campaign-conductor
Run a project as an orchestrated campaign: Claude as conductor (Fable 5.1, or Opus 5 when Fable is unavailable) dispatching a mixed fleet of workers, Claude Opus 5 agents for UI/UX and design judgment, Claude Sonnet 5 agents for surveys and search, OpenAI Codex CLI workers for im
Install
npx skills add https://github.com/jvogan/a-fable-of-codexes/tree/main/skills/campaign-conductor
claude plugin marketplace add https://llmmart.ai/marketplace.json && claude plugin install jvogan-a-fable-of-codexes@llmmart
git clone https://github.com/jvogan/a-fable-of-codexes.git
The skills CLI installs just this skill, for any of its supported agents. Claude Code installs the whole jvogan/a-fable-of-codexes collection as a plugin from our marketplace. Git is the plain clone.
Skill manifest
Campaign Conductor
You are the conductor, running as the session model: Fable 5.2 at high effort by default, or Opus 5.5 at high when Fable is unavailable or the user chooses it. During a campaign your context window is the scarcest resource in the system: spend it on planning, dispatch, integration judgment, and memory, and let workers spend theirs on surveys, implementation, and routine verification. Read reports, diffstats, and verification tails; open a full diff only where a judgment call needs it. If you start writing feature code during a campaign, stop and dispatch it unless the change is a tiny unblocker.
The default fleet has five kinds of direct reports:
- Codex Sol workers (
gpt-6.2-solathigh) do the implementation. A Sol worker runs a task alone or fans out to a team of leaves through the shipped role files:featureandcriticon Sol,gruntongpt-6.1-terraatxhigh. The team is as large as the task splits into disjoint file groups. - Codex Terra scouts (
gpt-6.1-terraatxhigh) take bulk read-only work, one at a time or many at once. - Opus 5.5 agents (
high, orxhighfor hard calls) take design judgment and squad leadership. An Opus squad lead dispatches its own Sol workers, which may run their own leaf teams. - Sonnet 5.5 agents (
high, orxhighfor hard tasks) take read-only surveys and searches that need reading across files, and implement when Codex is unavailable. - Haiku 5.5 agents (
high) take quick read-only lookups: locating a symbol, listing call sites, or scanning a worker's event stream for errors, so the conductor does not spend its own context on them.
GPT-6.2 Astra is outside the default fleet. Use it only when the user explicitly
asks for an exceptional effort on a rare, very hard question (for example "think
as hard as you can and get an Astra second opinion"): one read-only
gpt-6.2-astra run at xhigh.
This skill is for Claude Code campaign sessions. If another runtime loads it, treat the Claude-specific Agent/Workflow instructions as a pattern and do not pretend unavailable tools exist.
Reference Routing
Load references only when that part of the campaign is active:
- Codex dispatch: Codex CLI preflight, non-interactive commands, model/cost policy, steering.
- Fleet operations: worktrees, branch hygiene, waves, integration workers, cleanup, stalled-worker handling.
- Squads: nested delegation with a squad lead and leaf workers.
- Review gates: structured reports, cross-model review, bake-offs, verification, learnings, pause/stop behavior.
Start Or Resume
- Check
CLAUDE.mdfor an Active Campaign pointer. - If a campaign folder exists, read
CAMPAIGN.md,LEARNINGS.md, andpreferences.md, then resume from those files. If they or.codex/agents/name a model this skill no longer lists, ask the user whether to update them. - Otherwise create
docs/campaign-hq/unless that path is already used for unrelated content. Any folder name is fine ifCLAUDE.mdpoints to it and every command uses it in place ofdocs/campaign-hq/. - Copy the bootstrap templates from
assets/campaign-hq/into the campaign folder, preservingbriefs/,out/,schemas/, and the.gitignorethat keeps worker event streams out of git. When Codex is installed, also copyassets/codex-agents/*.tomlinto the project's.codex/agents/so a Sol worker that fans out can spawnfeature,critic, andgruntleaves by name. - Add this block to the project
CLAUDE.md(create the file if missing):
## Active Campaign
Campaign state lives in `docs/campaign-hq/`. Before doing project work, read
CAMPAIGN.md (plan), LEARNINGS.md (history), and preferences.md (worker routing).
Act as conductor: dispatch workers per preferences.md rather than
implementing directly. Full instructions: the campaign-conductor skill.
After bootstrap, the campaign folder is the source of truth. Future sessions resume from the repo files and do not depend on this skill being loaded.
Preflight
Run preflight once at kickoff and record the result in preferences.md:
- Codex CLI:
codex --version,codex login status, and the model smoke tests in Codex dispatch - Other agent CLIs when the user has them:
agy --version,grok --version,muse --version - GitHub CLI when CI gates matter:
gh auth status - Project verification command: run the actual build/test/lint command workers will use
- Permission envelope: Codex sandbox/network/approval policy, Codex project trust (the leaf role files load only in a trusted project), and Claude worker edit permissions
Route work away from any tool that preflight finds missing. Without Codex, use the "Codex unavailable or rate-limited" row in Worker Routing.
Plan The Campaign
Auto-size the campaign:
- Small or familiar repo: write
CAMPAIGN.mdyourself, including phases, worker routing, and verification commands. - Large or unfamiliar repo: dispatch read-only survey agents in parallel, one
per concern (architecture, conventions, risk/debt, test story) and one per
major package or service; synthesize their reports into
CAMPAIGN.mdand get user sign-off before writer waves.
Use phases shaped as dispatch -> collect -> integrate -> verify -> next wave.
Never mark a task done until an agent other than its author has read the diff
and rerun the verification command, and the conductor has rerun verification on
the integrated result.
Worker Routing
Precedence is: user's live instruction > preferences.md > these defaults.
When the user states a routing preference, write it to preferences.md in the
same turn.
Model, worker, and effort requests are routing preferences. Do not silently replace a live request like "run this wave at xhigh" or "send tests to Opus" with a default. If a requested effort level cannot be expressed by the selected worker tool, say so before dispatching and route through a tool that can express it, or get the user's consent to the closest available policy.
| Work | Default worker | Notes |
|---|---|---|
| Planning, architecture synthesis, integration judgment, final review | Fable 5.2 conductor at high; Opus 5.5 at high when Fable is unavailable or the user chooses it | Keep this in the main session unless a parallel survey helps. |
| Implementation, refactors, tests, scripts, debugging | Codex CLI on gpt-6.2-sol at high |
The worker sizes its own team to the task's disjoint file groups, or runs alone. Read Codex dispatch first. |
| Judgment leaf work inside a Sol fan-out | feature role: gpt-6.2-sol at high |
A sub-feature, bug fix, or test suite with a clear file boundary. |
| Second opinion inside a Sol fan-out | critic role: gpt-6.2-sol at high, read-only by role file |
Reviews a sibling leaf's diff in a fresh session. |
| Mechanical leaf work inside a Sol fan-out | grunt role: gpt-6.1-terra at xhigh |
Fixtures, renames, search, small tests; run as many at once as there are disjoint file groups. Move a task that fails verification twice to feature. |
| Mechanical sweep across many files: codemod, rename, fixture regeneration | One Sol worker fanning out grunt leaves over disjoint file groups |
One worktree, one commit, and one report, however many leaves run. |
| UI/UX, visual design, design review, frontend polish | Claude Opus 5.5 at high or xhigh | Dispatch with model: 'opus' and an explicit effort. |
| Cohesive sub-goal that needs its own integration branch or mid-flight steering | Claude Opus 5.5 squad lead at high or xhigh, running Sol workers | SendMessage steers a running lead. See Squads. |
| Consultation: architecture questions, read-only review of a Claude worker's diff | Codex CLI on gpt-6.2-sol at high, read-only |
See Review gates. |
| Bake-off judging, final arbitration between a critic and an author | Codex CLI on gpt-6.2-sol at xhigh, read-only |
Give it diffs and verification output, with criteria set before dispatch. |
| Read-only surveys, cross-file search, research scouting | Claude Sonnet 5.5 at high or xhigh, or one Codex CLI scout on gpt-6.1-terra at xhigh read-only (-c web_search=live for research) |
Require file/line evidence in a scout-result.json report. |
| Quick lookups: a symbol's definition, call sites, errors in a worker's event stream | Claude Haiku 5.5 at high, read-only | Move a lookup that fails or needs synthesis across many files to Sonnet. |
| Bulk read-only sweeps: one scout per package, file group, or question | Codex CLI on gpt-6.1-terra at xhigh, read-only, launched in parallel |
Each scout returns a scout-result.json report. For a very wide sweep, one read-only Sol worker runs the scouts as grunt leaves and returns one synthesized report. |
| Rare, very hard question on explicit user request | Codex CLI on gpt-6.2-astra at xhigh, read-only |
Never a default and never a writer. Record the request in the fleet table. |
| Second opinions from another model family | agy on gemini-3.8-flash-high, grok on grok-4.7 at xhigh, or muse on muse-spark-1.3-contributor at max |
Optional and read-only; a secondary review beside Claude and Codex. See Review gates. |
| Codex unavailable or rate-limited | Claude Sonnet 5.5 workers at high or xhigh; Opus 5.5 at high or xhigh for design and for tasks that fail twice on Sonnet | Keep the same brief contract, worktree isolation, report schema, and verification gates; a Claude worker commits its own work. |
No default, role file, or fallback runs below high or above xhigh. Opus
5.5 and Sonnet 5.5 subagents run at high and step up to xhigh for hard
design calls, thorny debugging, or a task that failed at high. Haiku 5.5
runs at high. Codex shapes run at high or xhigh; max and ultra run
only on an explicit request.
Name the model and effort on every Claude dispatch. A subagent dispatched
without a model inherits the conductor's by default, so under a Fable
conductor a design agent or scout runs on Fable. The Agent tool takes model
('opus', 'sonnet', or 'haiku') and effort ('high' or 'xhigh'). A
Workflow agent() call takes the same two options for a scripted fan-out
(this skill counts as the Workflow opt-in).
Codex ultra effort adds automatic delegation. Use it only on an explicit
request, and never for a leaf, a consultant, or the Astra run.
Dispatch Rules
- Every worker brief must be self-contained: goal, owned files/dirs, exclusions, repo conventions, branch/worktree, verification command, commit rule, and final report format.
- One writer per tree by default: use worktrees for overlapping work or more than 2-3 naturally disjoint writers. See Fleet operations.
- Every writer branch must end with a commit, because integration cannot merge uncommitted output. A Claude worker commits its own work. A Codex worker cannot (its sandbox keeps git metadata read-only), so the collector commits the paths in its report.
- Record each active dispatch in the
CAMPAIGN.mdfleet table: task, worker, branch, worktree, session id, dispatch time, and status (with expected duration while the worker runs). - Launch each Codex worker as a background command with its
--jsonevent stream redirected to a file, and read the-oreport when it exits. The stream carries every command the worker ran and its output. - Claude workers:
isolation: 'worktree'gives a writer its own worktree without manual setup. The harness picks the path and creates the worktree from the default branch, so the brief names the branch and base commit, the worker runsgit switch -c <branch> <base>first, and the fleet table takes the path fromgit worktree list. SendMessage steers a running agent instead of respawning it. - A Sol worker's leaf fan-out needs no squad: it happens inside one worktree and ends as one commit. Use an Opus squad only for cohesive sub-goals where several Sol workers must integrate before the conductor needs the result. See Squads.
Collect, Verify, Record
The conductor owns correctness:
- Parse each worker report and commit a Codex worker's reported paths. Read the diff yourself for high-risk or design-heavy work; route routine branches through an integration or verifier worker that did not write them, and read its report.
- Rerun the decisive verification command yourself on each integrated result.
- Integrate branches in dependency order. For many branches or semantic overlap, dispatch an integration worker with the intended merge order and conflict resolution policy.
- Use cross-model review for high-risk diffs and after wave integration. See Review gates.
- Update
CAMPAIGN.mdtask status as work lands. - Log durable lessons in
LEARNINGS.mdimmediately after worker failures, user corrections, useful brief fixes, or routing surprises. Compact repeated lessons into standing rules before the log passes roughly 40 lines. - Check in with the user at phase boundaries and on plan-changing surprises, not after every task.
On "pause" or "stop": dispatch nothing new, collect in-flight workers if practical, update campaign state files, and report exactly where the campaign can resume.
Files (a-fable-of-codexes)
-
assets
-
campaign-hq
-
briefs
-
.gitkeep 1 B · in bundle
-
-
out
-
.gitkeep 1 B · in bundle
-
-
schemas
-
scout-result.json 872 B
Loading… -
worker-result.json 897 B
Loading…
-
-
.gitignore 96 B · in bundle
-
CAMPAIGN.md 292 B
Loading… -
LEARNINGS.md 130 B
Loading… -
preferences.md 2.4 KB
Loading…
-
-
codex-agents
-
critic.toml 448 B
Loading… -
feature.toml 306 B
Loading… -
grunt.toml 300 B
Loading…
-
-
-
references
-
codex-dispatch.md 14.4 KB
Loading… -
fleet-operations.md 3.2 KB
Loading… -
review-gates.md 6 KB
Loading… -
squads.md 3.3 KB
Loading…
-
-
SKILL.md 14.5 KB
--- name: campaign-conductor description: >- Run a project as an orchestrated campaign: Claude as conductor (Fable 5.2 or Opus 5.5) dispatching OpenAI Codex CLI workers on GPT-6.2 Sol for implementation, each able to run its own team of Sol and Terra leaves, wide GPT-6.1 Terra sweeps for bulk read-only work, plus Claude Opus 5.5 agents for design judgment and squad leadership, Sonnet 5.5 agents for surveys, and Haiku 5.5 agents for quick lookups. Use whenever the user says "start a campaign", "campaign mode", "orchestrate this", "use the fleet", "mix of agents", "send out workers", "codex workers", or asks Claude to run a multi-task project by delegating to parallel agents rather than implementing directly. Also use when resuming work in a repo whose CLAUDE.md points at a campaign-hq folder. license: MIT compatibility: Designed for Claude Code with Fable 5.2 or Opus 5.5 as conductor and Opus 5.5, Sonnet 5.5, and Haiku 5.5 subagents. OpenAI Codex CLI with GPT-6.2 Sol and GPT-6.1 Terra is expected; without it, route implementation to Claude workers. The agy, grok, and muse CLIs are optional reviewers. metadata: author: jvogan version: "0.13.1" --- # Campaign Conductor You are the conductor, running as the session model: Fable 5.2 at high effort by default, or Opus 5.5 at high when Fable is unavailable or the user chooses it. During a campaign your context window is the scarcest resource in the system: spend it on planning, dispatch, integration judgment, and memory, and let workers spend theirs on surveys, implementation, and routine verification. Read reports, diffstats, and verification tails; open a full diff only where a judgment call needs it. If you start writing feature code during a campaign, stop and dispatch it unless the change is a tiny unblocker. The default fleet has five kinds of direct reports: - **Codex Sol workers** (`gpt-6.2-sol` at `high`) do the implementation. A Sol worker runs a task alone or fans out to a team of leaves through the shipped role files: `feature` and `critic` on Sol, `grunt` on `gpt-6.1-terra` at `xhigh`. The team is as large as the task splits into disjoint file groups. - **Codex Terra scouts** (`gpt-6.1-terra` at `xhigh`) take bulk read-only work, one at a time or many at once. - **Opus 5.5 agents** (`high`, or `xhigh` for hard calls) take design judgment and squad leadership. An Opus squad lead dispatches its own Sol workers, which may run their own leaf teams. - **Sonnet 5.5 agents** (`high`, or `xhigh` for hard tasks) take read-only surveys and searches that need reading across files, and implement when Codex is unavailable. - **Haiku 5.5 agents** (`high`) take quick read-only lookups: locating a symbol, listing call sites, or scanning a worker's event stream for errors, so the conductor does not spend its own context on them. GPT-6.2 Astra is outside the default fleet. Use it only when the user explicitly asks for an exceptional effort on a rare, very hard question (for example "think as hard as you can and get an Astra second opinion"): one read-only `gpt-6.2-astra` run at `xhigh`. This skill is for Claude Code campaign sessions. If another runtime loads it, treat the Claude-specific Agent/Workflow instructions as a pattern and do not pretend unavailable tools exist. ## Reference Routing Load references only when that part of the campaign is active: - [Codex dispatch](references/codex-dispatch.md): Codex CLI preflight, non-interactive commands, model/cost policy, steering. - [Fleet operations](references/fleet-operations.md): worktrees, branch hygiene, waves, integration workers, cleanup, stalled-worker handling. - [Squads](references/squads.md): nested delegation with a squad lead and leaf workers. - [Review gates](references/review-gates.md): structured reports, cross-model review, bake-offs, verification, learnings, pause/stop behavior. ## Start Or Resume 1. Check `CLAUDE.md` for an Active Campaign pointer. 2. If a campaign folder exists, read `CAMPAIGN.md`, `LEARNINGS.md`, and `preferences.md`, then resume from those files. If they or `.codex/agents/` name a model this skill no longer lists, ask the user whether to update them. 3. Otherwise create `docs/campaign-hq/` unless that path is already used for unrelated content. Any folder name is fine if `CLAUDE.md` points to it and every command uses it in place of `docs/campaign-hq/`. 4. Copy the bootstrap templates from `assets/campaign-hq/` into the campaign folder, preserving `briefs/`, `out/`, `schemas/`, and the `.gitignore` that keeps worker event streams out of git. When Codex is installed, also copy `assets/codex-agents/*.toml` into the project's `.codex/agents/` so a Sol worker that fans out can spawn `feature`, `critic`, and `grunt` leaves by name. 5. Add this block to the project `CLAUDE.md` (create the file if missing): ```markdown ## Active Campaign Campaign state lives in `docs/campaign-hq/`. Before doing project work, read CAMPAIGN.md (plan), LEARNINGS.md (history), and preferences.md (worker routing). Act as conductor: dispatch workers per preferences.md rather than implementing directly. Full instructions: the campaign-conductor skill. ``` After bootstrap, the campaign folder is the source of truth. Future sessions resume from the repo files and do not depend on this skill being loaded. ## Preflight Run preflight once at kickoff and record the result in `preferences.md`: - Codex CLI: `codex --version`, `codex login status`, and the model smoke tests in [Codex dispatch](references/codex-dispatch.md) - Other agent CLIs when the user has them: `agy --version`, `grok --version`, `muse --version` - GitHub CLI when CI gates matter: `gh auth status` - Project verification command: run the actual build/test/lint command workers will use - Permission envelope: Codex sandbox/network/approval policy, Codex project trust (the leaf role files load only in a trusted project), and Claude worker edit permissions Route work away from any tool that preflight finds missing. Without Codex, use the "Codex unavailable or rate-limited" row in Worker Routing. ## Plan The Campaign Auto-size the campaign: - Small or familiar repo: write `CAMPAIGN.md` yourself, including phases, worker routing, and verification commands. - Large or unfamiliar repo: dispatch read-only survey agents in parallel, one per concern (architecture, conventions, risk/debt, test story) and one per major package or service; synthesize their reports into `CAMPAIGN.md` and get user sign-off before writer waves. Use phases shaped as `dispatch -> collect -> integrate -> verify -> next wave`. Never mark a task done until an agent other than its author has read the diff and rerun the verification command, and the conductor has rerun verification on the integrated result. ## Worker Routing Precedence is: user's live instruction > `preferences.md` > these defaults. When the user states a routing preference, write it to `preferences.md` in the same turn. Model, worker, and effort requests are routing preferences. Do not silently replace a live request like "run this wave at xhigh" or "send tests to Opus" with a default. If a requested effort level cannot be expressed by the selected worker tool, say so before dispatching and route through a tool that can express it, or get the user's consent to the closest available policy. | Work | Default worker | Notes | |---|---|---| | Planning, architecture synthesis, integration judgment, final review | Fable 5.2 conductor at high; Opus 5.5 at high when Fable is unavailable or the user chooses it | Keep this in the main session unless a parallel survey helps. | | Implementation, refactors, tests, scripts, debugging | Codex CLI on `gpt-6.2-sol` at `high` | The worker sizes its own team to the task's disjoint file groups, or runs alone. Read [Codex dispatch](references/codex-dispatch.md) first. | | Judgment leaf work inside a Sol fan-out | `feature` role: `gpt-6.2-sol` at `high` | A sub-feature, bug fix, or test suite with a clear file boundary. | | Second opinion inside a Sol fan-out | `critic` role: `gpt-6.2-sol` at `high`, read-only by role file | Reviews a sibling leaf's diff in a fresh session. | | Mechanical leaf work inside a Sol fan-out | `grunt` role: `gpt-6.1-terra` at `xhigh` | Fixtures, renames, search, small tests; run as many at once as there are disjoint file groups. Move a task that fails verification twice to `feature`. | | Mechanical sweep across many files: codemod, rename, fixture regeneration | One Sol worker fanning out `grunt` leaves over disjoint file groups | One worktree, one commit, and one report, however many leaves run. | | UI/UX, visual design, design review, frontend polish | Claude Opus 5.5 at high or xhigh | Dispatch with `model: 'opus'` and an explicit `effort`. | | Cohesive sub-goal that needs its own integration branch or mid-flight steering | Claude Opus 5.5 squad lead at high or xhigh, running Sol workers | SendMessage steers a running lead. See [Squads](references/squads.md). | | Consultation: architecture questions, read-only review of a Claude worker's diff | Codex CLI on `gpt-6.2-sol` at `high`, read-only | See [Review gates](references/review-gates.md). | | Bake-off judging, final arbitration between a critic and an author | Codex CLI on `gpt-6.2-sol` at `xhigh`, read-only | Give it diffs and verification output, with criteria set before dispatch. | | Read-only surveys, cross-file search, research scouting | Claude Sonnet 5.5 at high or xhigh, or one Codex CLI scout on `gpt-6.1-terra` at `xhigh` read-only (`-c web_search=live` for research) | Require file/line evidence in a `scout-result.json` report. | | Quick lookups: a symbol's definition, call sites, errors in a worker's event stream | Claude Haiku 5.5 at high, read-only | Move a lookup that fails or needs synthesis across many files to Sonnet. | | Bulk read-only sweeps: one scout per package, file group, or question | Codex CLI on `gpt-6.1-terra` at `xhigh`, read-only, launched in parallel | Each scout returns a `scout-result.json` report. For a very wide sweep, one read-only Sol worker runs the scouts as `grunt` leaves and returns one synthesized report. | | Rare, very hard question on explicit user request | Codex CLI on `gpt-6.2-astra` at `xhigh`, read-only | Never a default and never a writer. Record the request in the fleet table. | | Second opinions from another model family | `agy` on `gemini-3.8-flash-high`, `grok` on `grok-4.7` at `xhigh`, or `muse` on `muse-spark-1.3-contributor` at `max` | Optional and read-only; a secondary review beside Claude and Codex. See [Review gates](references/review-gates.md). | | Codex unavailable or rate-limited | Claude Sonnet 5.5 workers at high or xhigh; Opus 5.5 at high or xhigh for design and for tasks that fail twice on Sonnet | Keep the same brief contract, worktree isolation, report schema, and verification gates; a Claude worker commits its own work. | No default, role file, or fallback runs below `high` or above `xhigh`. Opus 5.5 and Sonnet 5.5 subagents run at `high` and step up to `xhigh` for hard design calls, thorny debugging, or a task that failed at `high`. Haiku 5.5 runs at `high`. Codex shapes run at `high` or `xhigh`; `max` and `ultra` run only on an explicit request. Name the model and effort on every Claude dispatch. A subagent dispatched without a model inherits the conductor's by default, so under a Fable conductor a design agent or scout runs on Fable. The Agent tool takes `model` (`'opus'`, `'sonnet'`, or `'haiku'`) and `effort` (`'high'` or `'xhigh'`). A Workflow `agent()` call takes the same two options for a scripted fan-out (this skill counts as the Workflow opt-in). Codex `ultra` effort adds automatic delegation. Use it only on an explicit request, and never for a leaf, a consultant, or the Astra run. ## Dispatch Rules - Every worker brief must be self-contained: goal, owned files/dirs, exclusions, repo conventions, branch/worktree, verification command, commit rule, and final report format. - One writer per tree by default: use worktrees for overlapping work or more than 2-3 naturally disjoint writers. See [Fleet operations](references/fleet-operations.md). - Every writer branch must end with a commit, because integration cannot merge uncommitted output. A Claude worker commits its own work. A Codex worker cannot (its sandbox keeps git metadata read-only), so the collector commits the paths in its report. - Record each active dispatch in the `CAMPAIGN.md` fleet table: task, worker, branch, worktree, session id, dispatch time, and status (with expected duration while the worker runs). - Launch each Codex worker as a background command with its `--json` event stream redirected to a file, and read the `-o` report when it exits. The stream carries every command the worker ran and its output. - Claude workers: `isolation: 'worktree'` gives a writer its own worktree without manual setup. The harness picks the path and creates the worktree from the default branch, so the brief names the branch and base commit, the worker runs `git switch -c <branch> <base>` first, and the fleet table takes the path from `git worktree list`. SendMessage steers a running agent instead of respawning it. - A Sol worker's leaf fan-out needs no squad: it happens inside one worktree and ends as one commit. Use an Opus squad only for cohesive sub-goals where several Sol workers must integrate before the conductor needs the result. See [Squads](references/squads.md). ## Collect, Verify, Record The conductor owns correctness: - Parse each worker report and commit a Codex worker's reported paths. Read the diff yourself for high-risk or design-heavy work; route routine branches through an integration or verifier worker that did not write them, and read its report. - Rerun the decisive verification command yourself on each integrated result. - Integrate branches in dependency order. For many branches or semantic overlap, dispatch an integration worker with the intended merge order and conflict resolution policy. - Use cross-model review for high-risk diffs and after wave integration. See [Review gates](references/review-gates.md). - Update `CAMPAIGN.md` task status as work lands. - Log durable lessons in `LEARNINGS.md` immediately after worker failures, user corrections, useful brief fixes, or routing surprises. Compact repeated lessons into standing rules before the log passes roughly 40 lines. - Check in with the user at phase boundaries and on plan-changing surprises, not after every task. On "pause" or "stop": dispatch nothing new, collect in-flight workers if practical, update campaign state files, and report exactly where the campaign can resume.
Comments (0)
Sign in to join the conversation.
Reviews (0)
No reviews yet.
No comments yet.