mirror of
https://github.com/obra/superpowers.git
synced 2026-07-23 10:44:01 +08:00
c97988d8d1
Codex SDD runs at frontier tiers spin out: subagents inherit the session's model and effort, review seats find real-but-endless defects every round, and the final whole-branch review has no reachable termination state. Field forensics across multiple real sessions traced the mechanism: spawn_agent's fork_turns defaults to "all" (full-context forks that also refuse model overrides), omitted model/effort params inherit the frontier parent, and dispatch rules loaded at session start do not survive context compaction — a compacted controller reverts to inheritance exactly when the session is longest and most expensive. Three coordinated changes: - codex-tools.md: every SDD spawn sets fork_turns "none" plus explicit model/reasoning_effort when the schema supports them (Codex 0.145+), with an inheritance warning for older builds; reviewer tier never exceeds implementer tier and fix rounds never get effort bumps. - task-brief/review-package print a dispatch tuple with their output (review-package grows --role for re-review/final-review), putting the routing values in front of the controller at the moment of dispatch — reprinted every round, so they survive compaction by construction. - SKILL.md: the final-review wave closing is policy, not a verdict — one fix dispatch, one scoped re-review, residuals adjudicated to the ledger; one-off review procedures never become standing; Model Selection defers to platform role tables where one exists. Validated live: a previously spun-out 14-task run (8h, 6 tasks, died mid-loop) re-executed to completion under these changes — bounded fix loops, single-cycle final review, ~20 consecutive correctly-routed dispatches across two compactions. Reviewer-tier matrix (125 cells) showed no calibration cost: clean-diff approvals and planted-defect recall are identical across tiers. Eval scenarios to follow before any upstream submission. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
83 lines
4.2 KiB
Markdown
83 lines
4.2 KiB
Markdown
## Subagent dispatch requires multi-agent support
|
|
|
|
Add to your Codex config (`~/.codex/config.toml`):
|
|
|
|
```toml
|
|
[features]
|
|
multi_agent = true
|
|
```
|
|
|
|
This enables `spawn_agent`, `wait_agent`, and `close_agent` for skills like `dispatching-parallel-agents` and `subagent-driven-development`. When using subagent-driven-development, close reviewer subagents when their review returns. Keep each implementer subagent open until its task's review passes — the fix loop resumes the implementer — then close it. If your harness cannot send another message to a spawned agent, dispatch each fix round as a fresh implementer carrying the brief, the report file, and the findings.
|
|
|
|
## SDD dispatch: pin fork_turns and the model on every spawn
|
|
|
|
Every `spawn_agent` call in subagent-driven-development sets
|
|
`fork_turns: "none"`. The parameter defaults to `"all"`, which forks your
|
|
entire session transcript into the child — the opposite of the fresh,
|
|
constructed context SDD requires — and a full-history fork also refuses
|
|
model and effort overrides. Never omit the parameter, and never pass
|
|
`"all"` or a turn count for an SDD dispatch.
|
|
|
|
Before Task 1, check your `spawn_agent` tool schema for `model` and
|
|
`reasoning_effort` parameters (present on Codex 0.145+).
|
|
|
|
**If the parameters exist**, set both explicitly on every dispatch. The
|
|
task-brief and review-package scripts print the exact values for each
|
|
dispatch as a `dispatch (codex spawn_agent):` line with their output —
|
|
copy that line's values onto the spawn_agent call verbatim, every time,
|
|
even late in a long session. The mapping they print: every SDD seat runs
|
|
`gpt-5.6-terra` — implementers and reviewers at `reasoning_effort: high`,
|
|
scoped re-reviews at `medium`. On Codex this mapping IS the Model
|
|
Selection section — including the final review, which stays on
|
|
terra/high rather than "most capable available." Never give a subagent
|
|
your session's model when you run a frontier config (sol at xhigh or
|
|
max): reviewer tier never exceeds implementer tier, and a fix round
|
|
never gets an effort bump. Rounds 4-5's "more capable model" means a
|
|
fresh implementer at the same tier; a task that genuinely needs more
|
|
than terra/high is a BLOCKED escalation to your human partner, not a
|
|
quiet tier climb. Inherited frontier-tier subagents are a measured cause
|
|
of SDD runs spinning out for hours: review seats that inherit a frontier
|
|
model at maximum effort find real-but-endless defects every round, and
|
|
fix diffs balloon instead of converging.
|
|
|
|
The model names here track Codex's `spawn_agent` allowlist (currently
|
|
`gpt-5.6-sol` and `gpt-5.6-terra`). When the allowlist changes, update
|
|
this file and the hint lines in task-brief and review-package together.
|
|
|
|
**If the parameters do not exist** (Codex 0.144 and earlier), every child
|
|
inherits your session's model and effort and no override is possible —
|
|
role files in `~/.codex/agents/` do not attach to spawns either. Say so
|
|
to your human partner before starting a plan of more than a few tasks,
|
|
and offer the choice: proceed with inheritance, or restart the session
|
|
at a lower effort so the whole run — controller and children — pays the
|
|
lower rate.
|
|
|
|
## Environment Detection
|
|
|
|
Skills that create worktrees or finish branches should detect their
|
|
environment with read-only git commands before proceeding:
|
|
|
|
```bash
|
|
GIT_DIR=$(cd "$(git rev-parse --git-dir)" 2>/dev/null && pwd -P)
|
|
GIT_COMMON=$(cd "$(git rev-parse --git-common-dir)" 2>/dev/null && pwd -P)
|
|
BRANCH=$(git branch --show-current)
|
|
```
|
|
|
|
- `GIT_DIR != GIT_COMMON` → already in a linked worktree (skip creation)
|
|
- `BRANCH` empty → detached HEAD (cannot branch/push/PR from sandbox)
|
|
|
|
See `using-git-worktrees` Step 0 and `finishing-a-development-branch`
|
|
Step 1 for how each skill uses these signals.
|
|
|
|
## Codex App Finishing
|
|
|
|
When the sandbox blocks branch/push operations (detached HEAD in an
|
|
externally managed worktree), the agent commits all work and informs
|
|
the user to use the App's native controls:
|
|
|
|
- **"Create branch"** — names the branch, then commit/push/PR via App UI
|
|
- **"Hand off to local"** — transfers work to the user's local checkout
|
|
|
|
The agent can still run tests, stage files, and output suggested branch
|
|
names, commit messages, and PR descriptions for the user to copy.
|