fullsend run and related CLI paths (runtime selection, allowlist, config-driven agents, status notifications, etc.).
• The directory layout used by reusable workflows (Prepare workspace, --fullsend-dir, harness resolution) should be internally consistent — either matching ADR 0033 (.fullsend/ as the fullsend root) or documented as an intentional deviation with explicit fallback rules.

How to reproduce

1. Run fullsend github setup owner/repo --vendor --runtime dummy on a test repository.
2. Confirm .fullsend/config.yaml contains runtime: dummy.
3. Trigger a triage run via the per-repo shim (install_mode: per-repo).
4. Observe that fullsend run does not load runtime from .fullsend/config.yaml (on main, Claude runs regardless).

Attempting to align config lookup by passing fullsend-dir: .fullsend in reusable-triage.yml (as explored in #1982) causes a different failure: harness files are still prepared at workspace root, so fullsend run errors with harness file not found: tried triage.yaml and triage.yml.

Context

Discovered while implementing behaviour tests in #1982, which require runtime: dummy from per-repo config. Production per-repo triage may appear to work today because the default Claude runtime matches typical expectations and harness files are copied to workspace root at workflow runtime — but any per-repo config field that fullsend run should honor is effectively unreachable.

Related: ADR 0033 per-repo installation mode, --vendor install path (CollectPerRepoInstallFiles), reusable workflow Prepare workspace step.

[fullsend-ai-triage]

🤖 Finished Triage · ✅ Success · Started 10:08 PM UTC · Completed 10:11 PM UTC
Commit: 0a95cac · View workflow run →

[fullsend-ai-triage]

Triage Summary

Severity: High · Category: Bug

This is a well-documented config path mismatch affecting all per-repo installations.

Problem

Per-repo setup writes config to .fullsend/config.yaml, but fullsend run defaults --fullsend-dir to the workspace root and looks for config.yaml there — never reading the per-repo config. All per-repo config fields (runtime, allowed_remote_resources, create_issues, status notifications) are silently ignored.

The reusable workflow Prepare workspace step compounds this by always materializing harness files at the workspace root regardless of install_mode, making it impossible to simply pass --fullsend-dir=.fullsend without breaking harness resolution.

Root Cause Hypothesis

The per-repo installation path correctly writes config under .fullsend/, but the runtime invocation path was never updated to set --fullsend-dir=.fullsend for per-repo mode. The Prepare workspace step also never implemented the ADR 0033 layout (.fullsend/harness/, etc.), so the two halves are inconsistent.

Recommended Fix

1. Update the reusable workflow Prepare workspace step to materialize harness/agents/skills under .fullsend/ when install_mode: per-repo
2. Pass --fullsend-dir=.fullsend for per-repo agent runs
3. Ensure fullsend run config loading resolves the correct path in both modes

Key files: reusable-triage.yml (and sibling workflows), internal/cli/ (config loading), internal/appsetup/ (CollectPerRepoInstallFiles)

Proposed Test

Scenario: Per-repo config is loaded during agent run
  Given a repository with per-repo installation
  And .fullsend/config.yaml contains runtime: dummy
  When a triage agent run is triggered
  Then fullsend run uses the dummy runtime
  And does not fall back to the default Claude backend

Edge cases:
  - Config exists at both .fullsend/config.yaml and repo root (precedence)
  - Missing config file (graceful fallback)
  - Per-org mode remains unaffected

The reproduction steps provided in the issue are clear and actionable. Related work in #1982 discovered this while implementing behaviour tests with runtime: dummy.

---

Labels: Adding component/harness (workspace layout and harness resolution involved) and priority/high (all per-repo config silently ignored); existing component/install and type/bug labels are correct.

[fullsend-ai-coder]

🤖 Finished Code · ❌ Failure · Started 10:12 PM UTC · Completed 10:31 PM UTC
Commit: 0a95cac · View workflow run →

[ggallen]

Inline Code Review — Issue #2970

Confirmed Root Causes

Verified four contributing factors in the codebase:

---

1. Workspace Prepare step always targets workspace root

Files: All 6 reusable-*.yml — e.g. reusable-triage.yml:94-99

for dir in ${LAYERED_DIRS}; do
  if [[ -d "${SRC}/${dir}" ]]; then
    mkdir -p "${dir}"              # ← always workspace root
    cp -r "${SRC}/${dir}/." "${dir}/"
  fi
done

Harness/agents/skills are copied to workspace root regardless of install_mode. Only the customized overlay source differs (customized vs .fullsend/customized at line 101). ADR 0033 §5 specifies per-repo should populate under .fullsend/.

---

2. --fullsend-dir never set to .fullsend for per-repo mode

File: action.yml:314

FULLSEND_DIR="${FULLSEND_DIR:-${GITHUB_WORKSPACE}}"

No reusable workflow passes the fullsend-dir input when invoking the action (e.g. reusable-triage.yml:157-167 — no fullsend-dir: key). The input defaults to "" (action.yml:22), which resolves to $GITHUB_WORKSPACE.

Result: fullsend run always looks for config at $GITHUB_WORKSPACE/config.yaml, never .fullsend/config.yaml.

---

3. Config parsed as OrgConfig, not PerRepoConfig

File: internal/cli/orgconfig.go:23-37, called from run.go:200-201

orgConfigPath := filepath.Join(absFullsendDir, "config.yaml")
orgCfg := tryLoadOrgConfig(orgConfigPath, printer)

tryLoadOrgConfig calls config.ParseOrgConfig. But per-repo installs write a PerRepoConfig — a different struct type (config.go:407-414) missing Dispatch, Inference, Defaults, and Repos fields. Even if the path were corrected, the parser silently drops per-repo-only fields like Roles.

Second loading site: run.go:2312-2314 (setupStatusNotifier) also reads config as OrgConfig with the same path+type mismatch.

Note: loadAgentConfig in agent.go:73-93 (used by fullsend agent add/list/update/remove) already handles both formats via format detection — so the pattern for a fix exists in the codebase.

---

4. Runtime selection is hardcoded (not a config field)

File: internal/runtime/runtime.go:65-68

func Default() Backend {
    r := ClaudeRuntime{}
    return Backend{Runtime: r, Transcripts: r}
}

Neither OrgConfig nor PerRepoConfig structs have a runtime field. The runtime: dummy use case from the issue requires both a config field and a runtime registry, which are part of PR #1982, not main.

---

Code Agent Failure Analysis (run 28624853135)

The code agent's approach was correct — it modified the right 10 files:

• 6 reusable workflows: Added DEST prefix variable for per-repo workspace layout + passed fullsend-dir input to the action
• orgconfig.go: Added fallback from ParseOrgConfig → ParsePerRepoConfig with an OrgConfigFromPerRepo adapter
• config.go: Added OrgConfigFromPerRepo bridge function
• Tests for both

The agent exited code 0, but post-script pre-commit hooks rejected the commit:

• pinact (SHA-pin check) — Failed
• lint markdown links — Failed

The changes never made it to a PR branch.

---

Fix Approach

The code agent's approach aligns with ADR 0033 §5. A fix should:

1. Parameterize Prepare workspace by DEST — per-repo populates .fullsend/{agents,harness,...}, per-org populates at root (all 6 reusable workflows)
2. Pass fullsend-dir: .fullsend when install_mode == 'per-repo' in the action invocation step
3. Add OrgConfigFromPerRepo adapter and update tryLoadOrgConfig/requireOrgConfig to fall back to ParsePerRepoConfig (same pattern as loadAgentConfig in agent.go)
4. Fix the post-script issues — ensure pinact pins are correct and no broken markdown links

[ggallen]

/fs-code Generate a PR again, taking this comment (#2970 (comment)) into account, while also fixing the pinact and lint markdown links issues.

[fullsend-ai-coder]

🤖 Finished Code · ❌ Failure · Started 7:00 PM UTC · Completed 7:16 PM UTC
Commit: 0a95cac · View workflow run →