Wires sandcastle's native `resumeSession` into the dispatch loop so
the implementer walks into task N already knowing what task N-1
discovered — repo layout, helper signatures, gate output, prior diff.
No scratchpad / no hand-curated context file; the agent's own Claude
Code conversation log is the carrier.
Three guardrails keep it bounded:
- Story boundary reset. `currentSession` is dropped whenever
findNextTask returns a different story id. New domain ≈ new
context — keeps story 03 from inheriting story 02's residue.
- Token-threshold reset. After each approved slice, sum the
implementer's last-iteration usage (inputTokens +
cacheCreationInputTokens + cacheReadInputTokens — caching saves
dollars but doesn't free window space). If above
SANDCASTLE_SESSION_TOKEN_RESET (default 140000 ≈ 70% of Sonnet
4.6's 200k), drop the session before the next task. Configurable
via env.
- Context-exhausted safety net. If the model rejects with
"prompt is too long" / "context_length_exceeded" / similar, the
retry loop drops the session and re-runs the attempt fresh
exactly once. Doesn't count against SANDCASTLE_MAX_ATTEMPTS
(different failure mode).
Reviewer always runs fresh — each approve/reject decision should be
independent of prior tasks to keep the gate honest. Within a single
slice's reject-fixup retries, the implementer also carries forward
across attempts (so attempt 2 sees attempt 1's reasoning + the
reviewer notes), but that's per-slice cumulative, not cross-slice.
runOneSlice now returns { sessionId, usage } so executeDispatch can
make the carry-or-reset decision per slice.
Danijel Martinek
d5c01209ea