fix(slackbot): avoid corrupting final delivery when the streamed-char offset lands mid-token

[axis-guzus]

Problem

Agent answers delivered to Slack are intermittently corrupted — characters are dropped or words merge (observed in production: "market cap" → "marketountries", "defillama.com" → "vama.com", "authenticate" → "thenticate"). The content is correct server-side; only the Slack render is mangled.

Root cause

continuationText (services/slackbot/src/centaur/final-delivery.ts) blind-slices the canonical result_text at slackbot_streamed_answer_chars:

return text.slice(offset).trimStart();

That offset counts the source characters streamed live (the per-delta source-char accumulator in agent-session.ts / codex-session.ts), which is a different string-space than result_text:

• markdown normalization adds boundary characters to the live stream that aren't in the source count;
• whitespace / paragraph reflow changes lengths;
• the harness can revise the final answer — the code even logs slack_codex_canonical_answer_correction when the canonical answer length differs from what was streamed.

When the offset is misaligned, text.slice(offset) lands mid-token and drops characters. The offset is accumulated server-side with max(...), so overshoot is sticky.

Fix

continuationText only has the offset count, not the streamed string, so it can't verify the prefix. This guards the slice: only treat the offset as a split point when it lands on a whitespace boundary in result_text; when it's misaligned, fall back to posting the full answer (re-showing the already-streamed prefix is strictly preferable to emitting corrupted text). The aligned suffix-continuation path (live delivery cut off at a clean boundary) is unchanged.

Tests

Adds a regression test (offset lands mid-token → the full answer is posted, not a corrupted slice). The existing "posts only the missing suffix when live delivery was cut off" test still passes.

bun test src/centaur/final-delivery.test.ts  →  6 pass, 0 fail

Note

A more complete fix would plumb the streamed prefix text through (or compute the continuation server-side, where both strings are available) so the split point can be verified exactly rather than heuristically. Happy to follow up if you'd prefer that direction — this change is a minimal, safe guard against the user-visible corruption.

[Zygimantass]

We've merged the Rust rewrite in #344, meaning the old API and Slackbot services are deprecated. As such we're closing this PR - if you think this is a mistake, please reopen the PR and ping me.