transcript.js reads block.input on tool_result blocks, silently dropping all tool output from compaction

[Shaderzx]

Version: morph-compact@0.2.8 (Claude Code plugin, installed via claude plugin install morph-compact@morph)

Summary

hooks/lib/transcript.js reads block.input when extracting tool_result content blocks. Anthropic tool_result blocks carry their payload in block.content, not block.input — input is a field on tool_use. The condition is therefore always false and every tool result in the transcript is silently dropped before the history is sent to the Compact API.

Net effect: Morph only ever receives assistant/user prose plus [tool: name] breadcrumbs. No file contents, no command output, no error messages, no search results. For a coding agent that is the majority of the useful context, and it's exactly the material that most benefits from verbatim compaction rather than lossy re-summarization.

Fails silently — no error, no warning. The only visible symptom is an implausibly small inputChars in the state file written by pre-compact.js.

Root cause

hooks/lib/transcript.js, in extractText():

      case "tool_result":
        if (typeof block.input === "string" && block.input) {   // <-- always undefined
          parts.push(`[result]: ${block.input}`);
        }
        break;

tool_result blocks in a Claude Code transcript have exactly these keys:

tool_result keys: ['content', 'tool_use_id', 'type']
has 'input': False

block.content is either a string or an array of content blocks ({type: "text", text: "..."}), depending on the tool.

Evidence

Measured by running the shipped parseTranscript() against a real ~/.claude/projects/<project>/<session>.jsonl, before vs. after fixing the field name:

BEFORE: messages=93   chars=35129
AFTER:  messages=137  chars=72302

44 messages consisted entirely of tool results and were dropped in full; total content roughly doubled. That transcript contained 40 tool_use blocks and 40 matching tool_result blocks — none of the results survived extraction.

A separate session recorded stats: {messageCount: 54, inputChars: 7952, ...} — about 147 chars per message, which is what tipped me off.

Repro

1. Run a Claude Code session that makes several tool calls producing substantial output.
2. Trigger /compact.
3. Inspect the state file written by pre-compact.js ($TMPDIR/compact-hook/<session_id>) before session-start.js consumes it.
4. stats.inputChars is far smaller than the real conversation; the resulting summary contains [tool: X] placeholders with no corresponding results.

Suggested patch

      case "tool_result": {
        // tool_result carries `content` (string | block[]), not `input`
        const c = block.content;
        if (typeof c === "string") {
          if (c) parts.push(`[result]: ${c}`);
        } else if (Array.isArray(c)) {
          for (const inner of c) {
            if (inner?.type === "text" && inner.text) {
              parts.push(`[result]: ${inner.text}`);
            }
          }
        }
        break;
      }

Verified locally: node --check passes and the measurement above is the result of this exact change.

Secondary note (not fixed here)

The tool_use case keeps only [tool: ${block.name}] and discards block.input, so the arguments are lost too — which file was read, which command ran, which pattern was searched. That may well be deliberate compression, but combined with the dropped results it means a compacted session retains no record of what the agent actually did, only that it did something. Worth a second look while the above is being fixed.