| name | self-diagnostic |
| description | An 8-step procedure for verifying that Retainer's tools and specialist delegates are wired correctly end-to-end. Run when the operator asks for a self-check, a health check, or "is everything working?". Do not run unprompted. |
| agents | cognitive |
Self-diagnostic — procedural
When the operator asks for a self-check, run this procedure in
order. Each step touches a real surface and reports what
happened. The diagnostic IS the runtime — there is no separate
"test mode." If a step fails, that failure is a real signal
about the running system, not a test artifact.
The point of this skill is to drill the habit of trusting and
using the surface I actually have. Skip nothing. Don't
substitute reasoning for tool calls.
Step 1 — memory_write
Write a single ephemeral fact:
- key:
self_diagnostic_marker
- value:
<current ISO timestamp from the sensorium clock>
- scope:
ephemeral
- confidence:
1.0
- description:
self-diagnostic write, will be cleared
Record whether the write succeeded.
Step 2 — memory_read
Read back the marker key. The returned value should match what
I wrote in step 1. Record whether it matched.
Step 3 — memory_clear_key
Clear self_diagnostic_marker so the diagnostic doesn't leave
clutter in the operator's fact store. Record whether the clear
succeeded.
Step 4 — agent_observer
Dispatch with instruction: "Recall the most recent narrative
entry. One-line summary." The observer should reply with a
short summary of the most-recent cycle (the one that drove
this diagnostic, or the previous one).
If the observer reply is empty or generic ("no entries"), that
is still a successful dispatch — the chain works; the index is
just thin. Record success on dispatch, note any anomaly in the
content.
Step 5 — agent_researcher (skip-if-absent)
If agent_researcher appears in my tool list, dispatch with
instruction: "Search for retainer and return one source."
The researcher should reply with a result (URL + brief
summary).
If agent_researcher is NOT in my tool list, this means no
search-API keys are configured in the workspace. Record as
"skipped — researcher not configured" and continue. This is a
valid workspace state, not a failure.
Step 6 — agent_taskmaster
Dispatch with instruction: "List my pending tasks." The
taskmaster should reply with the Taskwarrior list table
verbatim (or no pending tasks for an empty store). Either
outcome is a successful dispatch.
Step 7 — agent_scheduler
Dispatch with instruction: "List my scheduled jobs." The
scheduler should reply with active jobs sorted by next-fire
ascending (or no active jobs for an empty workspace). Either
outcome is a successful dispatch.
Step 8 — read_skill
Call read_skill with the location of THIS skill (its path
appears in <available_skills> under <name>self-diagnostic </name>). The tool should return the body of this file.
Record whether the read succeeded — proves the skill substrate
itself is functional from inside a live cycle.
Final report
Reply with a structured summary, one line per step:
Self-diagnostic results
1. memory_write — ✅ ok | ❌ <error>
2. memory_read — ✅ ok (matched) | ⚠ ok (mismatch: got X, wanted Y) | ❌ <error>
3. memory_clear_key — ✅ ok | ❌ <error>
4. agent_observer — ✅ dispatched | ❌ <error>
5. agent_researcher — ✅ dispatched | ⏭ skipped (not configured) | ❌ <error>
6. agent_taskmaster — ✅ dispatched | ❌ <error>
7. agent_scheduler — ✅ dispatched | ❌ <error>
8. read_skill — ✅ ok | ❌ <error>
Overall: <healthy | degraded — see failures above>
Use the literal characters ✅ ⚠ ⏭ ❌ — the operator parses
this output by eye, and the symbols make pass/fail scannable.
When NOT to run
- Unprompted. The diagnostic costs ~7 LLM calls (one per
dispatch + one for the final reply); never run it without an
operator ask.
- Mid-task. If the operator is in the middle of asking for
something else, finish that first.
- Repeatedly within a session. Once per session is plenty;
results don't change minute-to-minute.
Failure modes I might see
- Step 4 / 5 / 6 / 7 returns "agent not available." The
delegate tool is missing from my tool list. That's a wiring
break — surface it plainly; the operator needs to know.
- Step 8 returns an error about path validation.
read_skill
is restrictive about which paths it accepts. If the skill's
<location> from the system prompt fails this check,
something is wrong with how the skill was discovered.
- Repeated LLM errors during a dispatch. Adapter-level
decode bug or rate-limit hit. The error message will name the
cause; report it verbatim.
