Skillscript

Post-v0.19.0 friction patches. Three convergent fixes from live-test
signal:

A. Shared resolveRuntimeConfigFromEnv() — structural cleanup

Closes adopter CR f2549ddf + follow-up aeccddac. Pre-v0.19.1
each env→option translation lived inside the CLI cascade — so
programmatic adopters using bootstrap() or DashboardServer
directly hit silent default-off behavior on every SKILLSCRIPT_* knob.
v0.18.9 patched the shell allowlist instance feature-specifically;
v0.19.0 immediately reproduced the same bug with two new env vars
(SKILLSCRIPT_EVENT_INGRESS_ENABLED + SKILLSCRIPT_EVENT_INGRESS_AUTH_TOKEN).

The structural fix: a shared resolveRuntimeConfigFromEnv() resolver
in src/runtime-env-resolver.ts covering ALL SKILLSCRIPT_* knobs:

port / host
mcpCallerIdentityHeader
enableUnsafeShell
forceAlwaysDraft
pollIntervalSeconds
absoluteTimeoutMs
maxRecursionDepth
shellAllowlist
eventIngressEnabled
eventIngressAuthToken

Both bootstrap() and DashboardServer call it. New SKILLSCRIPT_*
knobs added to the resolver inherit env support across CLI +
programmatic paths automatically — the bug class closes, not just
this instance.

Perry's explicit-wins guard (from 42de3d72) generalized:

• opts.X === undefined → fall back to env
• opts.X === any defined value (including false, 0, [], "")
  → authoritative; env does NOT widen

Adopters who explicitly pass shellAllowlist: [] or
eventIngressEnabled: false to assert specific posture get exactly
that, regardless of ambient env. Security-load-bearing.

B. Dynamic declarative-trigger registration

Closes Perry F1 (f68eb84d) + adopter Finding 2 (d538f7df).
Pre-v0.19.1, declarative triggers (# Triggers: cron|event:) only
wired at boot via wireDeclarativeTriggers. Skills authored
mid-session via skill_write weren't live until next restart —
contradicted the "POST /event accepts the new event_name now"
expectation external POSTers have.

skill_write + skill_status MCP handlers now call a new
Scheduler.syncDeclarativeTriggersForSkill(name, parsedTriggers, parsedVars, status) that:

• Drops prior declarative triggers for the skill
• Re-registers them from the new parsed source IF status === Approved
• Status transitions Approved→Disabled drop; Disabled→Approved restore

Author skill → /event accepts immediately. No restart.

C. Imperative register_trigger derives params from # Vars:

Closes Perry F2 (f68eb84d). Declarative wiring derived event-
trigger params from the skill's # Vars: (v0.19.0 feature); imperative
MCP register_trigger didn't apply the same derivation. Result: POST
/event with params for an imperatively-registered event 400'd with
"unknown params" because registration.params = [].

register_trigger handler now loads the named skill, parses, derives
params from # Vars: for event-source registrations. Cron registrations
unchanged (cron has no params).

D. Schema enum trim regression guard

Closes Perry F3 (f68eb84d). v0.19.0 already trimmed the MCP
register_trigger / list_triggers source enums to ["cron", "event"] in source + dist; Perry's live test likely hit a stale
cached state. v0.19.1 adds explicit regression tests asserting the
enum literals so future trigger-source changes can't silently
desynchronize the schema and the parser.

E. Doc add: # Autonomous: true for event/cron mutation skills

Closes adopter Finding 1 (d538f7df). Not a v0.19.x regression
— it's the pre-existing v0.14.x mutation gate. But event-fired skills
are a new category where adopters first discover this. Adopter-playbook
"Trigger model" section now has an explicit subsection with worked
examples covering:

• Skill-level # Autonomous: true for cron/event-fired mutation skills
• Per-op approved="<reason>" kwarg as an alternative
• Why ?? / ask() doesn't apply (no interactive author in cron/event
  contexts)

Tests

23 new tests in v0.19.1-dynamic-registration-and-env-resolver.test.ts:

• 7 env-resolver parse cases (every SKILLSCRIPT_* knob; invalid values
  fall through; explicit false stays false; empty allowlist string;
  whitespace trim)
• 4 bootstrap env-fallback cases (env populates Scheduler config;
  explicit opts win over env including [] deny-all + false)
• 4 DashboardServer env-fallback cases (event ingress + port + host +
  auth token from env)
• 4 dynamic-registration cases (skill_write registers immediately;
  Draft → no registration; Approved → Disabled drops; Disabled →
  Approved restores)
• 2 imperative-param-derivation cases (event derives from #Vars; cron
  doesn't)
• 2 schema-enum regression cases (register_trigger + list_triggers
  source enum is ["cron", "event"])

Plus 1 existing test updated (dogfood-t6b.test.ts test 6) to reflect
the new dynamic-registration semantics (heartbeat's declarative cron
trigger registers immediately on Approved → list_triggers shows 2
entries after the test 6 imperative-register, not 1).

1814 tests total; typecheck clean.

Banking note (feedback memory 9d969eb1)

"Operator-controllable knobs ALL flow through a shared env resolver;
never patch the env→option translation per-feature." Two-release thrash
(v0.18.8 → v0.18.9 patch → v0.19.0 reintroduces → v0.19.1 generalizes)
is the structural-vs-per-feature lesson. Future SKILLSCRIPT_* knobs
extend the resolver; the bug class is closed.

Docs

• docs/adopter-playbook.md — new "# Autonomous: true for event/
  cron skills" subsection under "Trigger model"
• No other doc changes; existing trigger-model + env-var docs already
  cover the surfaces

No breaking changes (v0.19.0's removal of session/agent-event/file-watch/
sensor sources stands; v0.19.1 doesn't change parser behavior).