| name | codi-agents |
| description | Whale-spawns-Codi-Agents orchestration dispatcher. Routes T0-T2 orchestration tasks (prompt sharpening, persona detection, calibration scanning, hallucination pre-flight, lightweight jury) AWAY from the active premium model (whale = Sonnet/Opus/Claude) and TO cheap Codi Agents (Gemini Flash via OAuth, GPT-5.4 via codex CLI). Fixes the token-burn bug where premium reasoning capacity was being spent on mechanical orchestration. FAIL-HARD: if no Codi Agents are reachable, BLOCKS — never silently falls back to whale. Activate when the user says "codi agents", "spawn agents", "delegate to agents", "stop burning tokens", "cheap orchestration", or when any other skill needs a T0-T2 verdict on prompt-sharpen / persona-detect / calibration-scan / hallucination-preflight / t0t2-jury.
|
| metadata | {"version":"4.12.0","author":"Anand Vallamsetla","tier":"core","plugin_number":9} |
BEGIN CODI-AGENTS
Codi Agents — Whale-Spawns-Agents Orchestration Dispatcher
Plugin #9, Core tier. Operationalizes the Whale-and-Agents primitive: the
active premium model (whale) dispatches mechanical orchestration tasks to
cheap Codi Agents, never burns its own tokens on T0-T2 work. Constitution
anchors: 3D Execution Axiom (Optimal Cost), OPERATIONAL DISCIPLINE (right-size
by task class), FAIL-HARD invariant (no silent fallback to premium).
Why this exists
Premium reasoning capacity is precious. The whale (Sonnet/Opus/Claude in the
active session) costs $0.01-$0.05 per turn. When it spends those tokens on
mechanical tasks — rewriting a vague prompt, classifying a domain, scanning
for sycophancy phrases, picking a risk label, deciding if a draft is
ship-ready at T0-T2 — every cycle is a waste of the partner's most valuable
capability. The user's directive 2026-04-27: "I am also burning tokens, I
need small agent swarm to run without fail."
Codi Agents fixes this by routing every T0-T2 orchestration task to a
cross-family cheap-agent cascade (the same harness judge-panel uses), with
FAIL-HARD discipline so the whale never silently absorbs the cost.
What counts as a Codi Agent (and what does NOT)
| Tier | Status | Examples |
|---|
| Local (free) | Disabled — enable when local LLM installed | DeepSeek-R1 7B, Llama 3.1 8B, Mistral 7B, Phi-4 via Ollama |
| Cheap OAuth CLI | Primary | Gemini Flash via agy, GPT-5.4 via codex CLI (OAuth) |
| Premium API | NEVER | gpt-4o, claude-* anything, gemini-pro |
| Active session model (whale) | NEVER for orchestration | Sonnet, Opus, Haiku |
Haiku is excluded — it's still Anthropic-tier, billed against the user's
Claude quota. The point of Codi Agents is to leave the Claude lane untouched
for T3-T4 work.
The five orchestration tasks
Each task is one rubric in the existing judge-panel harness
(scripts/judge_panel.ts). The harness does the cross-family cheap-agent
cascade; this skill is the prose contract that names the tasks and routes
calls. No parallel Python harness — REUSE.
| Task | Rubric slug | Input | Output (in flags[0]) |
|---|
| Prompt sharpening | prompt-sharpen | A vague user prompt | The sharpened prompt verbatim, OR ALREADY_SHARP, OR NEEDS_USER_INPUT: <what> |
| Persona detection | persona-detect | A user prompt / task | Persona slug (architecture, product, legal, life-coach, etc.); if multi-domain, flags[0..1] = top 2 |
| Calibration / sycophancy scan | calibration-scan | LLM-generated text | Each flagged phrase verbatim, one per flag entry |
| Hallucination pre-flight | hallucination-preflight | A user prompt (BEFORE LLM responds) | flags[0] = risk label (FACTUAL/LEGAL/MEDICAL/FINANCIAL/TEMPORAL/CITATION/NONE), flags[1] = grounding action (WEB_SEARCH/PRIMARY_SOURCE/USER_CONFIRMATION/ARXIV_RECENT/PATENT_DB/NONE_NEEDED) |
| T0-T2 lightweight jury | t0t2-jury | Any artifact (code/prose/plan/decision) | One-line pass/fail/uncertain reason |
All five rubrics return the standard judge-panel JSON:
{verdict, confidence, flags, ...cascade metadata}. Callers parse stdout.
How the active model invokes Codi Agents
bun run plugins/co-dialectic/skills/judge-panel/scripts/judge_panel.ts \
--rubric <slug> \
--artifact-file <path> \
--silent
Or inline:
bun run plugins/co-dialectic/skills/judge-panel/scripts/judge_panel.ts \
--rubric prompt-sharpen \
--artifact "make a thing that does the stuff" \
--silent
The harness handles the cascade (≥2 cross-family small Codi Agents parallel;
one big-agent tiebreaker on disagreement / low confidence). The whale parses
the JSON final_verdict + all_flags and proceeds. The whale does NOT
re-reason about the orchestration task in its own context.
Silent mode is the default for Codi Agents
When Codi Agents is invoked from another skill or from the active model's
orchestration path, ALWAYS pass --silent. Conversational framing is for
explicit user-invoked judge-panel runs, not background orchestration.
Health check — fires at session start (and on demand)
When co-dialectic activates, Codi Agents probes the agent pool and reports
inline. The probe is two independent checks; all run in parallel.
command -v agy >/dev/null && agy --version >/dev/null 2>&1
command -v codex >/dev/null && codex --version >/dev/null 2>&1
The active model runs these directly via Bash (no separate script needed).
A passing probe is exit 0 + non-empty body. Tally the count.
Note: Local Ollama is excluded until a local LLM is installed. To re-enable,
add curl -s --max-time 2 http://localhost:11434/api/tags as check #3 and
restore the 3-agent status entries below.
Status line (codi conversation surface)
The active model emits ONE line at session start, and on any health-state
transition during the session:
| Agent count | Status line |
|---|
| 2 (Gemini + codex) | 🤖 Codi Agents: full — Gemini Flash + GPT-5.4 active |
| 1 (only one reachable) | ⚠ Codi Agents: degraded — only <which-one> active |
| 0 (NONE reachable) | ❌ Codi Agents: unavailable — orchestration BLOCKED — see remediation |
FAIL-HARD contract
Per the FAIL-HARD invariant (Constitution Ground Zero): if zero Codi Agents
are reachable, this dispatcher BLOCKS and surfaces remediation. It NEVER
silently routes the orchestration task back to the whale. Soft fallback to
the active premium model is the exact failure mode this skill exists to fix.
When agent_count == 0, the active model emits this exact block to the user
and refuses to proceed with the orchestration task:
❌ Codi Agents: unavailable
Cannot dispatch orchestration tasks (prompt-sharpen, persona-detect,
calibration-scan, hallucination-preflight, t0t2-jury) — no cheap-agent
endpoint is reachable.
Remediation (any one restores the pool):
1. Gemini OAuth CLI (free with Gemini Pro subscription):
npm i -g @google/generative-ai-cli # or vendor instructions
gcloud auth login
2. OpenAI Codex CLI (free with ChatGPT Plus / Pro subscription):
brew install codex # or vendor install
codex login
Re-run the same command after restoring the agent pool.
The active model does NOT proceed with the user's orchestration request
until at least one Codi Agent is back. T3-T4 work (architectural decisions,
irreversible artifacts) the whale handles directly — that's not what
Codi Agents gates.
When the active model SHOULD invoke Codi Agents
Per v4.1 Protocol 11a, Codi Agents is now the AUTO-ROUTING TARGET for
mechanical sub-tasks (when codi agent-swarm on, default). Activation surface
expanded from explicit-only to: explicit OR auto-routed by Protocol 11.
Dispatch mandate table — ALWAYS dispatch these patterns, NEVER absorb inline (Seam 3 fix — 2026-05-05):
The whale absorbing these patterns inline is the #1 token-burn failure mode
(observed 2026-05-05: whale did 200-line Rumsfeld audits, gap analysis, prompt
sharpening all inline — zero Codi Agents dispatched). The mandate is structural:
if the task matches a row below, the whale MUST spawn an Agent() call with
the rubric, not inline-generate.
| Task pattern | Rubric | Spawn or inline? |
|---|
| User pasted vague prompt → whale about to ask clarifying questions | prompt-sharpen | MUST spawn — offer sharpened version from Codi Agent |
| User prompt spans clear domain | persona-detect | MUST spawn — adopt persona from agent verdict |
| Whale draft contains "great", "absolutely", "perfect", "amazing", "excellent" | calibration-scan | MUST spawn — strip flagged phrases, regenerate |
| User asks about a fact, citation, current date, recent event, legal/medical/financial claim | hallucination-preflight | MUST spawn — fire grounding action, THEN respond |
| T0-T2 artifact ready to ship | t0t2-jury | MUST spawn — one-line pass/fail before emitting |
| Audit/gap-analysis task ≤200 lines | t0t2-jury | MUST spawn — not an inline Rumsfeld audit |
| Pattern-match/scan across a document section | t0t2-jury | MUST spawn — mechanical scan, not whale reasoning |
Whale-inline anti-pattern (observed, from session 2026-05-05):
- ❌ Whale doing a 200-line Rumsfeld audit of content-flywheel.md inline — should have been
t0t2-jury dispatch
- ❌ Whale doing gap analysis of social distribution engine inline — should have been
t0t2-jury dispatch
- ❌ Whale absorbing prompt-sharpening turn without calling
prompt-sharpen rubric
- ❌ Whale doing persona detection from topic keywords without calling
persona-detect
Litmus test before any inline analysis: "Is this a pattern-match, scan, or mechanical judgment task? If yes — dispatch to Codi Agent, do NOT inline."
Never, for these patterns (T3-T4 — whale's job):
- Architectural decisions (system design, schema choices, irreversible plans)
- Outreach to real humans (cold emails, posts to subscribers)
- Code that ships to production
- Patent or legal artifacts
- Anything in the EMERGENT SYSTEM IMMUNITY T3-T4 tiers — those go through
the full judge-panel cascade, not the lightweight Codi Agents dispatch
The boundary between Codi Agents and judge-panel is the stakes-tier:
Codi Agents = T0-T2 orchestration; judge-panel = T3-T4 verification. Same
harness, same cascade, different rubrics, different escalation thresholds.
How to verify Codi Agents are healthy (one command)
( command -v agy >/dev/null && echo "agy: ✓" || echo "agy: ✗" ) ; \
( command -v codex >/dev/null && echo "codex: ✓" || echo "codex: ✗" )
Two lines. Any one ✓ = Codi Agents operational. Zero ✓ = BLOCKED per
FAIL-HARD.
Smoke test — verify the skill end-to-end
bun run plugins/co-dialectic/skills/judge-panel/scripts/judge_panel.ts \
--rubric prompt-sharpen \
--artifact "make a thing that does stuff with the data" \
--silent
Expected: JSON object with final_verdict: "pass" and all_flags[0]
containing a sharpened version of the prompt (specific input/output, named
data type, intent clear). Cost reported in cost_usd_estimate should be
< $0.001 — that's the whole point.
Relationship to other skills
- judge-panel (plugin #4): Codi Agents REUSES the harness, adds five
rubrics. T0-T2 dispatch goes through Codi Agents; T3-T4 verification goes
through judge-panel. Same code path, same cascade discipline, different
use-site.
- hallucination-detector (plugin #3): the
hallucination-preflight
rubric in Codi Agents runs BEFORE response generation; the full
hallucination-detector runs AFTER. Pre-flight is cheap (T0-T2); post-flight
is full verification (T3-T4) when stakes warrant.
- calibration-auditor: the
calibration-scan rubric is the lightweight
passive scan. Heavy calibration audits (cross-session sycophancy patterns)
remain calibration-auditor's job.
- co-dialectic core: when the user enables Cruise or Drive mode,
Codi Agents dispatch becomes the default for orchestration steps; whale
handles only T3-T4. When Quiet mode, Codi Agents still fires but verdicts
are silent.
Anti-patterns
- ❌ Routing T3-T4 work to Codi Agents — use judge-panel's full cascade.
- ❌ Silently falling back to the whale when agent pool is down — FAIL-HARD
violation, defeats the whole token-burn fix.
- ❌ Caching agent verdicts across sessions — verdicts are session-scoped;
the underlying state may have changed.
- ❌ Adding new rubrics by writing a parallel harness — extend
judge_panel.ts's RUBRICS dict, never duplicate the cascade code.
- ❌ Using Haiku as a Codi Agent — it's still Anthropic-tier; defeats the
cost separation.
- ❌ Skipping the health probe at session start — the user needs to know
the agent pool is up BEFORE the first orchestration task fires, not after
a silent failure.
END CODI-AGENTS
