[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M7EX6P8DCTP09680K3XPQPSN":3,"skill-same:01M7EX6P8DCTP09680K3XPQPSN":56},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":18,"publishedAt":19,"linkOnly":20,"markdownContent":21,"markdownBody":22,"frontMatter":23,"refName":26,"commitSha":27,"skillDirectory":28,"forksCount":14,"repositoryHtmlUrl":29,"repositoryHomepageUrl":6,"ownerHtmlUrl":30,"ownerBlogUrl":6,"ownerDescription":31,"updatedAt":32,"install":33,"resources":45},"01M7EX6P8DCTP09680K3XPQPSN","plan",null,"Work through a plan file one item a session, with ele carrying the place across \u002Fclear; pick up the active plan, write a new one, or convert a spec's open items into plans.","nicolaspearson\u002Fclaude-ephant","nicolaspearson","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F8216388?v=4","01M7EX60Q0XJ6YXNG31B27ZNSK","Nicolas Pearson","01M7EX60QN5C49E2KH9ZR9YNEW",0,"GPL-3.0","https:\u002F\u002Fgithub.com\u002Fnicolaspearson\u002Fclaude-ephant\u002Fblob\u002Fmain\u002Fplugin\u002Fskills\u002Fplan\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"---\nargument-hint: \"[new \u003Cname> | convert \u003Cspec>]\"\ndescription: Work through a plan file one item a session, with ele carrying the place across \u002Fclear; pick up the active plan, write a new one, or convert a spec's open items into plans.\ndisable-model-invocation: true\nname: plan\n---\n\n# Plan\n\nA plan is a Markdown file under `.documentation\u002Fplans\u002F` in the repository whose code it changes. The file is the list and its checkboxes the state; ele stores only which plans are active. `ele plan show` parses and reports, and `ele plan check` lists the warnings `show` counts; every judgement below is yours. A plan is shared, so never write an ele note id or session into it.\n\nThe format:\n\n- Front matter: `kind: plan`, `created`, and `spec` when the plan implements a spec in the same repository. Nothing derived, such as counts\n- Sections: `## Why`, `## Requirements`, `## Context` (facts, dated), `## Decisions`, `## Actions`, `## Inbox`, `## Out of scope`\n- `## Requirements`, optional, numbered, in the user's words: what the plan must deliver, the list the review loop checks the items against. Only the user changes it. A plan written before the section has none\n- A Context line not verified against the code or data says \"not checked\"\n- An architectural Decision carries a `rejected: \u003Coption>, because …` line for each option it turned down, so an ADR drafted at the close has its considered options\n- An item is a `- [ ]` or `- [x]` line and its indented body\n  - Each action but the closing one leads with a short lowercase kebab name, `- [ ] judge-blocks: …`, written once and never changed, unique across active and closed plans. `ele plan start` refuses a name another active plan holds or a closed plan used, and an action starting `word: ` is read as named `word`; an Inbox line is never named: its leading `word: ` or date is not a name\n  - Each action but the closing one carries a `done when …` line in its body: what will be observable once it is done\n  - Progress goes in the body as dated lines, `as of \u003Cdate>: …` or `next: …`, so a session stopping mid-item needs no handoff\n  - A tick carries one line saying what was asserted, at which boundary and at which commit\n  - A `blocked:` line holds an item back: `blocked: judge-blocks, mining-reports\u002Freport-figures; \u003Creason>` names items in this plan or `\u003Cplan>\u002F\u003Cname>` in another of the same project, `\u003Cplan>` being its file name without `.md`, and `ele plan show` frees the item once each is ticked and flags one done, missing, or left unticked by a closed plan; a line with anything but names before its `;` is prose and holds the item until removed by hand, and `ele plan check` warns on an empty or malformed list and a cycle. ele reads only the plans this project has started or closed, so a `\u003Cplan>` it has neither of is flagged: start a draft it names, and write a wait on a plan in another repository as prose with no comma before its `;`, as a comma there reads as a malformed list\n  - `after: 2026-10-22` or `after: 2026-10-22 19:28Z` (UTC) holds an item until then, and `ele plan show` flags it to drop once passed\n- The last action folds each Decision through `\u002Fele:plan-audit close \u003Cpath>`, then what the plan built, and its out-of-scope lines as a `Not doing:` list in the section the plan implemented, into its spec; with no spec, it proposes each out-of-scope line as a `decision` through `\u002Fele:remember`. Then it runs `ele plan done \u003Cpath>`, which stores the plan's item names and whether each was ticked, and only then deletes the plan, since `done` reads the file. Never tick it while the inbox holds an item; `ele plan check` says when every other action is done, and warns when it is ticked with items there\n- `## Inbox` takes work captured mid-session, each item dated with where it came from\n- `## Out of scope` holds what triage ruled out, one line each: `- \u003Cgist>: \u003Cwhy> (ruled \u003Cdate>, from \u003Csource>)`. A ruled-out item returns only if the plan's why is redrawn\n\n## Asking in rounds\n\n`new` and `convert` settle questions in rounds. A round asks every question whose prerequisites are settled, numbered, each with your recommended answer, worded so that yes accepts it; a question that hangs on one still open waits for a later round, since asked early it is asked again. Facts are yours to find, as Grounding says: ask the user for decisions, and for a fact neither settles. A decision goes in `## Decisions`, with its reason, of each plan that relies on it; a fact about an item goes under that item as a dated line. The rounds end when no question is open and the user confirms the whole; only then does the step that asked them go on. A question `convert` step 1 leaves as a note is not asked\n\n## Grounding\n\nBefore the first round of `new` or `convert`, read the repository's ADRs, found by the ADRs section of `reference\u002Fplan-practices.md` in the plugin root, the directory holding `skills\u002F`. An accepted one is a settled constraint: a question it answers is not asked, and a wish it rules out is said to be so, superseding it being a question of its own. Another status is context. Before the first round, and before a later one resting on facts not yet found, launch `ele:explorer` agents for those facts, one per angle, at most 3, in one message, however small the repository. Read no code yourself before the first round, since a session that reads it launches no explorer, and so runs no recall. Brief each by absolute path: `reference\u002Fexplorer.md` in the plugin root, the file to follow; the user's `CLAUDE.md`, under `$CLAUDE_CONFIG_DIR` when set, else `~\u002F.claude`, and the repository's, the convention files; and give it its question and angle. Where `ele:explorer` is not listed, launch `general-purpose` with the same brief. A fact that goes in `## Context` unverified, including one an explorer lists under Not checked, says \"not checked\".\n\n## Designs\n\nBefore the round of `new` step 1 or `convert` step 2 that puts a design question meeting plan-audit Close mode's architectural test (hard to reverse, surprising without context and the result of a real trade-off, all three), launch two `ele:architect` agents in one message, each under a different constraint pulling that trade-off its own way, such as the smallest change against the shape easiest to extend. Brief each by absolute path: `reference\u002Farchitect.md` in the plugin root, the file to follow; the convention files Grounding names; and give it the explorers' findings the question rests on, the question and its constraint. Where `ele:architect` is not listed, launch `general-purpose` with the same brief. Put the question with each design as an answer, your recommendation first. The chosen design goes in `## Decisions`; each design not chosen, and each other option the user turned down, under it as `rejected: \u003Coption>, because …`, the reason from that design's Costs. A design question failing the test gets no architects, since each is a full design run, on opus for `ele:architect`. Neither does one raised once a draft exists, as an `open` review-loop finding or in `convert` step 5's re-audit round: ask it as any other question, since by then a draft exists whose design the question can be weighed against.\n\n## new \u003Cname>\n\n1. Ground the rounds, then settle in rounds whatever of the why, the requirements, the items and their design questions the conversation has not already settled; each decision it already made goes in `## Decisions` with its reason. Done when the user has confirmed the whole\n2. Write `.documentation\u002Fplans\u002F\u003Cname>.md` from `references\u002Ftemplate.md`, with today's date and the plan's own path in the closing action. `## Requirements` holds what the rounds confirmed the plan must deliver, numbered, in the user's words; each architectural Decision its `rejected:` lines, as Designs says. With no spec, drop the `spec` line and replace the closing action's \"then what this plan built and its out-of-scope lines as `Not doing:` into the spec\" with \"propose each out-of-scope line as a `decision` through `\u002Fele:remember`\". Done when `ele plan check` on the file lists no warning\n3. Run the review loop over the file. Done when the loop's summary is shown and its questions answered\n4. Show the file, then on approval run `ele plan start .documentation\u002Fplans\u002F\u003Cname>.md`. Done when it prints the plan as started, or the user has declined\n\n## convert \u003Cspec>\n\n1. List the spec's open items with their sections. Leave out those already in a plan, and those gated on a measurement or an undecided question: they stay notes, a `question` proposed through `\u002Fele:remember` where none holds one. Done when every open item is listed as to plan, already planned, or left as a note\n2. Ground the rounds, one explorer's angle being each item's status against the code or data it names rather than the spec's word for it, and ask in rounds about those it left under Not checked, confirming the spec's words each draft will quote as its requirements, and settling the design questions the items leave open. Done when each item's status is checked or ruled by the user, the requirements confirmed, and no design question open\n3. Draft plans grouped by what ships together and split by repository, each written as `new` writes one but not started, its items carrying the spec's detail and its `## Requirements` quoting the spec items it plans, as confirmed. Run the review loop over every draft at once, so the drafts and the spec they name are one corpus: contradictions between items, and artefacts no item produces or consumes. Done when the loop's summary is in hand\n4. Propose the plans, each with its items; where every open item went, left-out ones included; and the loop's summary and open questions. Done when the user has answered\n5. After the user answers, delete every draft not approved, all of them on a rejection. Apply the user's changes to the rest. Where any draft's items changed, re-audit every remaining draft with `ele:plan-audit`'s Drafts mode, as the loop's Audit step runs it, and put its findings and questions to the user in a round before starting; and in the change that lands them remove the spec's \"still to build\" and \"not built\" prose for their items only, leaving the design. Start the plans in this repository; for another, print the `ele plan start` command to run there. Done when no unapproved draft remains, every re-audit finding is answered, and every approved plan is started or its command printed\n\n## Review loop\n\n`new` and `convert` run it over their unstarted drafts, before any is shown for approval: an author reads its own draft as confirmation. At least 2 passes and at most 3. Before pass 1, run `mktemp -d` alone and copy each draft into the directory it prints, `\u003Cbase>` below, for the diff at the end; write that path into each later command, since Claude Code's Bash tool keeps no variable between calls. Each pass:\n\n1. **Audit.** Run the `ele:plan-audit` skill as `drafts \u003Cpath>...`, every draft's path, marking this checkout's repository ours and current; a draft for another repository, from `convert`'s split, leaves its repository's question to the audit. It adds every active plan of the repository itself and asks nothing. Invoke it afresh every pass: a revision is unaudited until the skill has run over it, and its text loaded in an earlier pass is not a run. Done when its report is in hand\n2. **Review.** Launch one `ele:plan-reviewer` agent a draft, all in one message. Brief each by absolute path only: `reviewers\u002Fplan.md` beside this file, the reviewer file to follow; the draft; its `spec:`; the user's `CLAUDE.md`, under `$CLAUDE_CONFIG_DIR` when set, else `~\u002F.claude`, and the repository's; each ADR Grounding found; `reference\u002Fplan-practices.md`; and from pass 2 the settled file below. Never this session's reasoning, its account of the draft or what changed: the review leans on the reviewer's own reading. Done when every reviewer has reported\n3. **Ids.** Number each finding `A\u003Cpass>.\u003Cn>` for the audit's and `R\u003Cpass>.\u003Cn>` for the reviewers', in report order across the pass. A finding matching a settled one keeps its id and its level, unless it names evidence that pass did not have. Done when every finding has an id and a level\n4. **Verdicts.** Before any edit, give each finding a verdict with its reason: `fixed`, naming the revision; `refuted`, naming the evidence it is wrong; `deliberate`, saying why it stays; or `open`, a question for the user. A finding against a requirement, and a requirement question, is `open`: the loop never edits `## Requirements`, since editing one to clear a finding makes the check circular. Append each to `\u003Cbase>\u002Fsettled.md` with its id, level, verdict and reason. Done when every finding is in the file with its verdict\n5. **Revise.** Make the `fixed` revisions, then `(cd \u003Croot> && ele plan check \u003Cpath>)` each draft, `\u003Croot>` being the checkout holding it: the check refuses a path outside the working directory's checkout. Done when the drafts carry every `fixed` revision and the check lists no warning\n\nAfter pass 2, stop when no blocker or major is `open` and pass 2 fixed no blocker or major, such a fix being unreviewed until pass 3; else run pass 3, the last. Then show the user every finding as a table of id, level, verdict and reason, minors listed without blocking; ask the `open` ones in a round, by the rules of Asking in rounds; and show `diff -u` of each draft from its copy in `\u003Cbase>`. An answer that changes a requirement is the user's edit; one that changes an item is made and shown in a second diff. Done when the user has seen the table and diffs, and every `open` finding is answered\n\n## Pickup\n\nWith no argument. Made for a return after days away, so trust nothing the plan says about the code until it is re-derived.\n\n1. **Which plan.** `ele plan show`. With several active it lists each with its counts and next item: ask which, then `ele plan show \u003Cpath>`. None active: say so and stop. Done when one plan is chosen\n2. **Already done?** Check the next item against the commits and uncommitted changes since the plan's last tick: `git log` and `git diff` on the paths the item names. Work found that delivers it: run `ele:checkpoint` in branch scope, its base the commit that ticked the plan's previous item, or with no tick yet the commit that added the plan file, `git log --diff-filter=A -1 --format=%H -- \u003Cpath>`. The checkpoint proposes the tick after its gate and review; never propose it here. A tick written, stop until the user commits it, then run this step again on the new next item, since an uncommitted tick names no base. No tick proposed, or the tick declined: this item is the one in hand. Done when the item in hand has no work found that delivers it, or its checkpoint proposed no tick\n3. **Task card.** Before any work, show the plan's path, the item's title and body, its blocked reasons, and related notes from `ele recall` on two or three of the item's rarer words, since every word must match. Ask to start. A wrong plan is caught here, before work begins. Propose no edit to the plan here: a stale premise seen now waits for the premise check. Done when the user says to start, or names another item\n4. **Premise check.** On a yes, before building:\n   - Re-derive each claim the item makes about code through one `ele:explorer` agent, reading no code yourself first: brief it as Grounding briefs one, its question the item quoted whole and the plan's path, its angle the item's claims about code. Each claim takes the explorer's cite, as `file:line` or, where the repository refuses line numbers in a plan, its package and identifier; a claim it lists under Not checked is marked \"not checked\"\n   - Re-read each blocked item's reasons and propose removing those now settled\n   - Run `ele plan check \u003Cpath>` and propose a fix for each warning: re-measure a stale `as of` line, drop a done blocker or a passed `after:`, add a missing name or done-when, untick a closing action ticked while another is open or the inbox holds items, triage an old inbox item at step 5, rewrite a line citing ele data (delete provenance, describe what was observed, move a dependency into the plan)\n   - Check that a decision the item relies on lives in one place, and that a status or count it states is dated or derived\n   - Write what you find under the item as dated lines\n   - Where a finding makes the item's title or a `## Context` line wrong, propose the reword as its own question, the old text beside the new, and write it on approval. The yes to start is not approval for it. Declined, the finding stays a dated line only\n   - Done when every claim is cited or marked \"not checked\", every `ele plan check` warning has a proposed fix, and the findings are written under the item\n5. **Inbox.** Triage this plan's inbox by the Inbox triage section of `plan-audit\u002FSKILL.md`, in the skills directory holding this skill's base directory, and apply what the user approves. Done when every inbox item has a proposal and each approved one is applied\n6. **Work the item.** Once the work is done, run `ele:checkpoint`: it judges whether the change completes the item and proposes the tick with its line, after its gate and review. This step never proposes a tick itself; stopping short of done, it proposes a dated progress line. Done when the checkpoint has run, or the progress line is proposed\n\nA session that compacted mid-item re-reads the item from the file before carrying on.\n","# Plan\n\nA plan is a Markdown file under `.documentation\u002Fplans\u002F` in the repository whose code it changes. The file is the list and its checkboxes the state; ele stores only which plans are active. `ele plan show` parses and reports, and `ele plan check` lists the warnings `show` counts; every judgement below is yours. A plan is shared, so never write an ele note id or session into it.\n\nThe format:\n\n- Front matter: `kind: plan`, `created`, and `spec` when the plan implements a spec in the same repository. Nothing derived, such as counts\n- Sections: `## Why`, `## Requirements`, `## Context` (facts, dated), `## Decisions`, `## Actions`, `## Inbox`, `## Out of scope`\n- `## Requirements`, optional, numbered, in the user's words: what the plan must deliver, the list the review loop checks the items against. Only the user changes it. A plan written before the section has none\n- A Context line not verified against the code or data says \"not checked\"\n- An architectural Decision carries a `rejected: \u003Coption>, because …` line for each option it turned down, so an ADR drafted at the close has its considered options\n- An item is a `- [ ]` or `- [x]` line and its indented body\n  - Each action but the closing one leads with a short lowercase kebab name, `- [ ] judge-blocks: …`, written once and never changed, unique across active and closed plans. `ele plan start` refuses a name another active plan holds or a closed plan used, and an action starting `word: ` is read as named `word`; an Inbox line is never named: its leading `word: ` or date is not a name\n  - Each action but the closing one carries a `done when …` line in its body: what will be observable once it is done\n  - Progress goes in the body as dated lines, `as of \u003Cdate>: …` or `next: …`, so a session stopping mid-item needs no handoff\n  - A tick carries one line saying what was asserted, at which boundary and at which commit\n  - A `blocked:` line holds an item back: `blocked: judge-blocks, mining-reports\u002Freport-figures; \u003Creason>` names items in this plan or `\u003Cplan>\u002F\u003Cname>` in another of the same project, `\u003Cplan>` being its file name without `.md`, and `ele plan show` frees the item once each is ticked and flags one done, missing, or left unticked by a closed plan; a line with anything but names before its `;` is prose and holds the item until removed by hand, and `ele plan check` warns on an empty or malformed list and a cycle. ele reads only the plans this project has started or closed, so a `\u003Cplan>` it has neither of is flagged: start a draft it names, and write a wait on a plan in another repository as prose with no comma before its `;`, as a comma there reads as a malformed list\n  - `after: 2026-10-22` or `after: 2026-10-22 19:28Z` (UTC) holds an item until then, and `ele plan show` flags it to drop once passed\n- The last action folds each Decision through `\u002Fele:plan-audit close \u003Cpath>`, then what the plan built, and its out-of-scope lines as a `Not doing:` list in the section the plan implemented, into its spec; with no spec, it proposes each out-of-scope line as a `decision` through `\u002Fele:remember`. Then it runs `ele plan done \u003Cpath>`, which stores the plan's item names and whether each was ticked, and only then deletes the plan, since `done` reads the file. Never tick it while the inbox holds an item; `ele plan check` says when every other action is done, and warns when it is ticked with items there\n- `## Inbox` takes work captured mid-session, each item dated with where it came from\n- `## Out of scope` holds what triage ruled out, one line each: `- \u003Cgist>: \u003Cwhy> (ruled \u003Cdate>, from \u003Csource>)`. A ruled-out item returns only if the plan's why is redrawn\n\n## Asking in rounds\n\n`new` and `convert` settle questions in rounds. A round asks every question whose prerequisites are settled, numbered, each with your recommended answer, worded so that yes accepts it; a question that hangs on one still open waits for a later round, since asked early it is asked again. Facts are yours to find, as Grounding says: ask the user for decisions, and for a fact neither settles. A decision goes in `## Decisions`, with its reason, of each plan that relies on it; a fact about an item goes under that item as a dated line. The rounds end when no question is open and the user confirms the whole; only then does the step that asked them go on. A question `convert` step 1 leaves as a note is not asked\n\n## Grounding\n\nBefore the first round of `new` or `convert`, read the repository's ADRs, found by the ADRs section of `reference\u002Fplan-practices.md` in the plugin root, the directory holding `skills\u002F`. An accepted one is a settled constraint: a question it answers is not asked, and a wish it rules out is said to be so, superseding it being a question of its own. Another status is context. Before the first round, and before a later one resting on facts not yet found, launch `ele:explorer` agents for those facts, one per angle, at most 3, in one message, however small the repository. Read no code yourself before the first round, since a session that reads it launches no explorer, and so runs no recall. Brief each by absolute path: `reference\u002Fexplorer.md` in the plugin root, the file to follow; the user's `CLAUDE.md`, under `$CLAUDE_CONFIG_DIR` when set, else `~\u002F.claude`, and the repository's, the convention files; and give it its question and angle. Where `ele:explorer` is not listed, launch `general-purpose` with the same brief. A fact that goes in `## Context` unverified, including one an explorer lists under Not checked, says \"not checked\".\n\n## Designs\n\nBefore the round of `new` step 1 or `convert` step 2 that puts a design question meeting plan-audit Close mode's architectural test (hard to reverse, surprising without context and the result of a real trade-off, all three), launch two `ele:architect` agents in one message, each under a different constraint pulling that trade-off its own way, such as the smallest change against the shape easiest to extend. Brief each by absolute path: `reference\u002Farchitect.md` in the plugin root, the file to follow; the convention files Grounding names; and give it the explorers' findings the question rests on, the question and its constraint. Where `ele:architect` is not listed, launch `general-purpose` with the same brief. Put the question with each design as an answer, your recommendation first. The chosen design goes in `## Decisions`; each design not chosen, and each other option the user turned down, under it as `rejected: \u003Coption>, because …`, the reason from that design's Costs. A design question failing the test gets no architects, since each is a full design run, on opus for `ele:architect`. Neither does one raised once a draft exists, as an `open` review-loop finding or in `convert` step 5's re-audit round: ask it as any other question, since by then a draft exists whose design the question can be weighed against.\n\n## new \u003Cname>\n\n1. Ground the rounds, then settle in rounds whatever of the why, the requirements, the items and their design questions the conversation has not already settled; each decision it already made goes in `## Decisions` with its reason. Done when the user has confirmed the whole\n2. Write `.documentation\u002Fplans\u002F\u003Cname>.md` from `references\u002Ftemplate.md`, with today's date and the plan's own path in the closing action. `## Requirements` holds what the rounds confirmed the plan must deliver, numbered, in the user's words; each architectural Decision its `rejected:` lines, as Designs says. With no spec, drop the `spec` line and replace the closing action's \"then what this plan built and its out-of-scope lines as `Not doing:` into the spec\" with \"propose each out-of-scope line as a `decision` through `\u002Fele:remember`\". Done when `ele plan check` on the file lists no warning\n3. Run the review loop over the file. Done when the loop's summary is shown and its questions answered\n4. Show the file, then on approval run `ele plan start .documentation\u002Fplans\u002F\u003Cname>.md`. Done when it prints the plan as started, or the user has declined\n\n## convert \u003Cspec>\n\n1. List the spec's open items with their sections. Leave out those already in a plan, and those gated on a measurement or an undecided question: they stay notes, a `question` proposed through `\u002Fele:remember` where none holds one. Done when every open item is listed as to plan, already planned, or left as a note\n2. Ground the rounds, one explorer's angle being each item's status against the code or data it names rather than the spec's word for it, and ask in rounds about those it left under Not checked, confirming the spec's words each draft will quote as its requirements, and settling the design questions the items leave open. Done when each item's status is checked or ruled by the user, the requirements confirmed, and no design question open\n3. Draft plans grouped by what ships together and split by repository, each written as `new` writes one but not started, its items carrying the spec's detail and its `## Requirements` quoting the spec items it plans, as confirmed. Run the review loop over every draft at once, so the drafts and the spec they name are one corpus: contradictions between items, and artefacts no item produces or consumes. Done when the loop's summary is in hand\n4. Propose the plans, each with its items; where every open item went, left-out ones included; and the loop's summary and open questions. Done when the user has answered\n5. After the user answers, delete every draft not approved, all of them on a rejection. Apply the user's changes to the rest. Where any draft's items changed, re-audit every remaining draft with `ele:plan-audit`'s Drafts mode, as the loop's Audit step runs it, and put its findings and questions to the user in a round before starting; and in the change that lands them remove the spec's \"still to build\" and \"not built\" prose for their items only, leaving the design. Start the plans in this repository; for another, print the `ele plan start` command to run there. Done when no unapproved draft remains, every re-audit finding is answered, and every approved plan is started or its command printed\n\n## Review loop\n\n`new` and `convert` run it over their unstarted drafts, before any is shown for approval: an author reads its own draft as confirmation. At least 2 passes and at most 3. Before pass 1, run `mktemp -d` alone and copy each draft into the directory it prints, `\u003Cbase>` below, for the diff at the end; write that path into each later command, since Claude Code's Bash tool keeps no variable between calls. Each pass:\n\n1. **Audit.** Run the `ele:plan-audit` skill as `drafts \u003Cpath>...`, every draft's path, marking this checkout's repository ours and current; a draft for another repository, from `convert`'s split, leaves its repository's question to the audit. It adds every active plan of the repository itself and asks nothing. Invoke it afresh every pass: a revision is unaudited until the skill has run over it, and its text loaded in an earlier pass is not a run. Done when its report is in hand\n2. **Review.** Launch one `ele:plan-reviewer` agent a draft, all in one message. Brief each by absolute path only: `reviewers\u002Fplan.md` beside this file, the reviewer file to follow; the draft; its `spec:`; the user's `CLAUDE.md`, under `$CLAUDE_CONFIG_DIR` when set, else `~\u002F.claude`, and the repository's; each ADR Grounding found; `reference\u002Fplan-practices.md`; and from pass 2 the settled file below. Never this session's reasoning, its account of the draft or what changed: the review leans on the reviewer's own reading. Done when every reviewer has reported\n3. **Ids.** Number each finding `A\u003Cpass>.\u003Cn>` for the audit's and `R\u003Cpass>.\u003Cn>` for the reviewers', in report order across the pass. A finding matching a settled one keeps its id and its level, unless it names evidence that pass did not have. Done when every finding has an id and a level\n4. **Verdicts.** Before any edit, give each finding a verdict with its reason: `fixed`, naming the revision; `refuted`, naming the evidence it is wrong; `deliberate`, saying why it stays; or `open`, a question for the user. A finding against a requirement, and a requirement question, is `open`: the loop never edits `## Requirements`, since editing one to clear a finding makes the check circular. Append each to `\u003Cbase>\u002Fsettled.md` with its id, level, verdict and reason. Done when every finding is in the file with its verdict\n5. **Revise.** Make the `fixed` revisions, then `(cd \u003Croot> && ele plan check \u003Cpath>)` each draft, `\u003Croot>` being the checkout holding it: the check refuses a path outside the working directory's checkout. Done when the drafts carry every `fixed` revision and the check lists no warning\n\nAfter pass 2, stop when no blocker or major is `open` and pass 2 fixed no blocker or major, such a fix being unreviewed until pass 3; else run pass 3, the last. Then show the user every finding as a table of id, level, verdict and reason, minors listed without blocking; ask the `open` ones in a round, by the rules of Asking in rounds; and show `diff -u` of each draft from its copy in `\u003Cbase>`. An answer that changes a requirement is the user's edit; one that changes an item is made and shown in a second diff. Done when the user has seen the table and diffs, and every `open` finding is answered\n\n## Pickup\n\nWith no argument. Made for a return after days away, so trust nothing the plan says about the code until it is re-derived.\n\n1. **Which plan.** `ele plan show`. With several active it lists each with its counts and next item: ask which, then `ele plan show \u003Cpath>`. None active: say so and stop. Done when one plan is chosen\n2. **Already done?** Check the next item against the commits and uncommitted changes since the plan's last tick: `git log` and `git diff` on the paths the item names. Work found that delivers it: run `ele:checkpoint` in branch scope, its base the commit that ticked the plan's previous item, or with no tick yet the commit that added the plan file, `git log --diff-filter=A -1 --format=%H -- \u003Cpath>`. The checkpoint proposes the tick after its gate and review; never propose it here. A tick written, stop until the user commits it, then run this step again on the new next item, since an uncommitted tick names no base. No tick proposed, or the tick declined: this item is the one in hand. Done when the item in hand has no work found that delivers it, or its checkpoint proposed no tick\n3. **Task card.** Before any work, show the plan's path, the item's title and body, its blocked reasons, and related notes from `ele recall` on two or three of the item's rarer words, since every word must match. Ask to start. A wrong plan is caught here, before work begins. Propose no edit to the plan here: a stale premise seen now waits for the premise check. Done when the user says to start, or names another item\n4. **Premise check.** On a yes, before building:\n   - Re-derive each claim the item makes about code through one `ele:explorer` agent, reading no code yourself first: brief it as Grounding briefs one, its question the item quoted whole and the plan's path, its angle the item's claims about code. Each claim takes the explorer's cite, as `file:line` or, where the repository refuses line numbers in a plan, its package and identifier; a claim it lists under Not checked is marked \"not checked\"\n   - Re-read each blocked item's reasons and propose removing those now settled\n   - Run `ele plan check \u003Cpath>` and propose a fix for each warning: re-measure a stale `as of` line, drop a done blocker or a passed `after:`, add a missing name or done-when, untick a closing action ticked while another is open or the inbox holds items, triage an old inbox item at step 5, rewrite a line citing ele data (delete provenance, describe what was observed, move a dependency into the plan)\n   - Check that a decision the item relies on lives in one place, and that a status or count it states is dated or derived\n   - Write what you find under the item as dated lines\n   - Where a finding makes the item's title or a `## Context` line wrong, propose the reword as its own question, the old text beside the new, and write it on approval. The yes to start is not approval for it. Declined, the finding stays a dated line only\n   - Done when every claim is cited or marked \"not checked\", every `ele plan check` warning has a proposed fix, and the findings are written under the item\n5. **Inbox.** Triage this plan's inbox by the Inbox triage section of `plan-audit\u002FSKILL.md`, in the skills directory holding this skill's base directory, and apply what the user approves. Done when every inbox item has a proposal and each approved one is applied\n6. **Work the item.** Once the work is done, run `ele:checkpoint`: it judges whether the change completes the item and proposes the tick with its line, after its gate and review. This step never proposes a tick itself; stopping short of done, it proposes a dated progress line. Done when the checkpoint has run, or the progress line is proposed\n\nA session that compacted mid-item re-reads the item from the file before carrying on.\n",{"name":5,"description":7,"argument-hint":24,"disable-model-invocation":25},"[new \u003Cname> | convert \u003Cspec>]",true,"main","5f6c9d0393b4eba4532ecd025ccf71628270b49d","plugin\u002Fskills\u002Fplan","https:\u002F\u002Fgithub.com\u002Fnicolaspearson\u002Fclaude-ephant","https:\u002F\u002Fgithub.com\u002Fnicolaspearson","Software Engineer","2026-11-14 17:59:58",{"source":8,"skillName":5,"commands":34,"prompt":44,"zipAvailable":25},[35,38,41],{"tool":36,"command":37},"npx","npx skills add nicolaspearson\u002Fclaude-ephant --skill plan",{"tool":39,"command":40},"bunx","bunx skills add nicolaspearson\u002Fclaude-ephant --skill plan",{"tool":42,"command":43},"pnpm","pnpm dlx skills add nicolaspearson\u002Fclaude-ephant --skill plan","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fnicolaspearson\u002Fclaude-ephant\u002Fblob\u002Fmain\u002Fplugin\u002Fskills\u002Fplan\u002FSKILL.md",[46,52],{"path":47,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":50,"sourceHtmlUrl":51},"references\u002Ftemplate.md","REFERENCE","参考文档",1010,"https:\u002F\u002Fgithub.com\u002Fnicolaspearson\u002Fclaude-ephant\u002Fblob\u002Fmain\u002Fplugin\u002Fskills\u002Fplan\u002Freferences\u002Ftemplate.md",{"path":53,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":54,"sourceHtmlUrl":55},"reviewers\u002Fplan.md",3623,"https:\u002F\u002Fgithub.com\u002Fnicolaspearson\u002Fclaude-ephant\u002Fblob\u002Fmain\u002Fplugin\u002Fskills\u002Fplan\u002Freviewers\u002Fplan.md",[57,64,71,78,85,92],{"skillNo":58,"name":59,"nameZh":6,"description":60,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":61,"categories":62,"keywords":63,"publishedAt":19},"01M7EX7BDBNMZQ1CWZ50PS2J4E","wrap","Wrap up a session by asking what to keep as notes, plan inbox items, progress lines or a recipe, which due reminders to close, and whether to hand off. Use when the user asks to wrap up or end a session, or asks \"anything to remember or hand off?\"; never start one unasked.","https:\u002F\u002Fgithub.com\u002Fnicolaspearson\u002Fclaude-ephant\u002Fblob\u002Fmain\u002Fplugin\u002Fskills\u002Fwrap\u002FSKILL.md",[],[],{"skillNo":65,"name":66,"nameZh":6,"description":67,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":68,"categories":69,"keywords":70,"publishedAt":19},"01M7EX77RDKPNA4FFKNVW1J53B","ticket-audit","Audit a tracker's ticket map against the code for stale premises and contradictions. Use when about to pick up a ticket or close a map, or when asked for a ticket audit; for ele plans, use ele:plan-audit.","https:\u002F\u002Fgithub.com\u002Fnicolaspearson\u002Fclaude-ephant\u002Fblob\u002Fmain\u002Fplugin\u002Fskills\u002Fticket-audit\u002FSKILL.md",[],[],{"skillNo":72,"name":73,"nameZh":6,"description":74,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":75,"categories":76,"keywords":77,"publishedAt":19},"01M7EX74KXXZV1710EX5R4NPKA","setup","Set up ele after installing it, or in a directory no owner covers, by proposing the ele owner commands that map it and running them once the user approves. Use when the user asks to set up or configure ele, when the start block says no owner covers the directory, or when ele status names no owner here.","https:\u002F\u002Fgithub.com\u002Fnicolaspearson\u002Fclaude-ephant\u002Fblob\u002Fmain\u002Fplugin\u002Fskills\u002Fsetup\u002FSKILL.md",[],[],{"skillNo":79,"name":80,"nameZh":6,"description":81,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":82,"categories":83,"keywords":84,"publishedAt":19},"01M7EX71QV7AT6Z64MGMKQEXQN","review","Review a merge request or pull request, a local branch or a commit with checkpoint's reviewers, and walk through its major findings, then any minor ones you pick, to publish, copy or skip. Uncommitted work is checkpoint's.","https:\u002F\u002Fgithub.com\u002Fnicolaspearson\u002Fclaude-ephant\u002Fblob\u002Fmain\u002Fplugin\u002Fskills\u002Freview\u002FSKILL.md",[],[],{"skillNo":86,"name":87,"nameZh":6,"description":88,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":89,"categories":90,"keywords":91,"publishedAt":19},"01M7EX6YVAP91CEV94VZ71VSNV","remember","Save a decision, an open question, a dated fact or a reminder for a later date so later sessions start with it. Use when the user says \"remember this\", \"we decided…\" or \"remind me on…\", or a skill or ele's Stop hook asks for notes.","https:\u002F\u002Fgithub.com\u002Fnicolaspearson\u002Fclaude-ephant\u002Fblob\u002Fmain\u002Fplugin\u002Fskills\u002Fremember\u002FSKILL.md",[],[],{"skillNo":93,"name":94,"nameZh":6,"description":95,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":96,"categories":97,"keywords":98,"publishedAt":19},"01M7EX6VECCCEGHWP6164WNEC4","recipe","Propose a reusable multi-step procedure this session ran as a user skill, and write it once the user approves; with audit, check the user's skills for commands, flags and versions that no longer exist. Use when the user asks to save a procedure as a recipe or skill, or to audit their skills for stale commands; \u002Fele:wrap offers a recipe too; never start one unasked.","https:\u002F\u002Fgithub.com\u002Fnicolaspearson\u002Fclaude-ephant\u002Fblob\u002Fmain\u002Fplugin\u002Fskills\u002Frecipe\u002FSKILL.md",[],[]]