fix: truncate step summary to stay under GitHub's 1024k limit

[sunnypatell]

long Claude sessions (30+ min, many tool calls) generate step summary markdown that exceeds GitHub's hard 1024k limit for $GITHUB_STEP_SUMMARY, causing:

$GITHUB_STEP_SUMMARY upload aborted, supports content up to a size of 1024k, got 1026k.

the error is non-fatal (Claude's work still succeeds), but it's noisy and prevents viewing the execution summary. users currently have to set display_report: false to avoid it entirely, losing the summary for all sessions.

fixes #927

what changed

added truncateStepSummary() in format-turns.ts that:

1. checks the markdown's byte size against a 1000k budget (leaves 24k headroom for content from earlier steps in the same job)
2. if over budget, cuts at the last section boundary (---) to avoid breaking mid-content
3. appends a truncation notice explaining the limit and pointing to display_report: false
4. handles multi-byte characters correctly via TextEncoder/TextDecoder (no mid-codepoint cuts)

applied in both the formatted report path and the raw JSON fallback path in writeStepSummary().

tests

5 new tests in format-turns.test.ts:

• under-limit content passes through unchanged
• over-limit content is truncated with notice
• cuts at last section separator when one exists in the latter half
• multi-byte characters (emoji) survive without corruption
• default limit matches GitHub's constraint

all 35 tests pass.

[Copilot]

Pull request overview

Adds step-summary truncation to prevent GitHub Actions $GITHUB_STEP_SUMMARY size-limit issues when rendering large Claude Code reports.

Changes:

• Introduces truncateStepSummary() to truncate markdown output to a byte budget and append a notice.
• Applies truncation when writing both formatted and fallback (raw JSON) step summaries.
• Adds unit tests covering truncation behavior, section-boundary cutting, and multi-byte character handling.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 5 comments.

File	Description
src/entrypoints/format-turns.ts	Adds truncateStepSummary() and a default byte budget for step-summary output.
src/entrypoints/run.ts	Wraps step-summary writes with truncateStepSummary() to avoid oversize reports.
test/format-turns.test.ts	Adds tests validating truncation and UTF-8/multi-byte behavior.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

You can also share your feedback on Copilot code review. Take the survey.

[sunnypatell]

addressing copilot's 5 inline comments (3 fixed in d5b55ae, 2 intentionally kept as-is):

fixed in d5b55ae:

1. budget going negative (format-turns.ts:443) - added a guard: if maxBytes <= noticeBytes, returns just the notice trimmed to fit. handles the edge case where the limit is absurdly small.

2. replacement char inflating encoded size (format-turns.ts:453) - added a re-verification loop that trims one character at a time if the re-encoded result exceeds the budget after TextDecoder replaces incomplete multi-byte sequences with U+FFFD.

3. test name mismatch (format-turns.test.ts:469) - renamed to "uses default limit with headroom below GitHub's 1024k constraint" to reflect the intentional 1000k budget.

4. tighter section boundary assertions (format-turns.test.ts:450) - added not.toContain("Section 3") and not.toContain("C".repeat(50)) to validate Section 3 content was excluded. also fixed the limit calculation so it actually triggers the boundary cut.

intentionally not changed:

5. checking existing summary file size (run.ts:108) - the 24k headroom (1000k vs 1024k) covers realistic earlier-step content. stat-ing the summary file on every write adds I/O and complexity for a scenario that effectively doesn't happen in practice (earlier steps would need to write 24k+ of summary, which is extremely unusual for steps that precede Claude). if this ever becomes a real problem, the fix would be to reduce the budget further, not to add file I/O.