mirror of
https://github.com/obra/superpowers.git
synced 2026-07-23 10:44:01 +08:00
3921dc9998
Experiment branch — NOT for merge without eval evidence (writing-skills
discipline; RED/GREEN campaign to follow if the manual run validates).
Two targeted changes against the measured Codex 5.6-era spinout
(36% of SDD runs >8h, PRI-2672):
1. codex-tools.md — SDD dispatch rules: fork_turns "none" on every
spawn (the default "all" forks the whole transcript and refuses
model/effort overrides — S2's terminal 13-agent final wave was all
full-history forks at sol/xhigh); capability-check for 0.145+ spawn
params with an explicit terra/high role table (re-review terra/medium),
no sol seats, no effort escalation between fix rounds; honest
inheritance warning for <=0.144 where routing is impossible (T0-probed
on 0.144.4: schema is {task_name, message, fork_turns} only, role
TOMLs inert).
Role-table bet: census shows task seats already ran terra/high in the
spinout sessions — the surgical bet is fork_turns:none + no-sol-seats
+ wave termination, not task-seat downgrade. T1 cross-review matrix
will refine reviewer tiers (may support terra/medium or lower).
2. SKILL.md final review — wave closure is policy, not a verdict:
strong reviewers find real defects indefinitely, so "review until
clean" never terminates; new breakage in the final fix diff joins
residual adjudication instead of opening wave two; one-off human
review procedures (competing reviewers, scoring) never become
standing. Two matching rationalization-table rows.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
81 lines
3.8 KiB
Markdown
81 lines
3.8 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:
|
|
|
|
| Role | model | reasoning_effort |
|
|
|------|-------|------------------|
|
|
| Implementer (all rounds) | `gpt-5.6-terra` | `high` |
|
|
| Task reviewer | `gpt-5.6-terra` | `high` |
|
|
| Scoped re-review | `gpt-5.6-terra` | `medium` |
|
|
| Final whole-branch review | `gpt-5.6-terra` | `high` |
|
|
|
|
On Codex this table IS the Model Selection mapping — 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. Frontier-tier subagents at inherited effort are the
|
|
measured top driver of Codex SDD runs spinning out to 8+ hours
|
|
(PRI-2672): review seats that inherited sol found real-but-endless
|
|
defects every round, and fix diffs ballooned instead of converging.
|
|
|
|
**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.
|