All the quality-pipeline configuration in one place. Three channels: env vars (persistent), slashes (session), /config quality (inspection).
/config qualitySingle command that lists the full pipeline state:
Typical output:
✨ AGENT HARNESS/QUALITY PIPELINE ─────────────────────────── CHATCLI_QUALITY_ENABLED : enabled Hooks registered : pre=0, post=2 ── Self-Refine (#5) CHATCLI_QUALITY_REFINE_ENABLED : disabled CHATCLI_QUALITY_REFINE_MAX_PASSES : 1 CHATCLI_QUALITY_REFINE_MIN_BYTES : 200 CHATCLI_QUALITY_REFINE_EPSILON : 50 CHATCLI_QUALITY_REFINE_EXCLUDE : formatter, deps, refiner, verifier CHATCLI_QUALITY_REFINE_CONVERGENCE_ENABLED : enabled CHATCLI_QUALITY_REFINE_CONVERGENCE_EMBEDDING : disabled CHATCLI_QUALITY_REFINE_CONVERGENCE_STRICT : disabled CHATCLI_QUALITY_REFINE_CONVERGENCE_CHAR_HIGH : 0.99 CHATCLI_QUALITY_REFINE_CONVERGENCE_JACCARD_HIGH: 0.95 CHATCLI_QUALITY_REFINE_CONVERGENCE_EMBEDDING_SIM: 0.92 ── Chain-of-Verification / CoVe (#6) CHATCLI_QUALITY_VERIFY_ENABLED : disabled CHATCLI_QUALITY_VERIFY_NUM_QUESTIONS : 3 CHATCLI_QUALITY_VERIFY_REWRITE : enabled CHATCLI_QUALITY_VERIFY_EXCLUDE : formatter, deps, shell, refiner, verifier ── Reflexion (#3) CHATCLI_QUALITY_REFLEXION_ENABLED : enabled CHATCLI_QUALITY_REFLEXION_ON_ERROR : enabled CHATCLI_QUALITY_REFLEXION_ON_HALLUCINATION: enabled CHATCLI_QUALITY_REFLEXION_ON_LOW_QUALITY : disabled CHATCLI_QUALITY_REFLEXION_PERSIST : enabled CHATCLI_QUALITY_REFLEXION_QUEUE_ENABLED : enabled CHATCLI_QUALITY_REFLEXION_QUEUE_WORKERS : 2 CHATCLI_QUALITY_REFLEXION_QUEUE_CAPACITY: 1000 CHATCLI_QUALITY_REFLEXION_QUEUE_MAX_ATTEMPTS: 5 CHATCLI_QUALITY_REFLEXION_QUEUE_STALE_AFTER : 168h0m0s Runtime state : queue=0 dlq=0 ── Plan-and-Solve / ReWOO (#2) CHATCLI_QUALITY_PLAN_FIRST_MODE : auto CHATCLI_QUALITY_PLAN_FIRST_THRESHOLD : 6 ── RAG + HyDE (#4) CHATCLI_QUALITY_HYDE_ENABLED : disabled CHATCLI_QUALITY_HYDE_USE_VECTORS : disabled CHATCLI_EMBED_PROVIDER : (not set) CHATCLI_EMBED_MODEL : (not set) CHATCLI_QUALITY_HYDE_NUM_KEYWORDS : 5 Vector index : (not attached) ── Reasoning backbone (#7) CHATCLI_QUALITY_REASONING_MODE : auto CHATCLI_QUALITY_REASONING_BUDGET : 8000 CHATCLI_QUALITY_REASONING_AUTO_AGENTS: planner, refiner, verifier, reflexion
Slash commands /thinking — reasoning override/thinking # show current state /thinking auto # clear override /thinking off # force no-thinking for next turn /thinking on # alias for /thinking high /thinking low | medium | high | max /thinking budget=N # nearest tier to N tokens
Details: #7 Reasoning Backbone .
/plan — force Plan-and-SolveSix forms (autocomplete via Tab)
/plan # arm flag; next /agent or /coder uses Plan-First /plan < free tas k > # arm + enter agent mode and execute /plan agent < tas k > # explicit equivalent of the previous form /plan coder < tas k > # enter coder mode (software engineer) and execute /plan preview < tas k > # dry-run: generate and render the plan WITHOUT executing /plan dry < tas k > # alias of preview
Recommended flow for large changes: /plan preview <task> → review the plan → /plan coder <same task> to execute.
Details: #2 Plan-and-Solve .
/refine — Self-Refine session toggle/refine # current state /refine on # enable for session /refine off # disable /refine once | next # enable (today identical to on) /refine auto | clear # clear override → use /config
Details: #5 Self-Refine .
/verify — CoVe session toggle/verify # current state /verify on # enable /verify off # disable /verify once # enable once /verify auto # clear override
Details: #6 CoVe .
/reflect — durable lesson queuePersist manual lesson (no LLM)
Status
Inspect the queue
DLQ operations
/reflect < free text of the lesso n > # Ex: /reflect when editing large Go files use Edit, not full rewrite
/reflect <free text> writes directly to memory.Fact (category=lesson, trigger=manual) without an LLM call . The queue subcommands (list/failed/retry/purge/drain) show and operate on automatic triggers (error, hallucination, low quality) buffered in the WAL awaiting async processing./reflect retry and /reflect purge feature dynamic autocomplete: Tab pulls live IDs from the DLQ with task preview + last error.
Full details: #3 Reflexion → Durable Queue .
Env vars — full reference Master switch Env var Default Values Effect CHATCLI_QUALITY_ENABLEDtruetrue|falseTurn off the whole pipeline. false = Pipeline.Run → agent.Execute direct, zero overhead
Self-Refine (#5) Env var Default Values Effect CHATCLI_QUALITY_REFINE_ENABLEDfalsetrue|falseEnable RefineHook CHATCLI_QUALITY_REFINE_MAX_PASSES1int ≥ 1 Hard cap on passes CHATCLI_QUALITY_REFINE_MIN_BYTES200int ≥ 0 Skip outputs smaller than N bytes CHATCLI_QUALITY_REFINE_EPSILON50int ≥ 0 Char-level fallback threshold CHATCLI_QUALITY_REFINE_EXCLUDEformatter,deps,refiner,verifierCSV Agents that don’t pass through refine
Semantic convergence cascade Env var Default Values Effect CHATCLI_QUALITY_REFINE_CONVERGENCE_ENABLEDtruetrue|falseCascade master switch (char→Jaccard→embedding) CHATCLI_QUALITY_REFINE_CONVERGENCE_EMBEDDINGfalsetrue|falseInclude embedding scorer (opt-in — costs $) CHATCLI_QUALITY_REFINE_CONVERGENCE_STRICTfalsetrue|falseStrict: refuse convergence without embedding CHATCLI_QUALITY_REFINE_CONVERGENCE_CHAR_HIGH0.990.0-1.0 Char short-circuit CONVERGED CHATCLI_QUALITY_REFINE_CONVERGENCE_CHAR_LOW0.30.0-1.0 Char short-circuit DIVERGED CHATCLI_QUALITY_REFINE_CONVERGENCE_JACCARD_HIGH0.950.0-1.0 Jaccard short-circuit CONVERGED CHATCLI_QUALITY_REFINE_CONVERGENCE_EMBEDDING_SIM0.920.0-1.0 Final embedding cosine threshold CHATCLI_QUALITY_REFINE_CONVERGENCE_CACHE_SIZE256int LRU cache size CHATCLI_QUALITY_REFINE_CONVERGENCE_CACHE_TTL_MIN5int Cache TTL (minutes) CHATCLI_QUALITY_REFINE_CONVERGENCE_BREAKER_THRESHOLD3int Failures before breaker opens
CoVe (#6) Env var Default Values Effect CHATCLI_QUALITY_VERIFY_ENABLEDfalsetrue|falseEnable VerifyHook CHATCLI_QUALITY_VERIFY_NUM_QUESTIONS3int 1-7 Number of verification questions CHATCLI_QUALITY_VERIFY_REWRITEtruetrue|falseRewrite output on discrepancy CHATCLI_QUALITY_VERIFY_EXCLUDEformatter,deps,shell,refiner,verifierCSV Agents that don’t pass through verify
Reflexion (#3) Env var Default Values Effect CHATCLI_QUALITY_REFLEXION_ENABLEDtruetrue|falseEnable ReflexionHook CHATCLI_QUALITY_REFLEXION_ON_ERRORtruetrue|falseFire on worker error CHATCLI_QUALITY_REFLEXION_ON_HALLUCINATIONtruetrue|falseFire on verified_with_discrepancy CHATCLI_QUALITY_REFLEXION_ON_LOW_QUALITYfalsetrue|falseFire on refine_low_quality CHATCLI_QUALITY_REFLEXION_PERSISTtruetrue|falseWrite to memory.Fact
Durable queue (WAL + worker pool + DLQ) Env var Default Values Effect CHATCLI_QUALITY_REFLEXION_QUEUE_ENABLEDtruetrue|falseEnable the queue (default). false = legacy mode (goroutine) CHATCLI_QUALITY_REFLEXION_QUEUE_WORKERS2int ≥ 1 Workers processing in parallel CHATCLI_QUALITY_REFLEXION_QUEUE_CAPACITY1000int In-memory cap CHATCLI_QUALITY_REFLEXION_QUEUE_DROP_OLDESTfalsetrue|falsetrue = drop oldest; false = blockCHATCLI_QUALITY_REFLEXION_QUEUE_BLOCK_TIMEOUT5sduration Enqueue timeout when queue is full CHATCLI_QUALITY_REFLEXION_QUEUE_MAX_ATTEMPTS5int ≥ 1 Total retries CHATCLI_QUALITY_REFLEXION_QUEUE_INITIAL_DELAY1sduration First retry delay CHATCLI_QUALITY_REFLEXION_QUEUE_MAX_DELAY5mduration Cap on exponential backoff CHATCLI_QUALITY_REFLEXION_QUEUE_JITTER0.20.0-0.5 Fractional jitter CHATCLI_QUALITY_REFLEXION_QUEUE_JOB_TIMEOUT2mduration Per-job timeout CHATCLI_QUALITY_REFLEXION_QUEUE_STALE_AFTER168hduration Old records discarded on replay CHATCLI_QUALITY_REFLEXION_QUEUE_BASE_DIR— path Queue directory override
Plan-and-Solve (#2) Env var Default Values Effect CHATCLI_QUALITY_PLAN_FIRST_MODEautooff|auto|alwaysWhen to fire CHATCLI_QUALITY_PLAN_FIRST_THRESHOLD6int 0-10 Minimum score for auto to fire
RAG + HyDE (#4) Env var Default Values Effect CHATCLI_QUALITY_HYDE_ENABLEDfalsetrue|falseEnable phase 3a (keyword expansion) CHATCLI_QUALITY_HYDE_USE_VECTORSfalsetrue|falseEnable phase 3b (vector search) CHATCLI_QUALITY_HYDE_NUM_KEYWORDS5int ≥ 1 Hypothesis keyword cap
Embedding providers (used by HyDE 3b) Env var Default Values Effect CHATCLI_EMBED_PROVIDERnullvoyage / openai / bedrock / nullEmbedding backend picker CHATCLI_EMBED_MODELprovider default string Voyage: voyage-3. OpenAI: text-embedding-3-small / -large. Bedrock: amazon.titan-embed-text-v2:0 (default), amazon.titan-embed-text-v1, cohere.embed-english-v3, cohere.embed-multilingual-v3. CHATCLI_EMBED_DIMENSIONSmodel native int OpenAI: truncate via Matryoshka (text-embedding-3-small=1536, -large=3072 native). Bedrock Titan v2: accepts 256 / 512 / 1024 (rejects others). Bedrock Titan v1 / Cohere v3: fixed dim, ignored. VOYAGE_API_KEY— string Required for provider=voyage OPENAI_API_KEY— string Required for provider=openai (uses the same chat key) BEDROCK_REGION / AWS_REGION / AWS_PROFILE / AWS credentialsper chat Bedrock — Required for provider=bedrock — reuses the same chain as the AWS Bedrock feature (IAM role, SSO, assume-role, profile).
Bedrock embeddings support Titan (single text per call — parallelized internally with an 8-worker pool) and Cohere v3 (native batch). Family dispatch is automatic from the model id prefix. See RAG + HyDE for the full architecture.Reasoning Backbone (#7) Env var Default Values Effect CHATCLI_QUALITY_REASONING_MODEautooff|auto|onAuto-attach policy CHATCLI_QUALITY_REASONING_BUDGET8000int (tokens) Thinking budget (Anthropic); mapped to tier on OpenAI CHATCLI_QUALITY_REASONING_AUTO_AGENTSplanner,refiner,verifier,reflexionCSV List for mode=auto
Per-agent overrides (apply to any built-in) Env var Effect CHATCLI_AGENT_<NAME>_MODELForce specific model for that agent CHATCLI_AGENT_<NAME>_EFFORTForce effort tier (low|medium|high|max)
<NAME> is the uppercase name. For the new ones: REFINER, VERIFIER.# Examples export CHATCLI_AGENT_REFINER_MODEL = "claude-haiku-4-5" export CHATCLI_AGENT_VERIFIER_MODEL = "claude-opus-4-8" export CHATCLI_AGENT_VERIFIER_EFFORT = "max"
Recommended presets Cheap dev (default)
Rigorous code review
Technical docs
Incident investigation
Batch autopilot
# No adjustments: chatcli out-of-the-box is already cheap. # ReAct + Reasoning auto for Planner + Reflexion on (only on errors)
Extra cost per turn: zero normally; +1 LLM call on rare errors.export CHATCLI_QUALITY_REFINE_ENABLED = true export CHATCLI_QUALITY_REFINE_MAX_PASSES = 2 export CHATCLI_QUALITY_VERIFY_ENABLED = true export CHATCLI_QUALITY_VERIFY_NUM_QUESTIONS = 5
Extra cost: +3-4 calls per worker with non-mechanical output. Use in critical PR reviews.export CHATCLI_QUALITY_REFINE_ENABLED = true export CHATCLI_QUALITY_VERIFY_ENABLED = true export CHATCLI_QUALITY_HYDE_ENABLED = true # Optional: HyDE vectors for max recall export CHATCLI_QUALITY_HYDE_USE_VECTORS = true # Pick an embedding provider (any of these works): # — Voyage (Anthropic-recommended) export CHATCLI_EMBED_PROVIDER = voyage export VOYAGE_API_KEY = pa- ... # — OpenAI (if OPENAI_API_KEY is already in the environment) # export CHATCLI_EMBED_PROVIDER=openai # — Bedrock (same AWS chain as chat, no extra API key) # export CHATCLI_EMBED_PROVIDER=bedrock # export CHATCLI_EMBED_MODEL=amazon.titan-embed-text-v2:0 # export BEDROCK_REGION=us-east-1
Extra cost: +1 HyDE hypothesis + refine + verify per doc generation. Polished and factually-checked output.export CHATCLI_QUALITY_REASONING_MODE = on export CHATCLI_QUALITY_REASONING_BUDGET = 16384 export CHATCLI_QUALITY_REFLEXION_ON_LOW_QUALITY = true
Extra cost: max thinking across all agents + generous lesson generation. Use in root-cause analysis.export CHATCLI_QUALITY_PLAN_FIRST_MODE = always export CHATCLI_QUALITY_REFINE_ENABLED = true export CHATCLI_QUALITY_VERIFY_ENABLED = true export CHATCLI_QUALITY_REASONING_MODE = on
Extra cost: all patterns always on. Use in workflows without a user in front (cron, CI).Turn everything off If something breaks in production and you need to return to pre-pipeline behavior instantly:
export CHATCLI_QUALITY_ENABLED = false
This makes Pipeline.Run degenerate to return agent.Execute(...) — byte-identical to pre-PR chatcli. Zero hooks, zero reasoning auto-attach, zero HyDE.
The master switch is the emergency exit. Individual toggles are for tuning scenarios; the master is for rollback.
Interaction with other configs The harness/quality pipeline does not conflict with, nor replace:
Feature Relationship Skills (with frontmatter effort:) Wins over applyAutoReasoning (skill hint already on ctx) Personas (custom agents) Can have effort: in their .md; per-agent override still works Multi-agent orchestration Dispatcher stays identical; pipeline only wraps each worker Policy (CODER policy rules) Applied inside the worker, independent of refine/verify/reflexion MCP tools Result of MCP tools goes through the pipeline like any other worker output
See also
Seven patterns overview Back to the hub.
Environment Variables (full) The whole ChatCLI env var matrix.
Command Reference Every slash.
/config (hierarchical) How /config <section> was structured.
