update_pull_request cannot use replace-island, but update_issue can — shared helper, asymmetric enums

[blavity-machine-user]

Summary

replace-island is reachable for update_issue but not for update_pull_request, even though the island-splicing logic lives in the shared body-update helper. The result is that a PR-body write can only be replace, append, or prepend — so any workflow that needs to update a section in place has to have the model regenerate the entire body, and a model that regenerates prose silently rewrites the parts it was not asked to touch.

Verified against main (latest release at time of writing: v0.86.2). Originally found on v0.81.6, and the gap is unchanged.

The asymmetry

pkg/workflow/js/safe_outputs_tools.json — operation enums:

Tool	operation enum
update_issue	replace, append, prepend, replace-island
update_pull_request	replace, append, prepend
update_release	replace, append, prepend

pkg/parser/schemas/main_workflow_schema.json agrees on the workflow-config side: update-pull-request.operation is ['append','prepend','replace'], so a caller cannot set it as a default either.

Meanwhile replace-island is implemented and exercised — it appears in pkg/workflow/safe_outputs_validation_config.go, in actions/setup/js/fuzz_update_body_harness.cjs, and in the compiled lock files of several workflows in this repo. So this reads as an enum that was extended for issues and not for pull requests, rather than a deliberate restriction.

Why it matters

Marker-delimited splicing is the only operation in that set that is deterministic. replace hands the model the job of reproducing everything it is not changing, and that is not a job models do reliably.

Concretely, with update-pull-request + replace, across 23 successful writes by one workflow:

• 6 of 9 PRs lost human-authored content. Worst case: a 194-line description reduced to 74 — 167 lines gone.
• The destructive writes correlated with splicing into an existing section (7 of 7), not with body length. Pure appends were harmless (0 lines removed).
• Even non-lossy writes corrupted prose: bold markers moved so a sentence changed meaning, and hard-wrapped lines were reflowed.
• Length-based detection would not catch it — one body kept every heading and near-identical character count while paraphrasing paragraphs away.

There is no prompt-level fix for this. "Echo the body back verbatim" is an instruction to the same component that is failing, and a deterministic splice is the only thing that removes the model from the copy path.

Ask

Expose replace-island for update_pull_request — in both safe_outputs_tools.json and main_workflow_schema.json — matching update_issue. If the omission is deliberate for pull requests, saying so in the docs would help, since the helper's presence suggests otherwise.

Happy to open a PR if a maintainer confirms the direction.

Related

#19347 was the same shape on a different surface: the handler and validation schema supported a field that the compiled tool schema never exposed, so the agent could not set it.