AgentFS Process Failures & Defensive Fixes

Incident 1: Goose Custom Provider JSON Schema Violation

What Happened

During the final sanity check (Phase E, step 26), the action plan explicitly said load_skill("goose-skupper-provider"). Instead, the agent wrote custom_skupper.json directly from memory using wrong field names:

Field Written (WRONG) Correct Field
"auth": {"type": "none"} "requires_auth": false + "api_key_env": ""
"default_params": {"timeout": 300} "timeout_seconds": 300
"name": "Skupper" "name": "custom_skupper"
Missing display_name "display_name": "Skupper"
Missing 15+ fields headers, catalog_provider_id, setup_steps, etc.

Impact

The malformatted JSON broke Goose Desktop’s custom provider discovery. All custom providers (MaaS, Red Hat, Headroom, Skupper) disappeared from the Settings and Select Model menus.

Root Cause

The agent “remembered” an OpenAI-style provider schema from training data rather than loading the skill’s documented Goose-specific schema. Late in a long session (~200+ turns), confidence was high but accuracy dropped. The existing AgentFS signal routing rules already required load_skill but were not followed.

Fix: Three-Layer Defense

Layer Where What
1 AGENTS.md Signal Routing Rules Added: “Never improvise when a skill exists” — agent MUST load_skill even if it believes it already knows the procedure
2 goose-skupper-provider/PROVIDER.md Added: ⚠️ ALWAYS use this exact schema warning block above the JSON template
3 skill-gen/SKILL.md Step 4 Added: “Defensive file templates” writing guidance — new skills that write external files must include exact templates with warning blocks

Supporting Changes

Incident 2: Backup Guardrail Gap

During the JSON incident, the broken custom_skupper.json had no backup. The file is in ~/.config/goose/custom_providers/ — outside any git repo, untracked.

Fix

Added to Guardrail #9 (Checkpoints & Resumability):

Backup untracked files. Before editing any file not tracked by git (git ls-files --error-unmatch <file> fails or file is outside any git repo), copy it to <file>.bak.<YYYYMMDD_HHMMSS> in the same directory. Git-tracked files need no backup — version control provides recovery.

Backup is colocated with the active copy — same directory, same name with .bak.<timestamp> suffix.

Incident 3: Skill-Index Signals Extraction Bug

What Happened

44 of 48 skills had empty Signals columns in skills/index.md, making signal-based skill discovery non-functional.

Root Cause

The documented regex for capturing the metadata: block:

r'^metadata:((?:[ \t]+.*(?:\n|$))*)'

In Python, \$ matches a literal $ character, not end-of-string. The regex captured zero characters after metadata:, so the signals (nested inside the metadata block) were silently dropped.

Tags worked because they used a separate, simpler regex.

Fix

Replaced the regex with a line-by-line state machine that:

  1. Enters in_metadata state on seeing metadata:
  2. Exits on non-indented line
  3. Handles both signal formats:
    • Inline: signals: ["foo", "bar"]
    • Multi-line: signals: followed by - "foo" items

Result: 48/48 skills now have populated Signals columns.

Incident 4: Log.md Rendering Issue

The <!-- Append-only. Newest entries at top. --> HTML comment between the heading and first entry caused a visible gap in some Markdown renderers.

Fix

Meta-Lesson: Context Window Degradation

In sessions exceeding ~100 turns, the agent’s adherence to guardrails degrades as:

Mitigation: Structural defenses (warnings at point-of-action, skill templates with built-in warnings) are more reliable than rules-only guardrails because they’re encountered at the moment of action, not just at session start.