Harness nudges and completion gates, by name

Generated by node scripts/nudge-index.mjs --update (checked by tests/nudge-index.test.ts; the descriptions live in WHEN in that script). A nudge is a user-role message the loop appends after a step (type: nudge in the turn log); a gate is a check at COMPLETED that can bounce the reply with its own nudge. The texts are in the files named; docs/prompts/nudges-and-completion-gate.txt has the long ones by constant. Benchmark-only gates live in bench-policy.ts and are installed only when an FG_* switch is set.

name	kind	defined in	when
bad_args	nudge	llm-loops.ts	the arguments of a tool call were not valid JSON or arrived damaged (call markers of the model itself inside the argument names)
bad_tool_format	nudge	llm-loops.ts	the model wrote a tool call together with a fabricated result
blocked_untried_fix	gate	bench-policy.ts	benchmark policy: BLOCKED declared before a fix was tried
broken_syntax	gate	completion-gates.ts	gate: COMPLETED while a file changed this turn (by the agent or a worker) no longer parses; names the error
broken_tool_call	nudge	llm-loops.ts	a tool call written as text (markup in the message), so nothing ran
calls_overflow	nudge	llm-loops.ts	more than 20 tool calls in one response: only the first 20 ran
completion_gate	nudge	llm-loops.ts	first COMPLETED after edits: the completion checklist (and, in benchmark runs, the graded-test bounce)
dup_cycle	nudge	llm-loops.ts	the same sequence of calls repeating in a cycle
empty_code	nudge	llm-loops.ts	execute_code called with no code argument
env_config_edits	gate	completion-gates.ts	gate: test or packaging configuration edited, or a new top-level package added, instead of fixing the code
env_failure	nudge	llm-loops.ts	execute_code failing repeatedly with the same error (3) or in total (8); also two inline python3 -c syntax failures in a row
failed_check_complete	gate	completion-gates.ts	gate: COMPLETED while the last test/compile check after the last change failed
invented_data	gate	completion-gates.ts	gate: the work created stand-in data or files nobody supplied
no_change_fix	gate	bench-policy.ts	benchmark policy: a "fix" that changed nothing
no_edit_complete	gate	bench-policy.ts	benchmark policy: COMPLETED on a task that needs an edit, with no edit made
output_cut_off	nudge	llm-loops.ts	the response was cut off by the output cap or ended in a degenerate repeated tail
output_stalled	nudge	llm-loops.ts	generation stalled mid-response and was discarded; nothing ran
quality_check	nudge	llm-loops.ts	runaway or low-quality response (repeated characters, long loops)
reactive_guidance	nudge	llm-loops.ts	a skill whose trigger matched the situation this step; forces a tool call
replace_fail	nudge	llm-loops.ts	repeated replace_in_file failures on the same file
rewrite_loop	nudge	llm-loops.ts	the same file written in full 4, 8, 12 ... times with no passing check between
same_error	nudge	llm-loops.ts	the same error repeated several times in a row
scratch_only_edits	gate	completion-gates.ts	gate: COMPLETED but only scratch/repro files were changed
step_validation	nudge	llm-loops.ts	a step output check (pseudo tool call, missing state line, ...) failed
steps_left	nudge	llm-loops.ts	few steps left in the turn
stuck_detected	nudge	llm-loops.ts	three identical full results in a row
syntax_watch	nudge	llm-loops.ts	experiment syntax_watch: a file the model changed has not parsed for 2 steps (again after 6); lists every file that does not parse
tool_failure	nudge	llm-loops.ts	two to four consecutive tool errors
tool_failures	nudge	llm-loops.ts	five or more consecutive failing tool calls
tool_repeat	nudge	llm-loops.ts	a tool call repeated within the turn: cached results are returned
turn_state	nudge	llm-loops.ts	protocol reminder: reply has no COMPLETED/BLOCKED line
worker_empty_report	nudge	workers.ts	a worker ended with an empty message
worker_env_fail	nudge	workers.ts	a worker's commands failing repeatedly (environment)
worker_no_tools	nudge	workers.ts	a worker answered without calling any tool
worker_runaway	nudge	workers.ts	a worker's response ran away (repetition / quality check)
worker_stuck	nudge	workers.ts	a worker got identical results repeatedly