Stage 2's main agent reads the sub-agent's own system prompt, and nothing uses it

[johanzander]

Problem

Stage 2's main agent reads .claude/agents/bess-analyst.md (25,681 B) as
REQUIRED READING item 3. That file is the bess-analyst sub-agent's own
system prompt, and no step in the workflow prompt tells the main agent to use
it for anything.

Follow-up to #650, which cut CLAUDE.md from the same floor. This is the
second-largest item left.

Measured

It is the sub-agent's system prompt. Frontmatter: name: bess-analyst,
tools: Read, Grep, Glob, Bash, WebFetch. So the main agent's read is a second
copy in a different context window, not a shared one.

No instruction consumes it. The six PROCESS steps are: get issue context →
identify the current problem → delegate to bess-analyst → verify the cited
file:line → publish → label. Item 3 is described as a "domain expertise
checklist", but nothing directs the main agent to apply it. Its one distinctive
task — step 4, "reading the file:line locations it cited … quote the actual
code" — needs the code, not a triage checklist.

The prompt already inlines the critical content. The task the prompt passes
to the sub-agent is a near-verbatim restatement of the file's
## CRITICAL: Separate Evidence from Claims — the section most plausibly
serving as a judging standard:

bess-analyst.md	already inline in the prompt
"The reporter's description is a hypothesis, not a diagnosis"	"The reporter's explanation is a HYPOTHESIS — do not assume it is correct"
"Never start from 'where in the code could this bug be?' … Start from what the debug bundle shows"	Step 2 "Triage the LATEST debug bundle FIRST"; Step 3 "Only after triage, read the design docs"
"verify that BESS Manager actually uses code path Y for this user's setup (inverter type, integration…)"	Step 2 "whether the reporter's HA integration matches the code path BESS Manager uses. Check inverter type (MIN vs SPH vs SolaX)"
"If the debug bundle shows fundamental issues … flag those FIRST"	Step 2 "If sensors are unavailable … that is likely the real problem"
"A design choice that is intentional and documented is not a bug"	Step 4(b) "the behavior isn't intentional per the design docs"

That content is currently loaded three times: sub-agent system prompt,
main-agent file read, and paraphrased inline.

The artifacts agree. All four landed Stage 2 analyses (#118, #252, #624,
#627) use the workflow's four headings — ## Root cause / ## Evidence /
## Proposed fix / ## Risks / open questions. None uses bess-analyst.md's
seven-item Output Format or its six-item type-B shape. #118's Evidence section
is exactly what step 4 asks for: quoted code at cited lines.

What is NOT duplicated

## Root-Cause Decomposition — the P-optimality / forecast-error /
control-noise split, with "do not average them into one blended verdict … that
has repeatedly produced wrong conclusions on this repo." That is a real
standard and it is not in the prompt. But it is a standard for the
investigator, which the sub-agent already has; the main agent holding an
unread copy changes nothing while no step asks it to audit against it.

Proposal

1. Drop item 3 from Stage 2's REQUIRED READING.
2. Because the decomposition standard is worth holding at the main-agent level,
   turn it into an instruction that actually fires — extend step 4 to reject a
   sub-agent report that blends the three categories into one verdict, or that
   cites no file:line.

Net: dead weight becomes a check. Saving is up to 25,681 B off the main agent
per turn (floor 92,438 → 66,757 B, −28%). "Up to", because whether a given
run complied with the read instruction is not observable — #646 showed
non-compliance happens.

Limits of this evidence

Inspection plus artifacts, not a controlled experiment. What is shown is that
nothing consumes the read and its critical content is already inline — not
that behaviour is provably identical without it. That is also the argument for
adding the step-4 check rather than only deleting: it leaves the main agent
with a stated standard instead of an unread file.

Explicitly out of scope

• Editing .claude/agents/bess-analyst.md itself. The sub-agent needs it whole.
• The other floor items (rules.md, architecture.md, the inline prompt) —
  all three are consumed by named steps.

Acceptance

• Stage 2 still delegates to bess-analyst and still publishes (Fix #646: make Stage 2 publish by command, and fail the job when it doesn't #649's
  verification step must stay green).
• Step 4 names a concrete rejection criterion rather than "independently verify".
• scripts/quality-check.sh passes, including Fix #646: make Stage 2 publish by command, and fail the job when it doesn't #649's publish-contract gate.