ADR 0030 — Cron context chaining (v53-F5)
Date: 2026-07-17 · Status: accepted
Context
Every tick was fire-and-forget: a disk-check script's finding could not reach the morning digest; composing them meant one monolithic script or hand-copying output. Hermes chains cron jobs; the operator asked for the same.
Decision
-
chainnames a source schedule;last_outputis what a tick produced. Synchronous castes only record output (note: its text; script: its stdout+stderr; digest: the composed summary — all capped at 4KB). A task-caste schedule can CONSUME a chain but not source one: its output lands after the tick, asynchronously (recorded bound; a post-run hook is future work if a field test wants it). -
Context, never instructions. The chained output arrives labeled —
[Context from schedule 'X']:— before the schedule's own instructions (task castes get the explicit[Your instructions]:separator; scripts get it as stdin: data into an operator-vetted command, never a new command). The memory-injection posture, applied to ticks. -
Acyclic, depth ≤ 3, validated at creation.
validate_chainwalks the chain at every creation surface (POST /api/schedules and the carded propose_schedule); a cycle, an unknown source, or a fourth level is a 400/tool error with the reason. Deeper chains are a sign the user should write one script. -
The per-job
modeloverride from the draft plan was CUT (the v53 review): three of four castes would store it unused, and it existed only for a hypothetical future conversation caste — stored-but-unused config is against the house YAGNI law (the v42 caste scar). It arrives WITH that caste, if that caste ever ships.
Consequences
script A → digest Bworks out of the box: B's digest opens with A's latest finding. Trust is unchanged — chaining moves DATA between schedules the operator already created through gated surfaces; the tick-time policy gates on dispatching castes are untouched.schedule_view(asdict) exposeschainandlast_outputon every schedule surface for free.