Adopt the evolved process docs from alkssh (their Phase 0 is the family
reference for SDD practice; adapted here for alkgen):
- sdd_process.md: POC placement two modes (branch mode vs standalone
scratch project), closed frontmatter enum variants (taskgraph 0.1.3)
with the three gotchas, impact table, POC specialist finding-first
deliverables
- agent specs: architect/taskgraph rules for OQ blocker trackers,
implementation-specialist conventions remapped to alkgen (no async,
wasm-clean, preserve_order, strict mode), coordinator spawn constraints
remapped, code-reviewer checks swapped for template/determinism rules,
poc-specialist two-mode workflow
- all adopted/edited files are em-dash-free (AGENTS.md convention 12)
- (b) direct codecs resolved AND validated: generated types+codecs,
ReadResult<Packet>, no serde/serde_json/alktype in output
- benchmark: ~10x faster than Value-shaped generated, ~15-20x faster
than interpreter, ~4x off raw no-alloc baseline (String/Vec allocs)
- F-6: names come from the BAST doc (plan erases $defs names);
OQ-GN-02 refined: plan for semantics, BAST for names
- V{key} variant naming for numeric discriminator keys
- clarification: alkcall is a library others build on; alkgen branches
1+2 are a dev tool generating static files into downstream projects
(protocol crates alktty/alktunnels/sftp, cloud integrations vast.ai/
runpod) from provided schemas (BAST or jsonschema/openapi)
- OQ-GN-01 resolved: module-level output for v1; fragments out
(defeats dev-tool purpose), crate generation out for v1 (composed
from modules later if asked). Residue: file partitioning policy
- OQ-GN-06 reframed: generated files land in projects that don't
depend on alktype today, so dependency footprint is consumer's
choice, not wire-shaped; leaning self-contained, POC decides
- branch 2 reframed: static counterpart of alkhttp from_* adapters
- plans for offsets/sizes (already computed+validated by the engine,
free at generation time since alktype is already a dep), BAST for
descriptive metadata; drift risk inverts into ordinary
dependency-upgrade tracking of alktype's plan API
- residue: exact plan->context projection is the Phase 1 context-
builder spec task
- OQ-GN-03 resolved as posture: interpreter/compiler duplication is
the point (dynamic vs static deployment profiles); alktype near
schema-method floor, ~20x off hand-rolled; alkgen's value prop is
recovering most of that gap. Drift stays the standing risk.
- OQ-GN-08 POC 1 calibration: three-way benchmark (plan-walk vs
generated vs hand-rolled) doubles as drift check + perf validation;
OQ-GN-05 boundary reframed as performance trade
glm-5.3-flash