Add robustness loops: quality validation, consensus, and critic-judge refinement

.env.example

@@ -96,3 +96,89 @@ CLK_TELEGRAM_ENABLED=false
 CLK_TELEGRAM_SKIP=false
 # Default workspace the bot operates on when /workspace has not been set.
 CLK_TELEGRAM_WORKSPACE=
+
+# ----------------------------------------------------------------------
+# Robustness loops (see README "Robustness loops" section)
+# ----------------------------------------------------------------------
+# Every layer below has an off-switch so you can dial cost.  All values
+# are forwarded into .clk/config/clk.config.json by kickoff.sh; you can
+# also edit that file directly to override per project.
+
+# Auto-consensus fan-out: every "careful" dispatch fans into N parallel
+# stochastic samples + chief coalescing, instead of a single dispatch.
+#   off          — only PROPOSE_CONSENSUS triggers fan-out (legacy)
+#   on_careful   — fan out stages marked careful=true (default)
+#   always       — fan out every non-chief dispatch (×N cost)
+# Cost: ~×consensus.max_samples per affected dispatch.
+CLK_ROBUSTNESS_AUTO_CONSENSUS=on_careful
+
+# Critic-judge refinement: draft -> critic -> revise inner loop.
+#   off            — only explicit refine: blocks in workflow YAML
+#   careful_only   — careful=true stages get one critic pass (default)
+#   all            — every non-chief, non-qa stage gets one critic pass
+# Cost: +1 critic dispatch + up to refine_max_rounds-1 worker revisions
+# per affected stage.
+CLK_ROBUSTNESS_AUTO_REFINE=careful_only
+
+# After every dispatch the response is scored for emptiness, malformed
+# ACTION/POST blocks, missing declared outputs, and low confidence;
+# recoverable failures re-dispatch with a repair preamble, escalating
+# to consensus on the final retry. Set to 0 to disable.
+CLK_ROBUSTNESS_MAX_QUALITY_RETRIES=2
+
+# Responses shorter than this many characters are treated as suspect-empty.
+CLK_ROBUSTNESS_MIN_RESPONSE_CHARS=40
+
+# Critic-judge inner loop bounds.
+CLK_ROBUSTNESS_REFINE_MAX_ROUNDS=3
+CLK_ROBUSTNESS_REFINE_ACCEPT_THRESHOLD=0.8
+
+# Inter-agent Q&A protocol: POST: question TO: <peer> URGENCY: blocking.
+# qa_parallel_judges optionally dispatches multiple peers in parallel.
+# max_qa_depth caps the chain length (peer A asks B asks C ...).
+CLK_ROBUSTNESS_QA_PARALLEL_JUDGES=1
+CLK_ROBUSTNESS_MAX_QA_DEPTH=3
+
+# Ralph / autoresearch plateau detection. After plateau_window
+# consecutive iterations without improvement, the loop escalates +
+# reframes; failing that, it terminates gracefully with done.md.
+#   off                     — never auto-terminate (legacy behavior)
+#   escalate_only           — force consensus on next iterations only
+#   reframe_only            — only ask chief to re-cast
+#   escalate_then_reframe   — both, then terminate (default)
+CLK_ROBUSTNESS_PLATEAU_WINDOW=3
+CLK_ROBUSTNESS_PLATEAU_ACTION=escalate_then_reframe
+
+# ----------------------------------------------------------------------
+# Prior-knob reference (already supported; documented here for parity)
+# ----------------------------------------------------------------------
+# Provider invocation timeout, seconds. 0 = use harness default (300s).
+CLK_PROVIDER_TIMEOUT_S=0
+# Per-provider-invocation no-output watchdog (kill if no stdout for N s).
+CLK_PROVIDER_NO_OUTPUT_TIMEOUT_S=0
+# Provider-level retry policy (transient errors: rate limits, timeouts, ...).
+CLK_PROVIDER_RETRY_MAX_RETRIES=10
+CLK_PROVIDER_RETRY_BACKOFF_S=5
+# Stage-level retry policy (whole stage repeats with a larger backoff).
+CLK_PROVIDER_RETRY_STAGE_MAX_RETRIES=10
+CLK_PROVIDER_RETRY_STAGE_BACKOFF_S=30
+# Chief supervise loop: cap on cycles before warning and stopping.
+CLK_SUPERVISE_MAX_CYCLES=20
+# Stochastic consensus caps (same scope as the auto-consensus knob).
+CLK_CONSENSUS_MAX_SAMPLES=6
+CLK_CONSENSUS_MAX_PARALLEL=4
+# Dynamic-roles cap (chief cannot mint more than this many dynamic agents).
+CLK_CASTING_MAX_DYNAMIC_ROLES=12
+# Auto-commit per agent-run after successful ACTION application.
+CLK_AUTO_COMMIT=true
+# File-action batch caps (per agent response).
+CLK_VALIDATION_MAX_FILES_PER_BATCH=25
+CLK_VALIDATION_WARN_FILES_PER_BATCH=5
+# Meta-prompt drafting (chief tightens worker prompts before dispatch).
+#   off | careful_only (default) | always
+CLK_META_PROMPT_DISPATCH=careful_only
+CLK_META_PROMPT_ROLE=careful_only
+# Per-stage checkpoint (chief CONTINUE/REDIRECT/ABORT after every stage).
+CLK_REVIEW_PER_STAGE=false
+# Recovery on unmet deps: chief recovery passes per stage.
+CLK_RECOVERY_MAX_PER_STAGE=3