…mpact input

estimateTotalChars now includes reasoning byte length so the threshold reflects
what Anthropic actually charges as input_tokens. Without this, sessions with
heavy reasoning (debug, planning) cross the model's degradation cliff (~200K
tokens) silently before compaction fires.

serializePart drops reasoning instead of forwarding to Morph compact API. Stale
reasoning is the LLM's prior internal monologue \u2014 not worth tokens to summarize
into the frozen block. The latest assistant turn's reasoning still lives in
the uncompacted tail until next compaction.

PLUGIN_VERSION bumped to 2.0.0-rate-reasoning-patch for visibility in logs.
package.json version bumped to 2.0.9-reasoning-patch.

Detects `/morph compact` (also /morph-compact, /morph_compact) at the start
of a user message via the chat.message hook, strips it from the part text,
and arms a module-level flag. The next experimental.chat.messages.transform
invocation bypasses both threshold checks (first-compact and re-compact paths)
and forces compaction. Flag is cleared regardless of success/failure.

Useful when the trigger threshold is set conservatively (e.g. 200K tokens for
Anthropic quality cliff) but the user wants to compact earlier — or to test
the compact path without waiting for the threshold to be crossed.

PLUGIN_VERSION → 2.0.1-rate-slash-compact
package.json   → 2.0.10-slash-compact

Previous placeholder "[/morph compact — manual compaction requested]" was
indistinguishable from a real user request to the LLM, which dutifully
produced a Markdown summary response (Goal / Constraints / Progress / ...).
The user perceived this as a second compaction running on top of Morphs.

Now:
- If the user typed extra text after the command, keep that text and drop
  the command part entirely.
- If the command was the entire message (the common case), replace it with
  an explicit no-reply directive phrased as an internal marker:
  "[internal: Morph compaction triggered in background. This is not a
  request; do not respond.]"

Side benefit: messages.transform still sees a user turn (no empty-message
backend issues), but the LLM no longer mistakes the placeholder for a
summarization request.

PLUGIN_VERSION -> 2.0.2-rate-slash-compact-silent
package.json   -> 2.0.11-slash-compact-silent

ROOT CAUSE: opencode (packages/opencode/src/session/prompt.ts:1566) invokes
the hook as

  yield* plugin.trigger("experimental.chat.messages.transform", {}, { messages: msgs })

and then ignores trigger.s return value, continuing to read msgs directly
on line 1572 (MessageV2.toModelMessagesEffect(msgs, model)).

The previous plugin code reassigned the local reference:

  output.messages = [...frozen, ...recent]

which only updated the wrapper object passed to the hook. msgs (the original
array opencode keeps using) was never mutated. The LLM kept receiving the
full conversation despite Morph reporting "828 messages -> 3 frozen".

EVIDENCE in trial DB ses_1c53c4870ffeSjEoAkn6vHiy53:

  Before first auto compact (00:06:01): cache.read=517972, total=518853
  After  first auto compact (00:08:04): cache.read=519046, total=521255

cache.read stayed flat. Anthropic saw the same prompt prefix. Compaction
was silently a no-op end-to-end.

FIX: helper replaceMessages(output, next) does

  output.messages.splice(0, output.messages.length, ...next)

which mutates the same array reference opencode holds. All four assignment
sites (under-threshold reuse, first compact, re-compact, re-compact failure
fallback) are routed through it. Spread is replaced with a push loop above
50k elements as a stack-safety guard.

PLUGIN_VERSION -> 2.0.3-rate-inplace-mutation
package.json   -> 2.0.12-inplace-mutation

…poisoned by parent

ROOT CAUSE: compactionState and forceCompactOnNextTransform were module-
level lets, shared across every session in the trial process. After the
parent session compacted, every subsequent LLM call in any session
(retries, subagents spawned via task tool) hit the same global. The
transform hook then ran `replaceMessages(output, [...frozen, ...uncompacted])`
where `frozen` came from the parent and `uncompacted` was the subagent's
tiny new messages array — overwriting the subagent's actual prompt with
the parent's frozen block.

OBSERVED: 3 explore subagents spawned at 01:35 with completely different
review prompts ALL answered the same unrelated question about FFI/WASM
flash API — which was the dominant topic captured in the parent's frozen
block from its 01:01 compaction.

NOTE: only became visible AFTER v2.0.3 (in-place replaceMessages). v2.0.0-
2.0.2 had the same global-state bug, but their `output.messages = [...]`
reassignment was ignored by opencode (it reads `msgs` not `output`), so
the poisoned messages array never reached the LLM. v2.0.3 fixed reference
preservation and exposed the latent global-state bug.

FIX: compactionStateBySession = Map<sessionID, CompactionState>; ditto
for forceCompactBySession. transform hook reads sessionID off
messages[0].info.sessionID (SDK guarantees the field). chat.message hook
reads sessionID directly from input. If sessionID is unavailable, transform
degrades to no-op — safer than touching another session's state.

Fix Morph handling of OpenCode compaction events

…hllm#9) (morphllm#20)

Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>

bun.lock was gitignored, so the SDK bump in morphllm#20 only updated
package-lock.json. Track bun.lock so bun users get the pinned
dependency tree too.

Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>

Desktop users can set `morph.apiKey` in opencode.json instead of the
MORPH_API_KEY env var. Env var still takes precedence.

Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>

The morphsdk WarpGrepClient.execute() async generator returns a Promise
from processAgentResult() via `return promise`. Node.js auto-awaits it
per ECMAScript spec, but Bun does not — the unresolved Promise becomes
the .next() value.

Without the explicit await, result was a Promise whose .success was
undefined, and formatWarpGrepResult returned the generic failure message:
"Search failed: search returned no error details."

The fix is a single line: `result = await value`. This is harmless on
Node.js (await on a non-Promise is a no-op) and resolves the Promise on
Bun.

Closes morphllm#24

…ee-fallback

fix: await async generator return value in warpgrep_codebase_search (Bun compatibility Issue)

Reconcile the per-session compaction work with main's native-compaction
rework. Both code paths now coexist:

- chat.message `/morph compact` slash command (force-compact) and the
  session.compacted event handler are kept as separate hooks; the event
  handler now clears per-session state (compactionStateBySession) instead
  of the removed global.
- sessionID is resolved once at the top of the transform hook so the
  native-compaction branch can use replaceMessages() (the in-place
  no-op fix) and clear its own session's frozen block.
- Compaction log lines keep the per-session context and adopt main's
  formatCompressionPercent() helper.
- PLUGIN_VERSION restored to main's value (dropped the dev label).

typecheck clean; 67/67 tests pass.

…r-session-state

fix(compact): make compaction effective, per-session state, and add /morph compact trigger