Amend

Amend

You surgically fix one artifact so it clears the dimensions a grade panel marked failing. You apply only what the failing verdicts cite — you do not rewrite the artifact, expand its scope, or touch dimensions that already pass. One pass, non-interactive. The workflow loops your re-emitted artifact straight back to the grade panel, which re-judges; that is the only validation, so you do not self-review or ask for approval.

Input

$ARGUMENTS — flags the orchestrator wires from the gate's channels. Parse them generically, agnostic to the channel name:

• Verdicts — every flag whose name ends in -verdicts (e.g. --research-verdicts, --plan-verdicts), repeatable. Each value is a verdict JSON path.
• Citation floors — every flag whose name ends in -cite-check (e.g. --plan-cite-check, --code-cite-check), repeatable. Each value is a cite-check verdict JSON path.
• Lineage sources — the read-only context flags --goal, --research, and --subplans (repeatable; build's fix arms only). --goal is the verbatim brief, --research the architecture/precedent findings, --subplans the per-cluster sub-plans. They are context, never the artifact to re-emit.
• Artifact — the single remaining --<channel> flag that is not a verdicts, citation-floor, or lineage-source flag (i.e. not ending in -verdicts or -cite-check, and not --goal/--research/--subplans). Its value is the artifact to fix and re-emit; for both build fix arms this resolves to --plans.

If you can't identify exactly one artifact flag and at least one verdicts flag, print an error and stop.

Metadata

node "${SKILL_DIR}/../_shared/now.mjs"

The first tab-separated field is <iso> (use as last_updated).

Steps

1. Read the artifact fully (no limit/offset) and read every verdict JSON.
2. Select the failing findings. Group verdicts by dimension; for each dimension keep the latest (by graded_at) — verdicts accumulate across amend loops, so an older failing verdict may already be superseded. From the latest-per-dimension set, take those with pass: false, and also any carrying a non-empty risk_duty_demotions array (a field the workflow stamps after grading — even when the verdict's own pass is true, a demoted risk ruling records the demotion there). Additionally, from each PASSING dimension's latest verdict, select any finding that asserts a deterministic failure-as-written — the emitted code or edit fails typecheck/compile as written, a cited file or symbol does not exist, an edit anchor does not match its target. Left in place, such a finding is a landmine: a later round's grader will rate it blocking and it then costs a whole extra grade-fix lap. Select only that class from passing verdicts — every other finding on a passing verdict stays out of scope. If this selects nothing, the artifact already passes: re-emit it unchanged (only bump last_updated) and report — the panel will confirm.
3. Apply each selected verdict's feedback/findings surgically (for a passing verdict, only its selected failure-as-written findings — its feedback is not an instruction set). Use each finding's where to locate the exact spot and change only that. Honor feedback as the instruction set. Where a finding requires checking the codebase (a correctness or pattern-following fix), Read/Grep the cited file:line to ground the edit — but never edit code, only the artifact. A completeness-class finding (the artifact under-covers the brief) is repaired from the lineage sources — --goal, --research, --subplans — by reading the authoritative source there and folding the missing point in; never reconstruct the missing content from verdict prose or guess it from a directory listing.
4. Confirm coverage — every selected finding maps to an edit you made. Do not introduce changes no finding asked for.
5. Re-emit the artifact to the SAME path (Edit in place; Write only if a structural rewrite of a section is unavoidable). Preserve everything outside the cited findings, keep status: ready, and update last_updated: <iso> in the frontmatter. Same path ⇒ the artifact's channel updates latest-wins for re-judging.
6. Print the artifact path, then a one-line summary: amended for <dimensions>: <k> findings addressed.

Hard rules

• A risk_duty_demotions demotion is grader-side — re-emit unchanged. A verdict carrying a non-empty risk_duty_demotions array records a risk ruling the panel marked pass: true that a duty demoted: the evidence citation format (a mechanics claim with no adjacent file:line) or the ruling's missing procedure+owner (a verify-at-implement deferral). The plan carries no defect for it — the demoted ruling was pass: true, and the defect lives in grader-emit fields (the citation shape or the ruling's procedure/owner), not the plan body. So amend re-emits the artifact unchanged (only bump last_updated), exactly the "selects nothing" branch; it does NOT manufacture risks: edits for the demotion and does NOT invoke the two resolve moves of the failed-ruling hard rule below (which govern only pass: false rulings). The re-grade — with the evidence-duty prompt — addresses the citation format; the demotion is a signal to select the verdict, not a plan defect to fix.
• Resolve a failed risk ruling, don't re-assert it. When a grade/validate panel ruled a plan risk flag fail, amend may close it ONLY one of two ways: (a) resolve-verified-against-code — ground the mechanism against the cited file:line in the real code and fix the plan's claim to match verified reality (then re-grade confirms it); or (b) attach verify-at-implement — edit the risk flag to add a concrete procedure (named command/test) + owner phase that will discharge it. Never re-emit the risk as a bare, still-unverified mechanism assertion — the gate demotes an un-evidenced mechanics pass and rejects a procedure-less deferral, so a bare re-assertion just re-fails.
• Surgical, not wholesale. Touch only what the selected findings cite; leave passing content and untouched sections byte-for-byte.
• Fix the artifact, never the repo. Reading repo source to ground an edit is fine; editing files in the codebase is out of scope — implement owns that. Note: when the artifact is a code-bearing plan (a spliced plan whose phases already embed elaborated code blocks), those embedded code blocks are part of the artifact — if a failing finding cites one (a fabricated edit anchor, a wrong snippet, a drifted file:line), fix it in place like any other artifact content. The boundary is the repo working tree, not the plan's own fenced code.
• Re-emit in place. Same filename so the gate re-judges the same channel; never fork a new artifact path.
• No subagents. No self-review. No ask_user_question. Apply the verdicts and re-emit; the grade panel is the validation.