Builder: edits on edit_workflow / edit_step; document reads bound by the window, not the 50k ceiling

[Evanfeenstra]

Why

The builder could not change one line (max_pages: 3 → 100) of a ~50k-char workflow made in the UI. Its own edit_workflow never had an ownership gate, but it only took the whole YAML back, and get_workflow is cut at the tool result cap (50k chars on a 1M window — the "about 50k characters" it reported is the truncation marker). So the builder published a throwaway workflow (exec cat → meta/publish-workflow) and hit the evolve surface's rule instead: a meta/* step republishes only over names it stamped ai (EVOLVE_SPEC §6). The chat transcript confirms the chain: tmp-republish-from-file published, run 1791566214864 returned Workflow "tweet-engagement-study" exists and is not agent-authored (publisher: none), relayed as "refused for workflows I didn't originally create".

What

• edit_workflow / edit_step take edits — exact-string replacements applied in order to the ACTIVE version's source, each required to match exactly once (applyEdits, src/authoring.ts). A no-match or a repeated match names the edit and the lines it hit, and publishes nothing; validation and the claims gate run on the result as before. yaml / code OR edits, not both. The fix for the tweet workflow is now one call: edit_workflow({ name, edits: [{ old: "max_pages: 3", new: "max_pages: 100" }] }).
• Document reads are bound by the window, not the ceiling. get_workflow and get_step (DOCUMENT_TOOLS, src/chat-store.ts) get AiDeps.documentMaxChars — the window's headroom with no env ceiling, 176k chars on a 1M window (every Claude 5 model), 16k on Haiku 4.5 — and are never re-cut on replay: truncateToolMessages skips them, since the number they were cut to follows the model and a fixed cap there would rewrite what the model already read (the append-only rule; a cache miss and, on preserved-thinking models, a 400). Every other result keeps the 50k ceiling.
• The system prompt tells the builder to use edits for a small change to a big workflow and never to work around a long YAML by other means.

Result	1M window	200k window
get_workflow, get_step	176k chars	16k chars
everything else	50k (unchanged)	16k

Unchanged

• The meta/publish-workflow ownership rule — it guards an unsupervised agent inside a run.
• The tool-result-spill plan stays proposed; its "error instead of a cut" exemption for documents is noted as superseded (with edits, a cut document is still editable).

Tests

npm test: 1386 pass. New: edits on both tools (one-line change, no-match, repeated match with line numbers, validation error after an edit, edits follow the active version after a rollback), the document cap at the source vs the ordinary cap, and the replay skip. Not exercised against a live model: the change is tool schema + execute logic, not provider streaming. tsc clean.