v0.33.0

v0.33.0 adds offline command specifications for agents, bounded official-MCP
alert history and configurable read deadlines. Existing commands retain their
backends and JSON contracts; no automatic source switching is introduced.

Offline command specifications

Use tv spec for the command index, then request a known path directly:

tv spec mcp bars
tv spec data values
tv spec screener open

Syntax comes from the running binary's clap definitions. Annotated commands also
provide prerequisites, sources, constraints, side effects, discovery examples
and interpretation limits. Semantic coverage currently includes 164 of 169
executable paths. Unannotated semantics stay explicit, and validation coverage
remains partial even for annotated commands. This is not an output-schema export
or an offline request validator. Use --help with older binaries.

The specifications distinguish selected-chart observations, account operations,
scanner data and official MCP. In particular, successful communication does not
prove complete data, chart readiness, successful persistence or authority to
perform follow-up actions. These descriptions expose existing behavior; they do
not silently fix or replace legacy operations.

MCP history and read deadlines

tv mcp alert history --symbol NASDAQ:AAPL
tv mcp --timeout 90 alert history --symbol NASDAQ:AAPL

Alert history returns mcp_alert_history.v1 with validated alert identity and UTC
firing times. Coverage remains unconfirmed. Notification messages and webhook
contents are not included. Native verification succeeded with a 90-second budget
and empty history; nonempty native normalization and success within the default
30-second deadline remain unverified. Nonempty response structure and fixtures
provide separate evidence, not complete history coverage.

--timeout accepts 1–180 seconds for provider reads, with a 30-second default.
Login, logout, status and mutations reject this override before credential or
provider access. There is no --timeout-secs alias. A longer deadline allows
more waiting; it does not resolve provider rate limits or guarantee success.
Official technical-indicator retrieval is deferred and is not added here.

Connection reuse and login guidance

MCP watchlist and alert mutations reuse their connection and discovered tool
catalog for readback. Synthetic measurements confirm less repeated setup; no
live-provider latency improvement is claimed. Mutation admission, deadlines,
readback failure reporting and protection against replay remain unchanged.

Explicit login explains browser and credential-store steps on an interactive
terminal. Captured stderr remains free of human progress messages, and usable
credentials are reused. Linux help now describes implemented Secret Service
support; graphical Linux OAuth remains unverified.

Portable skills

The seven runtime skills now have one editable source in root skills/ for
standard skill-manager discovery. Development links to those sources; archives
still contain real, independently attachable files under both agent skill roots.
Chart, market-data, Pine, Screener and account guidance includes the source and
partial-result limits described by tv spec.

Local discovery/install checks passed. Remote installation of this new layout
and Windows skill-manager execution remain unverified. Skill installation and
MCP credentials are separate; installing a skill does not authorize an account.