feat(usage)!: cost and first rounds per writer and reviewer cell, for the routing charts

README.md

@@ -1471,9 +1471,11 @@ harness, model, and effort, its Reviewer line ending in `, shadow` for the
 shadow, the session and round, the score and rationale, the tier it raised the
 risk to and why, its tally, each finding with
 its location, and the last round's findings with their dispositions. The
-tally, as `- Rounds: 3; findings 4 (high 1, medium 3): fixed 3, refuted 1, open 0`,
-counts the reviewer's rounds on the branch and each finding it raised once,
-by its latest severity, however many rounds keep it open. A finding counts
+tally, as `- Rounds: 3; first round 2 (medium 2); findings 4 (high 1, medium 3): fixed 3, refuted 1, open 0`,
+counts the reviewer's rounds on the branch, its first round's findings as that
+round graded them, and each finding it raised once,
+by its latest severity, however many rounds keep it open. A tally begun before
+it kept first rounds leaves that part out. A finding counts
 for the reviewer that raised it when a new size pick's reviewer fixes or
 refutes it, but one that reviewer keeps open becomes its own, since its round
 reports it again. `--new` drops the findings it leaves unsettled from the
@@ -1590,11 +1592,16 @@ prompt or reply. Since another repository may use the same branch name, the
 session must also be this repository's, as the repositories token-max resolved
 for it on its own host name, or, once its checkout is deleted, those resolved
 there for another session in that exact directory, as a Codex log records
-(never a parent's, which a nested clone may not share); otherwise it stays
+(never a parent's, which a nested clone may not share). A sub-agent or
+spawned session that names no PR's branch, as one in a detached checkout,
+takes the PR of the session that spawned it, unless its own repository is
+known to be another; otherwise it stays
 unattributed. A session is classifying when the PR's `## Review` record names
 it as the [risk classifier](#risk-tiers), reviewing when the record names it as
-a reviewer or Codex marks it a review, implementing when its first request since
-`--since` came before the PR opened, and iterating otherwise, so a session
+a reviewer or Codex marks it a review, its parent's role when both are on the
+same PR, so a review's sub-agents count as reviewing, implementing when its
+first request since `--since` came before the PR opened, and iterating
+otherwise, so a session
 that opened a PR before `--since` and kept working on it reads as iterating. Each PR shows its `## Models` size, its risk tier and
 whether a reviewer raised it, and its implementers, and as `unread:` each Implementer line it cannot read; each reviewer the
 `quest review` record names, with its latest round and score and its tally of
@@ -1622,19 +1629,36 @@ a bot resolving its own thread or a later review leaving it out, and as
 unanswered otherwise.
 
 Rollups follow: spend by size, role, harness, model, and effort; accuracy
-by size and implementer, which sums the tallies but the shadow's, so it counts
+by size and implementer, which sums the tallies but the shadow's, and their
+first rounds over the PRs whose every such tally records one, so it counts
 every finding the routed reviewers caught before merge, while the shadow's
 findings measure an effort step rather than the writer. Since a branch's tallies span every PR from it, only the
 PR whose record holds its latest tally, by the most rounds and then the newest,
 counts them, and a record written before tallies
 counts as unreviewed. An `Implementer:` line's own size, as
-`(iterate, XS)`, counts over the PR's; the writer A/B by size and arm, with
-the same outcomes and every session's spend, counting a sized PR only when
-each of its `Implementer:` lines names the same arm; each risk tier's PRs, how many merged,
-how many a reviewer raised, and how many a later PR's `## Caused by` names,
-which shows whether the tiers predict what breaks; and, over the merged PRs, each online
+`(iterate, XS)`, counts over the PR's. The writer A/B reports by size and arm,
+with the same outcomes and every session's spend, counting a sized PR only when
+each of its `Implementer:` lines names the same arm. Each implementer also carries its
+writing sessions' cost on the merged PRs they were found on, so cost per merged
+PR divides by those PRs, not the ones no host reported: a PR's spend on a
+model goes to its line of that model, of the same effort where both name one,
+and spend on a model no line names, such as a sub-agent on a smaller model,
+to the line that took the most. Spend that fits several lines goes to the
+first line when an implementing session spent it and to the later ones when an
+iterating one did. Local reviewers roll up the same way, by the
+PR's size and each reviewer's harness, model, effort, and shadow mark: PRs
+reviewed, tallies, first rounds, and reviewing cost per PR, where a session
+the record names, or its sub-agent, goes to that reviewer. Spend still left
+fitting several, such as a routed and a shadow reviewer at one effort whose
+sessions no record names, is split evenly among them and reported as `split`.
+Each risk tier's PRs, how many merged,
+how many a reviewer raised, and how many a later PR's `## Caused by` names
+show whether the tiers predict what breaks. Over the merged
+PRs come each online
 reviewer's coverage, final heads reviewed, findings per PR reviewed, and
-answers. Sessions on no PR here, from any
+answers. A hand-back counts only while its draft is open, so a merged PR
+never shows one.
+Sessions on no PR here, from any
 repository or agent token-max reads, total on an `Unattributed` line. `--pr` and
 `--quest` narrow the report to one PR or one quest's branch, and `--json`
 prints it as one object. Without `token-max` on `PATH`, it says how to install

quest/m0/routing-evidence/README.md

@@ -66,9 +66,14 @@ Readout, the agent's choices for the implementer to revisit:
   chooses to.
 - Carry the tallies into [the model check](/quest/m1/model-check.md)'s
   baseline.
-- Publish the arms' cost and quality as
-  [the cost-against-quality charts](/quest/m0/routing-evidence/charts.md) quest describes,
-  and link them with the recommendations.
+- Chart the arms' cost against quality as
+  [the model check](/quest/m1/model-check.md)'s chart procedure describes,
+  one point per arm, and link the charts with the recommendations. Its
+  `quest usage --json` rows give each arm's cost per merged PR and
+  first-round findings, and each shadow pair's cost and findings per PR.
+  Only tallies begun after
+  [moq-dev/quest#721](https://github.com/moq-dev/quest/pull/721) merged
+  record a first round, so count first rounds only over those.
 
 Baseline from the 2026-11-11 audit, at API list price:
 

quest/m0/routing-evidence/handbacks.md

@@ -0,0 +1,35 @@
+# [S] Count hand-backs on merged PRs
+
+## Goal
+
+`quest usage` counts a PR as handed back when any version of its body had a
+`## Waiting` (or `## Decisions needed`) section while it was a draft, merged
+PRs included, so the writer A/B's and the model check's hand-back measure
+sees every PR rather than only drafts still open.
+
+## Plan
+
+Decided 2026-11-14 (user, /quest-run, from
+[moq-dev/quest#721](https://github.com/moq-dev/quest/pull/721)'s
+follow-ups): add this quest rather than drop the measure.
+
+Today `handedBack` in `src/cli/usage.ts` reads only the current body of an
+open draft. `quest ready` and hand edits remove `## Waiting` before merge,
+so every merged PR reads 0.
+
+Recommended (agent's choice): read each PR's body history from GitHub
+(GraphQL `userContentEdits`, whose entries carry each earlier body) where
+`quest usage` already reads the PR, and count a hand-back when an earlier
+body had a hand-back section, using `handedBack` from `src/core/claim.ts`.
+It works on PRs merged before this lands and needs no change to the skills
+or `quest ready`. The alternative, having `quest ready` record hand-backs
+in `## Models`, counts only PRs readied after it lands and trusts every
+writer to keep the record. Batch the history reads with the existing PR
+query so `--since` over a week stays within the GraphQL rate limit.
+
+Count each PR once, however many times it was handed back, and report the
+count in the `accuracy` and per-PR rows as today.
+
+## Related
+
+- [Model check](/quest/m1/model-check.md) - charts hand-backs per PR

quest/m1/model-check.md

@@ -48,6 +48,81 @@ checkpoints and outages to count by hand:
   maintainer with its tally, as a decision in the PR, rather than changing
   it unasked. Replacing a stale model id with the same vendor's newer one,
   below, needs no decision, as this check always did.
+- Chart cost against quality for every writer and reviewer cell, as decided
+  2026-11-11 (user, /quest-plan, moq-dev/quest#642's follow-ups), so a
+  routing change is read off the frontier rather than tables. The agent
+  running the check makes the charts itself, from the saved
+  `quest usage --since <date> --json` (with `--ssh` for each host that runs
+  agents) and the benchmark table below. The charts are an artifact of the
+  run, and no chart code goes into `quest`. It draws them as Mermaid
+  charts in its PR body, which GitHub renders and any harness can write
+  (user, 2026-11-14, /quest-run on
+  [moq-dev/quest#721](https://github.com/moq-dev/quest/pull/721); rejected:
+  images in a secret gist, a claude.ai page made by hand). Mermaid's
+  quadrant charts scale both axes to 0-1, so each title gives the scales
+  and each point's label carries its PR count. The body also puts every
+  plotted number in tables, so a reviewer can check the charts. Each
+  chart names its date range, the command, and each point's PR count.
+  Points too thin for the run's variance estimate stay on the chart,
+  marked thin in their label, never dropped. Cells whose
+  `Implementer:` effort is `unknown` stay separate points. The charts:
+  - Writers, one point per `accuracy` row (size, model, effort): cost per
+    merged PR, `spent.cost / spent.pulls`, against each quality measure
+    on its own chart. The measures are first-round findings per PR
+    (`firstRounds.findings` over `firstRounds.pulls`), review rounds per
+    reviewed PR (`tally.rounds / reviewed`), hand-backs per PR
+    (`handedBack / pulls`, which counts only drafts still open until
+    [hand-backs on merged PRs](/quest/m0/routing-evidence/handbacks.md)
+    lands), and later
+    `## Caused by` escapes per merged PR (`fixes / merged`).
+  - Local reviewers, one point per `reviewers` row (PR size, harness,
+    model, effort, shadow): cost per PR reviewed, `spent.cost /
+    spent.pulls`, against findings per reviewed PR (`tally.findings` over
+    `reviewed`) and the share fixed, `tally.fixed` over the findings
+    settled one way or another (fixed, refuted, or settled).
+  - Online reviewers, which run on flat plans and so have no cost per
+    review, on a chart of their own: one point per `online` row, coverage
+    (`reviewed / merged`) against findings per reviewed PR, labelled with
+    its acceptance (`accepted / findings`).
+  - Published benchmarks, one chart per suite in the table below: score
+    against that suite's cost per task, for the same models and efforts,
+    with the frontier drawn. Benchmark and PR metrics differ, so compare
+    where each cell sits on each chart's frontier, not its values; flag a
+    cell on the frontier in one and not the other.
+- Published benchmarks, refreshed each run: rows with a stated effort, each
+  with its source and the date it was read. Cost is per task in the suite's
+  own terms; `~` marks a value derived from the source's prose.
+
+  | Suite | Model | Effort | Score | $ per task | Read |
+  |---|---|---|---|---|---|
+  | [mypriors](https://mypriors.com/adamholter) 10-benchmark index, subscription cost | Haiku 5.5 | high | 30 | 0.0034 | 2026-11-09 |
+  | mypriors | Haiku 5.5 | xhigh | 34.6 | 0.0053 | 2026-11-09 |
+  | mypriors | Sonnet 5.5 | medium | 30.7 | 0.012 | 2026-11-09 |
+  | mypriors | Opus 5.5 | low | 34.7 | 0.015 | 2026-11-09 |
+  | mypriors | Opus 5.5 | medium | 48 | 0.037 | 2026-11-09 |
+  | mypriors | Opus 5.5 | high | 51.7 | 0.052 | 2026-11-09 |
+  | mypriors | Opus 5.5 | xhigh | 57.3 | 0.092 | 2026-11-09 |
+  | mypriors | gpt-6.1-sol | low | 34.8 | 0.0094 | 2026-11-09 |
+  | mypriors | gpt-6.1-sol | medium | 46.3 | 0.021 | 2026-11-09 |
+  | mypriors | gpt-6.1-sol | high | 50.3 | 0.034 | 2026-11-09 |
+  | mypriors | gpt-6.1-sol | xhigh | 52.5 | 0.045 | 2026-11-09 |
+  | mypriors | Fable 5.2 | medium | 56 | 3.8 | 2026-11-09 |
+  | [Anthropic's cost guide](https://platform.claude.com/docs/en/about-claude/models/optimizing-for-cost-and-intelligence), SWE-bench Pro 478-problem subset, % solved | Sonnet 5.5 | medium | ~80.6 | ~0.057 | 2026-11-13 |
+  | Anthropic, SWE-bench Pro | Sonnet 5.5 | high | 87.6 | ~0.096 | 2026-11-13 |
+  | Anthropic, SWE-bench Pro | Opus 5.5 | low | 87.4 | 0.11 | 2026-11-13 |
+  | Anthropic, SWE-bench Pro | Opus 5.5 | medium | 92.8 | 0.21 | 2026-11-13 |
+  | Anthropic, SWE-bench Pro | Opus 5.5 | high | 95.3 | 0.29 | 2026-11-13 |
+  | Anthropic, SWE-bench Pro | Opus 5.5 | xhigh | 96.7 | 0.74 | 2026-11-13 |
+  | [Artificial Analysis](https://artificialanalysis.ai/models/comparisons/claude-sonnet-5-5-high-vs-claude-opus-5-5-low), Terminal-Bench 4.0, % | Haiku 5.5 | xhigh | 29.3 | 0.25 | 2026-11-11 |
+  | Artificial Analysis, Terminal-Bench 4.0 | Sonnet 5.5 | medium | 29.8 | 0.78 | 2026-11-11 |
+  | Artificial Analysis, Terminal-Bench 4.0 | Sonnet 5.5 | high | 43.9 | 1.64 | 2026-11-13 |
+  | Artificial Analysis, Terminal-Bench 4.0 | Opus 5.5 | low | 31.3 | 0.98 | 2026-11-13 |
+
+  Anthropic gives Sonnet high's cost per solved task ($0.11), Sonnet
+  medium as about 7 points lower at about 59% of high's cost per task, and
+  calls its figures directional. Its guide also has rows on internal
+  suites without a cost per task, which stay out. The `max` rows stay out
+  too, since `quest agent` and `quest review` refuse `max`.
 
 - Read [routing evidence](/quest/m0/routing-evidence/README.md)'s readout
   and baseline first: its shadow reviews and writer A/B measure the cells
@@ -80,25 +155,16 @@ checkpoints and outages to count by hand:
     effort (6 of the 20 lines said `unknown`).
   - Changed after that run rather than on tallies (maintainer, 2026-11-09):
     XS writes on Haiku xhigh instead of high. Tally that cell fresh. The
-    source was [mypriors.com/adamholter](https://mypriors.com/adamholter)'s
-    score against subscription cost per task, a 10-benchmark index read
-    off its chart (score, then $ per task), which this check compares each
-    cell with again:
-    - Haiku 5.5: high 30, $0.0034; xhigh 34.6, $0.0053.
-    - Sonnet 5.5: medium 30.7, $0.012; below the frontier at every effort,
-      so rejected for XS then, though Anthropic's
+    source was the mypriors rows of the benchmark table above:
+    - Sonnet 5.5 sat below the frontier at every effort, so it was rejected
+      for XS then, though Anthropic's
       [model guide](https://claude.dev/blog/which-claude-model-should-you-build-with/)
       favors it for scoped coding.
-    - Opus 5.5: low 34.7, $0.015; medium 48, $0.037; high 51.7, $0.052;
-      xhigh 57.3, $0.092; max 58.9, $0.15.
-    - gpt-6.1-sol: low 34.8, $0.0094; medium 46.3, $0.021; high 50.3,
-      $0.034; xhigh 52.5, $0.045.
-    - Fable 5.2: medium 56, $3.8; max 64.6, $26.5. Rejected for XL, which
-      stays on Opus xhigh, on the frontier; zero-data-retention
-      organizations may also be refused Fable.
-    - S looks lopsided: Opus low scores well below its Sol medium fallback
-      (34.7 against 46.3) for $0.006 less a task. Check its tallies before
-      moving it.
+    - Fable 5.2 was rejected for XL, which stays on Opus xhigh, on the
+      frontier; zero-data-retention organizations may also be refused
+      Fable.
+    - S looked lopsided: Opus low scored well below its Sol medium fallback
+      (34.7 against 46.3) for $0.006 less a task.
   - Changed again (maintainer, 2026-11-11): XS writes on Sonnet high
     instead of Haiku xhigh, since two Haiku XS agents in a row on
     [moq-dev/quest#643](https://github.com/moq-dev/quest/pull/643) died