fix(opencode): port 8 isolated bug fixes from upstream opencode - #199

This migration rebuilds `workspace` to add the `name` column, but its
INSERT...SELECT unconditionally read `name` from the old table. On a
database that predates this column, the migration itself would fail
with "no such column: name". Check `PRAGMA table_info` first and fall
back to an empty string when the column isn't there yet.

Ported from upstream commit db86d03 ("fix(core): tolerate missing
workspace names"). Pure migration-safety fix; no schema or behavior
change for databases that already have the column, and unrelated to
Batuta/Breniac/desktop.

Co-Authored-By: Claude Sonnet 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01AGh7o4NA8jUjwakDzZo12e

resolveModelID() was double-prefixing two cases:
- Full Bedrock ARNs (`arn:...`) got a region prefix appended, which
  Bedrock rejects.
- Any model id containing "deepseek" (not just the R1 models that
  need cross-region inference) got prefixed with the region.

Add an early return for ARN ids and narrow the DeepSeek match to
"deepseek.r1".

Ported from upstream commit ac1758c ("fix: preserve Bedrock
DeepSeek model ids (anomalyco#34441)"). Isolated to Bedrock model id
resolution; no effect on other providers or on Batuta/Breniac/desktop.

Co-Authored-By: Claude Sonnet 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01AGh7o4NA8jUjwakDzZo12e

`headerTimeout` already accepted `false` to disable its timeout, but
`chunkTimeout`'s schema only accepted a positive integer even though
the runtime (aisdk.ts wrapSSE / provider.ts fetch wrapper) already
treats a non-number chunkTimeout as "no timeout". Users setting
`chunkTimeout: false` in provider config hit a schema validation
error instead of getting the already-supported disable behavior.

Widen the schema to `PositiveInt | Literal(false)`, matching
headerTimeout.

Ported from the schema portion of upstream commit 4eb29a6
("fix(opencode): default chunk timeout to five minutes (anomalyco#46890)").
This fork's fetch-wrapping code in packages/opencode/src/provider/
provider.ts already diverged from upstream's default-to-300000ms
change, so only the schema fix (which matches existing runtime
behavior) was ported — not the default-value wiring, to avoid
touching code this fork customized.

Co-Authored-By: Claude Sonnet 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01AGh7o4NA8jUjwakDzZo12e

The `x-parent-session-id` header was only attached to the non-opencode
provider branch, so subagent/child sessions running on opencode-hosted
models never told the backend which session spawned them.  Move the
header outside the provider branch so it's always sent when present.

Ported from upstream commit 611cc73 ("fix(opencode): send parent
session header (anomalyco#44752)"). Header-only change; does not affect
Batuta/Breniac/desktop request handling.

Co-Authored-By: Claude Sonnet 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01AGh7o4NA8jUjwakDzZo12e

Anthropic reports, via providerMetadata.anthropic.inputTransformations,
when it removed thinking blocks before the model saw the prompt. A
non-empty list here means opencode changed conversation history behind
a signed thinking block. Log a warning with the session/message/model
context so this churn is traceable instead of silently ignored.

Ported from upstream commit 3f39a32 ("feat(opencode): tolerate
Anthropic thinking block binding (anomalyco#46653)"). Purely additive logging
on step-finish; no change to request/response handling, and unrelated
to Batuta/Breniac/desktop.

Co-Authored-By: Claude Sonnet 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01AGh7o4NA8jUjwakDzZo12e

When a tool part transitions back through the "running" resolve path
(e.g. re-emitted deltas for a call already in progress), it reset
`time.start` to `Date.now()` every time, corrupting the reported tool
duration. Keep the original start time when the existing state is
already "running".

Ported from upstream commit 765ae64 ("fix(core): Fix for incorrect
time.start reset in tool call logging (anomalyco#32574) (anomalyco#32596)"). Isolated to
tool-part timing; no change to tool execution or permissions, and
unrelated to Batuta/Breniac/desktop.

Co-Authored-By: Claude Sonnet 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01AGh7o4NA8jUjwakDzZo12e

The per-change metadata always included `movePath`, even when the
change wasn't a rename, sending `movePath: undefined` explicitly
instead of omitting the key. Spread it in conditionally so the key is
absent for non-move changes.

Ported from upstream commit f7da00f ("fix(opencode): omit empty
apply patch move path (anomalyco#45329)"). Isolated to the patch-change summary
object; no change to patch application logic, and unrelated to
Batuta/Breniac/desktop.

Co-Authored-By: Claude Sonnet 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01AGh7o4NA8jUjwakDzZo12e

…imits

fork-map.md claimed the upstream remote is always pre-configured and
upstream-sync.md's compare step assumed dev and upstream/dev share a
common ancestor. Neither holds anymore: fresh sandbox clones don't have
the upstream remote, and origin/dev..upstream/dev has no merge-base
(upstream rewrote its history at some point), so commit-range diffing
silently lists their entire history instead of what's actually new.
Documents the tree-diff/byte-diff workaround used to port 8 real fixes
in this session, plus the sandbox proxy blocking pkg.pr.new/GitHub
tarballs that had nothing to do with those fixes but looked like a
regression at first.

Co-Authored-By: Claude Sonnet 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01AGh7o4NA8jUjwakDzZo12e