feat(ce-work-beta): adaptive effort selection for Codex delegation batches (#759)

Author: davidalee

plugins/compound-engineering/skills/ce-work-beta/SKILL.md

@@ -65,7 +65,8 @@ Store the resolved state for downstream consumption:
 - `sandbox_mode` -- `yolo` or `full-auto` (from config or default `yolo`)
 - `consent_granted` -- boolean (from config `work_delegate_consent`)
 - `delegate_model` -- string from config, or unset (defer to Codex config)
-- `delegate_effort` -- string from config, or unset (defer to Codex config)
+- `delegate_effort` -- string from config, or unset (defer to Codex config). Floor for per-batch effort selection; not passed directly to `codex exec`.
+- `effective_effort` -- per-batch derived value (`default | medium | high | xhigh`), computed before each batch from `delegate_effort` and the picked level per `references/codex-delegation-workflow.md` ("Per-Batch Effort"). Feeds the `codex exec` invocation in place of `delegate_effort`.
 
 ---
 

plugins/compound-engineering/skills/ce-work-beta/references/codex-delegation-workflow.md

@@ -88,6 +88,45 @@ On decline:
 
 Delegate all units in one batch. If the plan exceeds 5 units, split into batches at the plan's own phase boundaries, or in groups of roughly 5 -- never splitting units that share files. Skip delegation entirely if every unit is trivial.
 
+## Per-Batch Effort
+
+Each batch picks an effort level proportional to its complexity, then resolves against the config floor before invocation.
+
+**Effort levels — guidelines, not predicates**
+
+Pick the level that best fits the batch. These are signals to weigh, not boxes to tick — use judgment.
+
+- **default (no flag)** — trivial work with no behavioral change: a one-line config tweak, a rename, a typo or comment-only fix, a pure documentation update. Defers to the user's `~/.codex/config.toml` default (which is `medium` on a stock Codex install).
+- **`medium`** — small, well-scoped behavioral changes that stay clear of high-risk areas. A handful of files, a single concern, no novel architecture.
+- **`high`** — work that touches a high-risk area (auth/session logic, payments, database migrations, external API contracts, error handling with retries/fallbacks), or work spanning enough surface area that one mistake could cascade.
+- **`xhigh`** — architectural work: cross-cutting refactors, multiple high-risk areas in the same batch, changes that propagate broadly, or anywhere a wrong call meaningfully degrades the project.
+
+When in doubt, lean up one level — under-resourcing risky work costs more than over-resourcing routine work. Briefly note the picked level and the signal that drove it (e.g., "`high` — touches db/migrations") so the choice is auditable.
+
+A few edge cases worth handling explicitly:
+- **Test-only batches:** classify by what the tests *exercise*, not by file paths. Tests for auth flows, payment logic, or migrations get the same level the equivalent implementation work would get.
+- **Mixed-complexity batches:** the batch picks one level. If a single batch combines a typo unit and a payments rewrite, pick the higher level. If the spread feels wasteful, prefer splitting at the batching step (see Batching above) over averaging it out.
+- **Deletion-only batches:** classify by the risk of what is being removed, not by counts of remaining content. Removing an auth module is `high` even if the batch produces zero `Modify` content.
+- **Documentation- or comment-only batches:** `default`.
+
+**Floor and resolution — hard rules**
+
+Effort levels are ordered: `minimal < low < medium < high < xhigh`.
+
+Compute `effective_effort`:
+
+- If `delegate_effort` is unset: `effective_effort = picked_level`.
+- If `delegate_effort` is set: substitute `default` → `medium` in `picked_level`, then `effective_effort = max(picked_level, delegate_effort)`.
+
+Emit based on `effective_effort`:
+
+- `medium`, `high`, or `xhigh` → emit `-c 'model_reasoning_effort="<value>"'`.
+- `default` → omit the flag (defer to `~/.codex/config.toml`). Reachable only when `delegate_effort` is unset and the pick is `default`.
+
+Never pass the literal string `"default"` to `codex exec`.
+
+Store `effective_effort` as a per-batch derived state value (alongside the session-level `delegate_effort`) and use it in place of `delegate_effort` throughout the Execution Loop.
+
 ## Prompt Template
 
 At the start of delegated execution, create a per-run OS-temp scratch directory via `mktemp -d` and capture its **absolute path** for all downstream use. All scratch files for this invocation live under that directory. Do not use `.context/` — these scratch files are per-run throwaway that get cleaned up when delegated execution ends (see Cleanup below), matching the repo Scratch Space convention for one-shot artifacts. Do not pass unresolved shell-variable strings to non-shell tools (Write, Read); use the absolute path returned by `mktemp -d`.
@@ -239,7 +278,7 @@ codex exec \
 **Conditional flags** — only include each line when the corresponding skill-state value is set:
 
 - If `delegate_model` is set, insert `  -m "<delegate_model>" \` as a line before `$SANDBOX_FLAG`.
-- If `delegate_effort` is set, insert `  -c 'model_reasoning_effort="<delegate_effort>"' \` as a line before `$SANDBOX_FLAG`.
+- If `effective_effort` is `medium`, `high`, or `xhigh` (resolved via Per-Batch Effort above), insert `  -c 'model_reasoning_effort="<effective_effort>"' \` as a line before `$SANDBOX_FLAG`. When `effective_effort` is `default` (only possible when `delegate_effort` is unset and the pick is `default`), omit the line — never pass the literal string `"default"`.
 
 When either value is unset, omit its line entirely — Codex resolves the default from the user's `~/.codex/config.toml` (and ultimately the CLI's own built-in default). Do not substitute a placeholder string for unset values.