roadmodel 0.2.85__tar.gz → 0.2.86__tar.gz

{roadmodel-0.2.85 → roadmodel-0.2.86}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: roadmodel
-Version: 0.2.85
+Version: 0.2.86
 Summary: BYO-key CLI that recommends the right AI model, platform, and settings for a prompt.
 Project-URL: Homepage, https://roadmodel.ai
 Project-URL: Repository, https://github.com/nathanramoscfa/roadmodel
@@ -607,7 +607,8 @@ server and the reasoning-effort default in parity automatically.
 | `roadmodel recommend --model`    | Override the specific model ID on the chosen provider.                             |
 | `roadmodel recommend --user-context PATH` | Override the user-context.md location for this invocation.                |
 | `roadmodel score --category C --complexity X` | Rank (model, platform, effort) candidates in code for a classified task — no engine call. See [docs/scoring-model.md](docs/scoring-model.md). |
-| `roadmodel backup --category C --complexity X --model M --effort E` | Name the cross-maker backup for a pick, by the rule /recommend uses — no engine call. Prints the roadmap Settings Backup cell, and first a `MODEL` cell when a newer version of the pick's model (same platform, no dearer) scores level or better at its effort. |
+| `roadmodel backup --category C --complexity X --model M --effort E` | Name the cross-maker backup for a pick, by the rule /recommend uses — no engine call. Prints the roadmap Settings Backup cell, and first a `MODEL` cell when the pick's platform offers a newer version of its model (GPT-6 Sol → GPT-6.1 Sol). |
+| `roadmodel ladder --category C --complexity X [--novel] [--phase]` | Print a roadmap Settings table as the full ladder: Cost / Balanced / Quality, each with its effort and a cross-maker backup, plus the column the step runs and the rule for a tight or exhausted pool — no engine call. |
 | `roadmodel catalog show`         | Print the bundled `model-selector.txt` (use `--doc tier-cost-scale` for the price doc). |
 | `roadmodel catalog path`         | Print the on-disk path of the bundled catalog document (same `--doc` flag).        |
 | `roadmodel context show`         | Print the resolved `user-context.md` file.                                         |

{roadmodel-0.2.85 → roadmodel-0.2.86}/README.md

@@ -356,7 +356,8 @@ server and the reasoning-effort default in parity automatically.
 | `roadmodel recommend --model`    | Override the specific model ID on the chosen provider.                             |
 | `roadmodel recommend --user-context PATH` | Override the user-context.md location for this invocation.                |
 | `roadmodel score --category C --complexity X` | Rank (model, platform, effort) candidates in code for a classified task — no engine call. See [docs/scoring-model.md](docs/scoring-model.md). |
-| `roadmodel backup --category C --complexity X --model M --effort E` | Name the cross-maker backup for a pick, by the rule /recommend uses — no engine call. Prints the roadmap Settings Backup cell, and first a `MODEL` cell when a newer version of the pick's model (same platform, no dearer) scores level or better at its effort. |
+| `roadmodel backup --category C --complexity X --model M --effort E` | Name the cross-maker backup for a pick, by the rule /recommend uses — no engine call. Prints the roadmap Settings Backup cell, and first a `MODEL` cell when the pick's platform offers a newer version of its model (GPT-6 Sol → GPT-6.1 Sol). |
+| `roadmodel ladder --category C --complexity X [--novel] [--phase]` | Print a roadmap Settings table as the full ladder: Cost / Balanced / Quality, each with its effort and a cross-maker backup, plus the column the step runs and the rule for a tight or exhausted pool — no engine call. |
 | `roadmodel catalog show`         | Print the bundled `model-selector.txt` (use `--doc tier-cost-scale` for the price doc). |
 | `roadmodel catalog path`         | Print the on-disk path of the bundled catalog document (same `--doc` flag).        |
 | `roadmodel context show`         | Print the resolved `user-context.md` file.                                         |

{roadmodel-0.2.85 → roadmodel-0.2.86}/docs/claude-commands/roadmap-refresh.md

@@ -167,39 +167,42 @@ Print the phase · step · status table.
 
 The current step is the first step, in phase order, that does not read
 `Complete`. For the current step and every step after it, in every phase
-roadmap that exists, re-run the model selector exactly as the phase
-roadmap was written: `planning/model-selector.txt`, with prices from
-`planning/model-tier-cost-scale.md` and display rules from
-`planning/settings-display.md`, against `planning/user-context.md`. You
-are the engine — do not call any external API. Honor every availability
-exclusion in the selector, and include a backup model.
+roadmap that exists, re-select its Settings the way a phase roadmap is
+written now. Classify the step (its PRIMARY and SECONDARY category, its
+complexity, novel or not; its Model rationale names them), then:
 
+- Where `roadmodel` is installed, the Settings table is the ladder
+  `roadmodel ladder --category <c> --complexity <x> [--novel]` prints
+  for that classification (MCP tool: `roadmap_settings`), pasted as
+  printed with its **Runs:** line: Cost, Balanced and Quality, each with
+  its own effort and a backup from another maker, read with every usage
+  pool at headroom, so the step's runner picks the column from the pools
+  on the day. A single-column table becomes the ladder: that is a
+  difference. Its models are always the latest version of their series.
+- Where `roadmodel` cannot be reached, re-run the model selector as the
+  roadmap was written: `planning/model-selector.txt`, with prices from
+  `planning/model-tier-cost-scale.md` and display rules from
+  `planning/settings-display.md`, against `planning/user-context.md`.
+  You are the engine — do not call any external API. Honor every
+  availability exclusion in the selector, include a backup model, and
+  never name an older version of a series (GPT-6 Sol where GPT-6.1 Sol
+  runs, Opus 5 where Opus 5.5 does).
 - If the result differs from the step's Settings table, rewrite the
   table and its Model rationale, and add one line beneath the table:
   `> Settings updated <YYYY-MM-DD> (refresh): was <old model> · <old effort>. <why>.`
-  A changed backup alone counts as a difference; name it in that line.
-- Every table carries a `Backup` row directly under `Model`: the
-  selector's BACKUP with its own platform and dial in one cell
-  (`GPT-5.6 Terra — Codex · Intelligence Medium`), or `None`. Where
-  `roadmodel` is installed, the cell is the one `roadmodel backup
-  --category <c> --complexity <x> [--novel] --model <the step's model>
-  --effort <the step's effort>` prints for the step's classification —
-  the rule /recommend backs a pick up by, which takes the cheaper of two
-  substitutes the benchmarks cannot tell apart — and it outranks the
-  selector's BACKUP. When the command also prints a `MODEL` line, that
-  cell replaces the step's model — a newer version of the same model
-  (GPT-6 Sol → GPT-6.1 Sol), on the same platform, no dearer, that
-  scores level or better at the step's effort; the selector's letters
-  are measured at each model's top effort, and a step that runs lower
-  can reverse them. That is a difference: rewrite the table, its Model
-  rationale and its selection block with the `MODEL` cell, keep the
-  Backup cell printed with it, and add the `Settings updated` line.
-  A table written before roadmodel 0.2.41 has no such row: add it from the
+  For a ladder, the old model and effort are the column the step ran
+  (its **Runs:** line), or the single column it replaced. A changed
+  backup alone counts as a difference; name it in that line.
+- Every table carries a `Backup` row directly under `Model`, each cell
+  a backup with its own platform and dial
+  (`GPT-6.1 Sol — Codex · Intelligence Medium`), or `None`. A table
+  written before roadmodel 0.2.41 has no such row: add it from the
   backup the step already names (its Model rationale's closing
   sentence, or its entry in the Model selection blocks) — a layout
   fix, not a new pick, so it takes no `Settings updated` line.
-- Keep the step's `BACKUP:` line in the Model selection blocks equal to
-  its table's Backup row.
+- Keep the step's selection block in the Model selection blocks equal
+  to the column its **Runs:** line names (its `BACKUP:` line to that
+  column's Backup), and its Summary Table row to the same column.
 - If it matches, leave the step exactly as it is.
 - **Never touch a step that reads `Complete`.** Its settings are the
   record of what ran.
@@ -214,18 +217,20 @@ roadmap settings** table (roadmodel 0.2.54 on): the model and dials for
 the session that writes that phase's roadmap with `/roadmap-phase N`.
 For every phase whose roadmap does not exist yet — no
 `phaseNN-roadmap.md` or `phaseNN-<slug>-roadmap.md` for it — re-run the
-selector for that session, quality first: the strongest `planning`
-model the operator's access methods reach, at the deepest reasoning
-rung it exposes on its platform (`Max` in Claude Code), thinking on,
-never `Ultracode`, with a backup on its own platform and dial — unless
-`planning/user-context.md` states a rule for roadmap-writing sessions,
-which wins. Where `roadmodel` is installed, take the model and the
-backup, not their effort, from the PRIMARY and BACKUP lines of
-`roadmodel score --category planning --complexity high --budget best`:
-it ranks on measured benchmarks, so a stale catalog letter cannot steer
-them, and a model the user-context names for these sessions still wins.
-The table takes its platform's shape, as a step's does, with
-Conversation **New**, followed by one or two sentences on why.
+selector for that session, quality first: it runs the Quality column,
+the strongest `planning` model the operator's access methods reach, at
+the deepest reasoning rung it exposes on its platform (`Max` in Claude
+Code), thinking on, never `Ultracode`, with a backup on its own platform
+and dial — unless `planning/user-context.md` states a rule for
+roadmap-writing sessions, which wins. Where `roadmodel` is installed,
+the table is the ladder `roadmodel ladder --category planning
+--complexity high --phase` prints, pasted as printed with its **Runs:**
+line: it ranks on measured benchmarks, so a stale catalog letter cannot
+steer it, and a model the user-context names for these sessions still
+wins. A single-column table becomes the ladder: that is a difference.
+Without roadmodel, the table takes its platform's single-column shape,
+as a step's does. Conversation is **New**, followed by one or two
+sentences on why.
 
 - No table yet ⇒ add one. It is the phase's first pick, so it takes no
   `Settings updated` line.

{roadmodel-0.2.85/roadmodel/data → roadmodel-0.2.86/docs/claude-commands}/roadmap-step.md

@@ -101,6 +101,23 @@ Print one line: `Step {{STEP}} requires: Model <M> · Platform <P> ·
 <the dials the table carries, e.g. Effort <E> · Thinking <T>, or
 Intelligence <I>>. This session: <your own model> on <this surface>.`
 
+- **Ladder tables** (Cost / Balanced / Quality columns, roadmodel 0.2.86
+  on) carry every column the step may run, so pick one before the checks
+  below. The table's **Runs:** line names the column. Apply its rule to
+  the Usage-pool status table in `planning/user-context.md` (a row whose
+  reset time has passed reads `headroom`). Read the pool of the
+  subscription that pays for that column's platform:
+  - `headroom`: run that column.
+  - `tight`: run that column's Backup when the Backup's pool reads
+    `headroom`, else the Cost column.
+  - `exhausted`: run that column's Backup.
+
+  The chosen column (or Backup) is what the printed line names as
+  required, with the column named (`… · Balanced column`, or `… ·
+  Balanced column's Backup — Claude pool tight`). Any column's model on
+  that column's platform, or any column's Backup on its own platform,
+  continues too. Say which column you are, and which column the rule
+  selected.
 - **Backup.** The table's `Backup` row (roadmaps from roadmodel 0.2.41
   on) names the model to run when the primary is unavailable, with its
   own platform and dial. If you are that backup model on that platform,
@@ -155,9 +172,13 @@ Intelligence <I>>. This session: <your own model> on <this surface>.`
   rollout's `"effort"` is the dial the session runs on, so print it. If
   it differs from the table's, name the value to set. Continue.
 
-**Record what ran.** If the model or effort you proceed with differs from
-the table, the step's own PR rewrites its Settings table to match (§5),
-with one line beneath it:
+**Record what ran.** With a ladder table, running one of its columns or
+Backups changes nothing in it: the step's own PR adds one line beneath
+it, `> Ran <YYYY-MM-DD>: <column> — <model · effort>. <why>.` (why: the
+**Runs:** column, or the pool that moved it). When what ran is in none of
+its cells, or the table has a single column, and the model or effort you
+proceed with differs from the table, the step's own PR rewrites its
+Settings table to match (§5), with one line beneath it:
 `> Settings updated <YYYY-MM-DD>: was <old model> · <old effort>. <why>.`
 The roadmap on `main` is the ledger, so it must say what actually ran.
 Never rewrite the table of a step that reads `Complete`: that is history.

{roadmodel-0.2.85/roadmodel/data → roadmodel-0.2.86/docs/templates}/phase-roadmap-template.md

@@ -737,6 +737,30 @@ that owns it.}}
      "Sonnet 4.6 via Claude Code" reads very differently
      from "Sonnet 4.6 via Anthropic API" on the cost line.
 
+     THE LADDER — the default wherever roadmodel is
+     installed. A step's Settings table carries the full
+     recommendation /recommend shows: Cost, Balanced and
+     Quality, each with its own model, platform and effort
+     and a backup from another maker. The roadmap then
+     outlives a usage pool's state: when a pool reads
+     `tight` or `exhausted`, the step's runner picks
+     another column (the **Runs:** line's rule) instead of
+     the roadmap being refreshed. Paste the table and its
+     **Runs:** line exactly as `roadmodel ladder --category
+     <c> --complexity <x> [--novel]` prints them for the
+     step's classification (MCP tool: `roadmap_settings`).
+     Rows: Model / Backup / Platform / Effort / Thinking /
+     Conversation. The Effort cell names each column's own
+     dial (`Intelligence Medium` on Codex, `Effort High` on
+     Claude Code); a column whose surface has no such
+     control reads `—` there. Every model is the latest
+     version of its series, never an older one (GPT-6 Sol
+     where GPT-6.1 Sol runs, Opus 5 where Opus 5.5 does).
+     Set Conversation to **Continue** in every column only
+     when the step continues the previous one's session.
+     The single-column shapes below are the fallback where
+     roadmodel cannot be reached.
+
      SETTINGS TABLE SHAPE IS PLATFORM-SPECIFIC. Each
      surface exposes a different tuning UI, and the table
      MUST mirror what the operator will actually see in
@@ -824,6 +848,24 @@ that owns it.}}
      cannot apply.
      -->
 
+Settings table — ladder ({{the default wherever roadmodel is
+installed: the table and **Runs:** line `roadmodel ladder
+--category <c> --complexity <x> [--novel]` prints for the
+step's classification, pasted as printed. The three shapes
+after it are the fallback where roadmodel cannot be
+reached}}):
+
+| Setting      | Cost                        | Balanced                    | Quality                     |
+| ------------ | --------------------------- | --------------------------- | --------------------------- |
+| Model        | {{Model}}                   | {{Model}}                   | {{Model}}                   |
+| Backup       | {{Backup — platform · dial}} | {{Backup — platform · dial}} | {{Backup — platform · dial}} |
+| Platform     | {{Access method}}           | {{Access method}}           | {{Access method}}           |
+| Effort       | {{Its dial and level}}      | {{Its dial and level}}      | {{Its dial and level}}      |
+| Thinking     | {{On, or — with no toggle}} | {{On, or — with no toggle}} | {{On, or — with no toggle}} |
+| Conversation | **{{New or Continue}}**     | **{{New or Continue}}**     | **{{New or Continue}}**     |
+
+**Runs:** {{Cost / Balanced / Quality}} — the user-context's Budget priority is `{{cheap / balanced / best}}`. If that column's usage pool reads `tight`, run its Backup when the Backup's pool has headroom, else the Cost column; if it reads `exhausted`, run its Backup.
+
 Settings table — Effort + Thinking variant ({{use this
 shape when PLATFORM is "Claude Code" or any other
 reasoning-dial surface that is not Codex — ChatGPT app,
@@ -1228,29 +1270,28 @@ hygiene").}}
 
 **Deploys:** {{surface → environment, or "nothing beyond merge"}}
 
-{{Settings table — pick the variant that matches PLATFORM
-per the Step 1 guidance, and carry ONLY the rows that
-platform's panel exposes. Claude Code and the other
-reasoning-dial surfaces use Model / Platform / Effort /
-Thinking (On/Off) / Conversation. Codex uses Model /
-Platform / Intelligence / Conversation (no Max Mode, no
-Thinking row). Cursor uses Model / Platform / Max Mode /
-Conversation (no Effort row, no Thinking row). The shape
-below is the Effort + Thinking variant — swap it wholesale if
-PLATFORM is something else; never keep a row for a dial
-the surface does not have.}}
-
-| Setting      | Value                          |
-| ------------ | ------------------------------ |
-| Model        | {{Model name}}                 |
-| Backup       | {{Backup model — its platform · its dial, or None}} |
-| Platform     | {{Access method name}}         |
-| Effort       | {{Low/Medium/High/Extra High/Max/Ultracode}} |
-| Thinking     | {{On or Off}}                  |
-| Conversation | **{{New or Continue}}**        |
-
-**Model rationale:** 3-5 sentences. Name the PLATFORM and
-the subscription or API key that pays for it. For Claude
+{{Settings table — the ladder `roadmodel ladder` prints for
+this step's classification (the Step 1 guidance), pasted as
+printed with its **Runs:** line. Where roadmodel cannot be
+reached, the single-column variant that matches PLATFORM,
+carrying ONLY the rows that platform's panel exposes.}}
+
+| Setting      | Cost                        | Balanced                    | Quality                     |
+| ------------ | --------------------------- | --------------------------- | --------------------------- |
+| Model        | {{Model}}                   | {{Model}}                   | {{Model}}                   |
+| Backup       | {{Backup — platform · dial}} | {{Backup — platform · dial}} | {{Backup — platform · dial}} |
+| Platform     | {{Access method}}           | {{Access method}}           | {{Access method}}           |
+| Effort       | {{Its dial and level}}      | {{Its dial and level}}      | {{Its dial and level}}      |
+| Thinking     | {{On, or — with no toggle}} | {{On, or — with no toggle}} | {{On, or — with no toggle}} |
+| Conversation | **{{New or Continue}}**     | **{{New or Continue}}**     | **{{New or Continue}}**     |
+
+**Runs:** {{the line `roadmodel ladder` prints}}
+
+**Model rationale:** 3-5 sentences. With the ladder: the
+classification it was printed for (PRIMARY and SECONDARY
+category, complexity, novel or not) and why, then what sets
+the three columns apart for this step. Otherwise: name the
+PLATFORM and the subscription or API key that pays for it. For Claude
 Code and the other reasoning-dial surfaces, state why
 EFFORT is at the stated level and why THINKING is on or
 off. For Codex, state why INTELLIGENCE is at the stated
@@ -1366,23 +1407,21 @@ mechanically rather than left to memory.
 **Deploys:** nothing beyond merge — the CI matrix entry is live
 on merge.
 
-{{Settings table — pick the variant that matches PLATFORM
-per the Step 1 guidance, carrying ONLY the rows that
-platform exposes. There are exactly three shapes: the
-Effort + Thinking variant (Claude Code and every other
-reasoning-dial surface), the Codex variant (Intelligence
-alone — no Max Mode, no Thinking row), and the Cursor
-variant (Max Mode alone — no Effort row, no Thinking row).
-The shape below is the Effort + Thinking variant.}}
+{{Settings table — the ladder `roadmodel ladder` prints for
+this step's classification, pasted as printed with its
+**Runs:** line; where roadmodel cannot be reached, the
+single-column variant that matches PLATFORM.}}
 
-| Setting      | Value                          |
-| ------------ | ------------------------------ |
-| Model        | {{Model}}                      |
-| Backup       | {{Backup model — its platform · its dial, or None}} |
-| Platform     | {{Access method name}}         |
-| Effort       | {{Low/Medium/High/Extra High/Max/Ultracode}} |
-| Thinking     | {{On or Off}}                  |
-| Conversation | **New**                        |
+| Setting      | Cost                        | Balanced                    | Quality                     |
+| ------------ | --------------------------- | --------------------------- | --------------------------- |
+| Model        | {{Model}}                   | {{Model}}                   | {{Model}}                   |
+| Backup       | {{Backup — platform · dial}} | {{Backup — platform · dial}} | {{Backup — platform · dial}} |
+| Platform     | {{Access method}}           | {{Access method}}           | {{Access method}}           |
+| Effort       | {{Its dial and level}}      | {{Its dial and level}}      | {{Its dial and level}}      |
+| Thinking     | {{On, or — with no toggle}} | {{On, or — with no toggle}} | {{On, or — with no toggle}} |
+| Conversation | **New**                     | **New**                     | **New**                     |
+
+**Runs:** {{the line `roadmodel ladder` prints}}
 
 **Model rationale:** Why this model fits the mechanical
 translation of the prior verify-script template into this
@@ -1868,7 +1907,12 @@ corresponding row below.
 
 ## Summary Table
 
-<!-- Summary Table column "Reasoning dial" carries
+<!-- With ladder Settings tables, each step's row shows the
+     column its **Runs:** line names (the Model, Platform
+     and dial of that column); the other two columns live
+     in the step's own table.
+
+     Summary Table column "Reasoning dial" carries
      whichever dial the step's PLATFORM exposes, written
      in the surface's native vocabulary so reviewers can
      audit each row 1:1 against the panel. The per-step
@@ -1928,6 +1972,8 @@ output format for ease of audit by a roadmodel script.
 
 Include this section if your project has a roadmodel
 audit tool that consumes the text format. Skip it if not.
+With ladder Settings tables, each block restates the column
+the step's **Runs:** line names.
 -->
 
 <!-- These blocks follow OUTPUT CONTRACT VERSION 2 of

{roadmodel-0.2.85 → roadmodel-0.2.86}/docs/templates/planning-kit-how-to-use.md

@@ -28,6 +28,19 @@ MCP tool, or web service — the operator's editor session already pays for your
 reasoning (e.g. Claude Code funded by a flat Max plan = $0 marginal). Calling
 out would add cost for no benefit.
 
+**Where `roadmodel` is installed, the Settings table comes from it.** Classify
+the step (step 1 below), then paste the table and **Runs:** line
+`roadmodel ladder --category <c> --complexity <x> [--novel]` prints: the
+Cost / Balanced / Quality picks /recommend shows, each with its effort and a
+backup from another maker. It is local code — no API call, no engine — and it
+reads every usage pool at headroom, so the table outlives a pool's state: when
+a pool reads `tight` or `exhausted`, `/roadmap-step` runs another column
+instead of the roadmap being refreshed. Its models are always the latest
+version of their series (GPT-6.1 Sol, never GPT-6 Sol where both run). The
+project roadmap's phase tables come from `roadmodel ladder --category
+planning --complexity high --phase`. Steps 2–4 are the fallback where
+`roadmodel` cannot be reached.
+
 Concretely, for each roadmap step:
 
 1. **Classify** the step's task with the selector's

{roadmodel-0.2.85/roadmodel/data → roadmodel-0.2.86/docs/templates}/project-roadmap-template.md

@@ -38,25 +38,29 @@ STYLE RULES (the AI MUST follow)
     own tables in the phase roadmap. Pick it with the model
     selector, quality first. A phase roadmap fixes every step's
     scope, task prompt, acceptance criteria and Settings, so it
-    is the phase's highest-leverage session, and it runs once.
-    Take the strongest `planning` model the operator's access
-    methods reach, at the deepest reasoning rung that model
-    exposes on its platform (`Max` in Claude Code, which
-    `/roadmap-phase` sets by itself), with thinking on. Where
-    `roadmodel` is installed, take the two models, not their
-    effort, from the PRIMARY and BACKUP lines of `roadmodel
-    score --category planning --complexity high --budget best`:
-    it ranks on measured benchmarks, so a stale catalog letter
-    cannot steer them, and a model the user-context names for
-    these sessions still wins. Never `Ultracode`: it reasons at
-    `Extra high` and adds parallel agents, which one planning
-    document does not need. A rule
-    for roadmap-writing sessions in the user-context overrides
-    this default. The table takes its platform's shape (the
-    phase roadmap template's Settings variants) with a Backup
-    row, and its Conversation is always **New**. Once the phase
-    roadmap exists, the table is history: only a phase not yet
-    written is ever re-picked.
+    is the phase's highest-leverage session, and it runs once:
+    it runs the Quality column, the strongest `planning` model
+    the operator's access methods reach, at the deepest
+    reasoning rung that model exposes on its platform (`Max`
+    in Claude Code, which `/roadmap-phase` sets by itself),
+    with thinking on. Where `roadmodel` is installed, the table
+    is the ladder `roadmodel ladder --category planning
+    --complexity high --phase` prints, pasted as printed with
+    its **Runs:** line: Cost, Balanced and Quality, each with a
+    backup from another maker, so a pool that reads `tight` or
+    `exhausted` on the day moves the session to another column
+    instead of the table being refreshed. It ranks on measured
+    benchmarks, so a stale catalog letter cannot steer it, and
+    a model the user-context names for these sessions still
+    wins. Never `Ultracode`: it reasons at `Extra high` and
+    adds parallel agents, which one planning document does not
+    need. A rule for roadmap-writing sessions in the
+    user-context overrides this default. Without roadmodel, the
+    table takes its platform's single-column shape (the phase
+    roadmap template's Settings variants) with a Backup row.
+    Its Conversation is always **New**. Once the phase roadmap
+    exists, the table is history: only a phase not yet written
+    is ever re-picked.
   - Every phase carries a `**Status:**` line directly under its
     `### Phase` heading, before its **Goal:** — `Not started`
     when this roadmap is written. The phase's FIRST step flips it
@@ -242,19 +246,23 @@ dependency graph in §5 is honoured.
 **Phase roadmap settings** — the session that writes
 `phase01-roadmap.md` (`/roadmap-phase 1`):
 
-| Setting      | Value                                                |
-|--------------|------------------------------------------------------|
-| Model        | <Model name>                                         |
-| Backup       | <Backup model — its platform · its dial, or None>    |
-| Platform     | <Claude Code, or the access method that pays>        |
-| Effort       | <Deepest reasoning rung there — Max in Claude Code>  |
-| Thinking     | On                                                   |
-| Conversation | **New**                                              |
+| Setting      | Cost                   | Balanced               | Quality                                  |
+|--------------|------------------------|------------------------|------------------------------------------|
+| Model        | <Model>                | <Model>                | <The strongest planning model>           |
+| Backup       | <Backup — its platform · its dial> | <Backup — its platform · its dial> | <Backup — its platform · its dial> |
+| Platform     | <Access method>        | <Access method>        | <Claude Code, or the access method that pays> |
+| Effort       | <Its dial and level>   | <Its dial and level>   | <Deepest rung there — Effort Max in Claude Code> |
+| Thinking     | <On, or — with no toggle> | <On, or — with no toggle> | On                                 |
+| Conversation | **New**                | **New**                | **New**                                  |
 
-<!-- The Claude Code shape. Another platform takes its own
-shape from the phase roadmap template's Settings variants:
-Codex carries Intelligence and no Thinking row; Cursor
-carries Max Mode alone. -->
+**Runs:** Quality — a phase-roadmap session runs Quality (quality first). If that column's usage pool reads `tight`, run its Backup when the Backup's pool has headroom, else the Cost column; if it reads `exhausted`, run its Backup.
+
+<!-- The ladder `roadmodel ladder --category planning
+--complexity high --phase` prints, pasted as printed. Without
+roadmodel, a single-column table in its platform's shape from
+the phase roadmap template's Settings variants: Codex carries
+Intelligence and no Thinking row; Cursor carries Max Mode
+alone. -->
 
 <One or two sentences: what this phase's roadmap must get
 right, why this model at this rung, and which subscription or
@@ -298,14 +306,16 @@ key pays for the platform.>
 **Phase roadmap settings** — the session that writes
 `phase02-roadmap.md` (`/roadmap-phase 2`):
 
-| Setting      | Value   |
-|--------------|---------|
-| Model        | …       |
-| Backup       | …       |
-| Platform     | …       |
-| Effort       | …       |
-| Thinking     | On      |
-| Conversation | **New** |
+| Setting      | Cost    | Balanced | Quality |
+|--------------|---------|----------|---------|
+| Model        | …       | …        | …       |
+| Backup       | …       | …        | …       |
+| Platform     | …       | …        | …       |
+| Effort       | …       | …        | …       |
+| Thinking     | …       | …        | On      |
+| Conversation | **New** | **New**  | **New** |
+
+**Runs:** Quality — <the line `roadmodel ladder` prints>
 
 <Why, in one or two sentences.>
 
@@ -333,14 +343,16 @@ key pays for the platform.>
 **Phase roadmap settings** — the session that writes
 `phaseNN-roadmap.md` (`/roadmap-phase N`):
 
-| Setting      | Value   |
-|--------------|---------|
-| Model        | …       |
-| Backup       | …       |
-| Platform     | …       |
-| Effort       | …       |
-| Thinking     | On      |
-| Conversation | **New** |
+| Setting      | Cost    | Balanced | Quality |
+|--------------|---------|----------|---------|
+| Model        | …       | …        | …       |
+| Backup       | …       | …        | …       |
+| Platform     | …       | …        | …       |
+| Effort       | …       | …        | …       |
+| Thinking     | …       | …        | On      |
+| Conversation | **New** | **New**  | **New** |
+
+**Runs:** Quality — <the line `roadmodel ladder` prints>
 
 <Why, in one or two sentences.>
 

{roadmodel-0.2.85/roadmodel/data → roadmodel-0.2.86/docs/templates}/prompt-phase-roadmap.md

@@ -68,9 +68,19 @@ own model> on <this surface>.`
   take a rollout from another directory. Where no source names
   the model, print `unverified`, ask me which I picked, and go on with my
   answer.
+- A ladder table (Cost / Balanced / Quality columns): its **Runs:**
+  line names the column this session runs (Quality, for a phase
+  roadmap). Apply that line's rule to the Usage-pool status table in
+  `planning/user-context.md`: a pool that reads `tight` moves the run
+  to the column's Backup when the Backup's pool has headroom, else to
+  the Cost column; `exhausted` moves it to the column's Backup. That
+  column's Model, Platform and dials are the ones checked below; print
+  which column it is.
 - You are its Model on its Platform, or a newer version in the same
   line (Opus 5 → Opus 5.5, Fable 5 → Fable 5.1, GPT-6 Sol → GPT-6.1
-  Sol): continue, and say so on the line.
+  Sol): continue, and say so on the line. In a ladder table, any
+  column's model on that column's platform continues too; when it is
+  not the column the rule selects, say which one is.
 - You are its Backup on the backup's platform: continue — I have
   switched to it — and say so on the line.
 - Anything else — a different line, an older version, another
@@ -96,27 +106,31 @@ Carry forward anything `{{PRIOR}}` hands to this phase (its closing
 "inherits" line, its Not-in-scope items owned by this phase, and any
 carry-over checklist).
 
-For **each step's** Settings table and Model rationale, run the model
-selector in `@planning/model-selector.txt` (prices from
-`@planning/model-tier-cost-scale.md`, display rules from
-`@planning/settings-display.md`) against `@planning/user-context.md`.
-You are the engine — do not call any external API. Honor every
-availability exclusion in the selector, and write each step's backup
-model into its Settings table's Backup row, with the backup's own
-platform and dial. Where `roadmodel` is installed, the Backup row is
-the cell `roadmodel backup --category <c> --complexity <x> [--novel]
---model <the step's model> --effort <the step's effort>` prints for the
-step's own classification: the rule /recommend backs a pick up by, in
-code, which takes the cheaper of two substitutes the benchmarks cannot
-tell apart. The selector's BACKUP stands only where it is not
-installed. When the command also prints a `MODEL` line, that cell
-replaces the step's model: a newer version of the same model (GPT-6 Sol
-→ GPT-6.1 Sol), on the same platform, no dearer, that scores level or
-better at the step's effort. The selector's letters are measured at each
-model's top effort, and a step that runs lower can reverse them. Write
-the `MODEL` cell into the Model row, the Model rationale and the step's
-selection block, citing the newer model's figures at the step's effort,
-and keep the Backup cell printed with it.
+For **each step's** Settings table, classify the step first: its
+PRIMARY and SECONDARY category, its complexity, and whether it is novel
+problem-solving (the model selector's Steps 1 and 2 in
+`@planning/model-selector.txt`, against `@planning/user-context.md`).
+Where `roadmodel` is installed, the Settings table is the ladder
+`roadmodel ladder --category <c> --complexity <x> [--novel]` prints for
+that classification (MCP tool: `roadmap_settings`), pasted as printed
+with its **Runs:** line. It carries the full recommendation /recommend
+shows: Cost, Balanced and Quality, each with its own effort and a
+backup from another maker. The step's runner picks the column from the
+usage pools on the day, so the roadmap needs no refresh when a pool
+tightens. Its models are always the latest version of their series.
+The Model rationale explains the classification and what sets the
+columns apart for the step; the Summary Table row and the step's
+selection block restate the column the **Runs:** line names.
+
+Where `roadmodel` cannot be reached, run the model selector as the
+engine instead (prices from `@planning/model-tier-cost-scale.md`,
+display rules from `@planning/settings-display.md`) and write the
+single-column variant that matches the step's platform. You are the
+engine — do not call any external API. Honor every availability
+exclusion in the selector, write each step's backup model into its
+Settings table's Backup row with the backup's own platform and dial,
+and never name an older version of a series (GPT-6 Sol where GPT-6.1
+Sol runs, Opus 5 where Opus 5.5 does).
 
 Honor the template's style rules: 80-column prose, zero `{{...}}`
 tokens left, `**Status:** Not started` directly under the title and