a record-level extra_prompt reaches no prompt and reports nothing: the render site passes only config #2110

Origin: rant 2026-10-11T11:23:55.136401+08:00

Part: 1/1

A record-level extra_prompt reaches nothing and reports nothing — the render site passes only config

What is wrong

TaskHandler._build_evolution_prompt hands the templates "task": self._config, and _config is the record's config sub-dictionary (scheduler.py:794). All six built-in templates read the section as a guarded block:

{% if task.extra_prompt %}
### Task-specific Instructions (extra_prompt from tasks.yml)

{{ task.extra_prompt }}
{% endif %}

The daemon renders with jinja2.Undefined, so a name that is not in the context is the empty string: the condition takes the false branch and the heading and body both disappear — no error, no warning, no log line. An extra_prompt written at the record's top level, beside interval / enabled, is therefore dropped in silence.

Measured on master 60541e39 (2026-10-11, cyc20261011-193407) through the real builder, TaskHandler._apply_record(record) → _build_evolution_prompt():

Where the wrong position came from

The module's own schema comment advertised it:

  sandbox / description / extra_prompt — optional; anything else is ignored.

Two of those three are not the same as the third, and the file says so elsewhere:

• sandbox really is record-level (scheduler.py:773) — and is also accepted under config, with the record-level value winning (_resolve_sandbox, scheduler.py:853).
• extra_prompt is honoured under config only; the design that introduced it (PR emrg: task prompts: support per-task extra_prompt from tasks.yml config (rant 2026-08-24T21:59:15) #964) renders task as the config, and its test spells that out.

So the comment listed a field that works where it claims, beside one that does not, and a reader of the schema had no way to tell which was which. validate_task_record does not look at the position either, and the CRUD/GUI write path does not warn.

Why it matters

The host reported the cost (rant 2026-10-11T11:23:55, translated from the Chinese original): the promote task's 4601-char extra_prompt (in place since 2026-08-31) and the competition task's 3042-char one never entered a single prompt — nothing in llm.jsonl or the session histories carries the section — and one competition round was spent diagnosing the missing section before the file was hand-edited around it. The schema comment is what made the wrong position look right.

Feedback is the host's first ask, ahead of leniency: with the render silent, a task's operator gets no signal at all that their instructions are not being read.

The fix

emrg/server/scheduler.py:

1. _template_task_context(config, record) — the mapping the templates render against. extra_prompt is honoured in both positions; config wins when a record carries both, because it is the canonical home of the whole task namespace. Blank counts as absent on both sides, since {% if %} renders nothing for "".
2. The losing copy is named in a warning — with the task's name and the size of the ignored copy — so the remaining ambiguous case is audible rather than silent. Being ignored is not the defect; being ignored quietly is.
3. The schema comment is corrected: sandbox keeps its two positions with their precedence stated, extra_prompt now says both positions are honoured and which wins, and description no longer hides inside a triple that implied all three behaved alike.

The fallback exists for the file written against the old comment, not to make the top level a second equal spelling: config is what the templates' task namespace is built from, so it stays authoritative.

Verification

• Reversed direction, through the real builder: the three new pins fail on master (60541e39) while their two controls pass. Full suite on the fixed head: 4759 passed, 27 skipped.
• uv run python -c "from emrg.client.app import run_client" and uv run python -m emrg --help — rc 0.
• Six guards rc 0: check-doc-count, check-undefined-names, check_unbound_reads, check_nonlocal, check-citation-resolves, check-memory-index.
• Mutation arms: removing the fallback KILLED its pin; removing the ambiguity warning KILLED its pin; making a blank copy count as a value KILLED the blank-rule pin; a reworded schema comment (control) SURVIVED.