execution-plan-review

execution-plan-review

Present the per-issue ordering and execution assignments as one table, absorb the user's revisions, and keep the GitHub issues (the single source of truth cold agents read) in sync. milestoneplan is the read-only view of the same table; it routes Execution-block fixes back here.

Steps

1. Render the table from the issues, not from memory

Fetch every issue in the milestone (gh issue list --milestone ... --json number,title,body) and parse the ## Execution blocks:

Issue	C	Depends on	Runs after	Build model	Effort	Validate	fableplan first?	Plan effort

Cell rules:

• Validate is the effective <model> · <effort> the pipeline dispatches. The model derives from the [C<score>] band and is never stampable (a missing prefix routes as band 5); the effort is the band default unless a Validate effort: line stamps one. The validate-issue step 6 band table is the authority; do not restate it here. Show a stamp as Fable 5.1 · medium (stamped; band high); show a clamp as Opus 5 · high (stamped <tier> → high: low/medium are Fable-only).
• Plan effort defaults to high on fableplan first: Yes issues; a stamped line shows as <tier> (stamped). On a No issue show — only when no line is stamped; any stamped line, high included, shows as <tier> (inert — no plan stage runs), because the milestone pipeline logs it on every run and the source-of-truth table must not mask it.
• An absent ordering field shows as missing, never none, so legacy prose inference is not silently discarded.

Follow with 2–3 sentences on the pattern (dominant bands, which issues plan first at score ≥ 71, the review trigger) for a sanity check against the validate-issue step 6 band table.

2. Take revisions

• Shorthand: "11 should be medium", "12 depends on 8 and 9", "13 runs after 12", "clear 14's dependencies", "plan 17 at medium", "validate 271 at medium". Resolve row-vs-issue ambiguity against the table just shown; confirm in half a sentence when still ambiguous. A bare effort revision means the build effort; confirm in half a sentence when the issue also has a plan stage.
• External CLI builds ("build 275 with luna on codex at max", "277 grok on cursor high", "278 on cursor with cursor-grok-4.6-high") write - **Build model:** <Name> (Codex CLI) or <Name> (Cursor CLI), appending , <model-id> inside the parenthetical when the user names an id, and write the tier to Effort:. max is a Codex CLI-only tier: a Cursor max is written as xhigh, and a Claude max as xhigh on every Claude model, Fable included; say the clamp applied. Only Luna on Codex and Grok on Cursor have a default model id; any other name needs an explicit id, so ask for one in half a sentence before writing.
• "validate at " writes - **Validate effort:** <tier>; "plan at " writes - **Plan effort:** <tier>. A revision that names the band default for that stage removes the line instead of stamping it, and say so. A milestone-wide revision ("every Fable validate at medium", "all plans at medium") applies to every issue whose stage runs on Fable, in one batch.
• Push back once when a revision conflicts with the score band (validate-issue step 6 / prd-to-issues): fableplan below score 71, or dropping below Opus on a money/security/irreversible-deletion issue. One recommendation with the reason, then the user decides.
• Batch revisions; do not round-trip to GitHub per message.
• A Plan effort revision on a fableplan first: No issue is inert. Say so once and drop it, or ask whether they meant to turn fableplan on. A flip to Yes plans at high unless the same batch stamps a Plan effort: line. A flip Yes → No on an issue carrying a Plan effort line strips that line during write-back; say you dropped it. Offer once to strip any inert stamp already on an issue instead of carrying it forward silently.
• Preserve the edge kind: Depends on stays a hard prerequisite, Runs after stays ordering-only. Never move an issue between the fields to simplify the graph.
• Before writing: verify every referenced issue exists, reject self-references, deduplicate each list, reject a predecessor present in both fields, and recursively fetch referenced issues outside the milestone until the explicit ordering graph closes.
• Reject the whole batch before write-back if the combined graph (every Depends on and Runs after edge, including unchanged and external issues) contains a cycle.