parallel-evolution

Files (4)

• SKILL.md
• finish.mjs · scripts/
• node-tests.mjs · scripts/
• test-node-tests.mjs · scripts/

SKILL.md (first lines)

# Evolving code in parallel

## Active operator-selected direct workflow

When `AGENTS.md` selects direct work, handle the human request serially without
subagents or external inference agents while `/orchestration` is off. Explicit on
uses native task owners: each implements/verifies/merges via `integrate`, then
reports; coordinator followups wait for the report. Coordinator alone handles UI
publication, `/update` and Sentinel. See `docs/ORCHESTRATION-MODE.md`.
Self-review is allowed and must be named
honestly. Use focused checks, ordinary Git integration through canonical main,
remote readback and clean trees. Do not create factory admission receipts or
run a full release gate unless requested. Preserve original unknown effects;
clean Git is not runtime settlement. The parallel procedure below applies only
when the operator re-enables it.

Verify this policy with
`node <repo>/scripts/test-direct-workflow-instructions.mjs`. It runs the complete
instruction checker plus private policy-removal mutations, not a full release gate.
In an allocated required worktree, use the verified direct CLI first; do not
call `spell_run` to rediscover `workspace_execution_context_unsupported`.
Inspect workspace status only if unknown; an unavailable binding still blocks
execution. Preserve the same postchecks, binding and instruction byte budget.
See `docs/SKILLS.md` for supported-path routing; no blind retry or fabricated proof.

For progressive model-envelope changes, `progressive-envelope-proof` records
`node <repo>/scripts/test-progressive-envelope.cjs <built-wa> <fresh-evidence>`
plus `--post`; use the verified direct CLI in required workspaces. Also run
`test-mcp-recovery.cjs` source+embedded reconstruction/no-replay proof and
`test-web-shadow-embedded.cjs` for the four-call source/citation oracle.
See `<repo>/docs/PROGRESSIVE-ENVELOPE.md`. The tiny three-way shadow measures
serialized bytes and replay mechanics, not autonomous quality or paid tokens.

For native subscription retry-buffer failures, `subscription-buffer-recovery-proof`
records `test-subscription-buffer-recovery.cjs <built-wa> <fresh-evidence>` and
retained-hash `--post`. Required workspaces use this verified direct CLI.
Keep exact507/body/empty-parser boundaries and prior-tool no-replay; run existing
server/disconnect/overflow/keepalive regressions. Request bytes are not model tokens;
unknown usage stays unknown. See `<repo>/docs/SUBSCRIPTION-BUFFER-RECOVERY.md`.

For binding-aware discovery, `binding-discovery-proof` records
`test-binding-discovery.cjs <built-wa> <fresh-evidence>` and retained-hash `--post`.
Required workspaces use this verified direct CLI. Run source/embedded seed/restart,
stable exact schemas, profile/role fences, unchanged native refusals and causal
old-tools control; also MCP native/restart and typed-routing regressions.
See `<repo>/docs/BINDING-DISCOVERY.md`. Blocked metadata is not execution,
and the deterministic zero-refusal plan is not autonomous-quality proof.

For provider-measured compaction changes, `context-usage-proof` records
`test-context-usage.cjs <built-wa> <fresh-evidence>` and retained-hash `--post`.
Required workspaces use this verified direct CLI. The source/embedded seed/restart,
actual92% boundary, complete-prefix/schema/model/provider/scope fences and
old-policy negative control must pass, along with focused overflow/accounting
regressions. Preserve originals and reasoning; estimated cold/tail context is
not provider usage. See `<repo>/docs/CONTEXT-COMPACTION.md`.

For tagged prompt/read_many bootstrap or recoverable public HTTP changes,
`layered-web-proof` records `test-layered-web.cjs <built-wa> <fresh-evidence>
<baseline-wa>` and its retained-hash `--post`. Required worktrees use that same
verified direct CLI, not unsupported replay. Also run the TAP web/HTTP suites,
existing envelope/full bodies/read-batch/MCP no-replay/socket regressions and
shipping/instruction checks. Keep original responses and source-scope boundaries;
HTTP acceptance is not completion, challenge solving is human-only, and a browser
solve does not certify anonymous HTTP recovery. See `<repo>/docs/LAYERED-PROMPT.md`
and `<repo>/docs/WEB-HTTP-OUTCOMES.md`. Tags/read_many may cost initial bytes;
report actual task outcomes and context, not a claimed saving.

For capture/lean-bootstrap/typed-routing changes, `envelope-iteration-proof`
records `test-envelope-iteration.cjs <built-wa> <fresh-evidence> <exact-capture>`
and `test-envelope-outcomes.cjs <built-wa> <other-fresh-evidence>`, with both
retained-hash `--post` checks. Required worktrees use these verified native Node
CLIs, not unsupported spell replay. Preserve complete originals, JSON shapes,
role/profile/workspace/definition fences and the unchanged global output budget.
Task plans are deterministic native replay, not autonomous quality or token/billing
proof. Report discovery and total input plus tool-result context; grouping already
independent singles does not itself save a model round. See
`<repo>/docs/PROGRESSIVE-ENVELOPE.md` and its measurement.

For skill-metadata envelope changes, `skill-discovery-proof` records
`node <repo>/scripts/test-skill-discovery.cjs <candidate-wa> <fresh-evidence> <baseline-wa>`
plus `--post`; required workspaces use this verified direct CLI, not unsupported
replay. Also run complete skill-body and progressive-envelope proofs. Preserve
scope/snapshot refusals and shared safety policy; first-use discovery and lexical
misses remain risks. Bytes alone do not establish autonomous quality or billing.

For `read_many` packing changes, `read-batch-packing-proof` records
`node <repo>/scripts/test-read-batch.cjs <built-wa> <fresh-evidence> <capture>`
plus `--post`; required worktrees use that verified direct CLI. Verify whole
results/receipts, scope fences, source/embedded native restart and exact historical
payload parity. The global view budget stays unchanged; follow cached `next_cursor`
with read_many, not an artifact-recovery workaround. See `<repo>/docs/BATCH-TOOLS.md`.

For MCP changes, `mcp-native-proof` records
`node <repo>/scripts/test-mcp-native.cjs <absolute-built-wa> <fresh-evidence>`
and its `--post` hash check. Use that verified direct CLI in required workspaces,
not unsupported spell replay. Also run private stdio/HTTP/CDP peers and the MCP
Chromium proof, then the required UI suite. See `<repo>/docs/MCP.md`; probe is not
app identity/send approval, and unknown effects are never replayed.

For subagent action-specific key/help changes, `subagent-message-contract-proof`
records `node <repo>/scripts/test-subagent-message-contract.mjs <built-wa>
<fresh-evidence> <old-wa>` and retained-hash `--post`. Required worktrees use this
verified direct CLI, not unsupported spell replay. It checks source/embedded help,
Responses serialization, real facade refusals/key forwarding and private SQLite
steering deduplication; continuation execution is mocked. Run the existing policy
suite and native subagent tests too. No live child or inference is admitted.

For independent batch-tool changes run
`node <repo>/scripts/test-batch-tools.cjs <absolute-built-wa> <fresh-evidence>`,
then the same CLI with `--post` before its arguments. `batch-tools-proof` records
that verified source/embedded plus retained-hash sequence; required workspaces
still use the direct CLI, not unsupported spell replay. See `<repo>/docs/BATCH-TOOLS.md`.
Batch only already-known independent items; retain every outcome/continuation,
preflight file mutations, and never replay unknown effects or claim a transaction.

**wasm-agent policy (parallel mode):** routine merges use appropriate focused source checks and independent
exact-source review, including shared runtime changes. Only the user selects pre-release
gating (`lane-policy.json`); never infer a full/combined gate from paths. Gate/closing-spell
instructions below apply to explicitly requested pre-release verification. Routine work

From Magaav/wasm-agent, MIT.

More from Magaav/wasm-agent